What’s inside (version for the busiest people)

Meta develops Max humanoid with Fourier, targets household tasks
→ 106 lb prototype (1.5m tall, NVIDIA Jetson Thor) built in 3 months at Shanghai factory; VP Li-Chen Miller embedded 2 months at Fourier
→ Why partnership + speed matters: Meta positioning humanoid as physical extension of Muse AI agent, household assistance (long-term), no commercial sales planned (research focus)

Bedrock Robotics hits $4.1B valuation, autonomous excavators on paid sites
→ $600M Series C, $1B raised total; August milestone: fully autonomous excavators on live commercial job sites (water treatment Nevada, earthwork Texas, civil work Zachry, Haydon contractor)
→ Why retrofit-only model (no manufacturing capital) + labor supply crisis matters: 5+ years to master excavator operation, 40% construction workforce retiring in 5 years, customer base doubled, expanding West

Machenta emerges with $40M, single-demo robot learning from simulation
→ Imperial spinout (Professor Edward Johns CEO, Vitalis Vosylius CTO), largest UK robotics seed ever
→ Why simulation-only pre-training (zero real-world data, zero human demos) + one-shot learning matters: avoids cost/privacy of human data collection at scale, “physical vibe coding” (show once, execute) across varied tasks (packing, tidying, humanoids)

Tesla starts production Optimus on high-volume line, ramp slow
→ Fremont factory (Model S/X line decommissioned in 46 days), first production-quality Optimus units rolling; 1M robots/year line infrastructure mostly in place
→ Why manufacturing milestone mattering more than quantity: confirms commercial-grade design intent, but 20K/week target at least 1+ years away, ramp “somewhat bumpy” (Musk’s words)

ANYbotics delivers ANYmal X into explosive atmospheres
→ Customer deliveries to Equinor, DSM-Firmenich; four-legged inspection robot built ATEX (Europe) + IECEx (global) certified for oil/gas/chemical where spark = explosion
→ Why intrinsic safety vs enclosures matters: robot needs exposed sensors + docking + leg power, can’t use heavy enclosures, so every core component (12 leg motors, battery, compute, electronics) individually flameproof-certified (4 years dev, 10+ patents)

Physical Intelligence in talks to raise $1B at ~$20B valuation
→ Bloomberg 27 Oct: Avenir, Lux Capital leading; nothing signed (valuation under negotiation)
→ Why rapid climb ($2.4B Nov 2024 → $5.6B Nov 2025 → $11.5B earlier 2026 → ~$20B now) signals foundation-model thesis gaining credibility: single π0 family across embodiments vs one-per-machine, opened Pi API beta Oct 13 for external fine-tuning

Meta builds a humanoid robot with China’s Fourier! 🤯

Meta has developed a humanoid robot prototype with Chinese robotics startup Fourier and is already working on a second version, according to The Information, as the Facebook parent steps up efforts to bring AI into everyday household tasks.

The prototype, internally named Max, weighs about 106 pounds and was built by the two companies in roughly three months. It stands around 1.5 metres tall, runs on Nvidia’s Jetson Thor chip, and features a display capable of mimicking facial expressions, built-in speakers, and five-fingered hands designed to handle small objects.

Meta’s robotics team travelled to China to work with Fourier, with vice president of product for robotics Li-Chen Miller spending almost two months at the startup’s Shanghai factory. Max was demonstrated to some Meta employees at an internal robotics summit in late August. Work on a second prototype, Max 2, began in September, with a redesigned torso, improved hands, a projector, and additional sensors to help it navigate stairs.

The strategic framing is that Meta sees humanoid robots as a potential physical extension of its Muse AI agent, with the longer-term ambition of machines that assist with household tasks. The company also strengthened its robotics capabilities in May by acquiring Assured Robot Intelligence, a startup building AI models that help robots understand and adapt to human behaviour.

Meta told The Information that its work with Fourier is focused on internal research and experimentation, and that it has no plans to sell the units.

Read more here!

Bedrock Robotics raises $600M as autonomous excavators move onto paid job sites! 🏗️

Bedrock Robotics has raised a $600M Series C at a $4.1B post-money valuation, co-led by CapitalG (Alphabet’s growth fund) and Emergence Capital. New investors US Innovative Technology Fund, Bessemer Venture Partners and Temasek joined existing backers 8VC, Georgian, Valor Equity Partners, Xora, Eclipse, Perry Creek Capital and NVIDIA. The two-year-old company has now raised nearly $1B in total. In February it closed a $270M Series B at $1.75B, so its valuation has more than doubled in about eight months.

The round follows what Bedrock calls a defining milestone. In mid-August, its fully autonomous excavators started doing paid work on live commercial job sites, beyond testing and pilots. Deployments include a water treatment facility in Nevada, an earthwork site run by Champion Site Prep in Texas, and a 1.2 million cubic yard civil sitework project with Zachry Construction. The newest partner is Haydon, an Arizona general contractor. Haydon plans to use Bedrock-equipped machines on advanced manufacturing plants, water and wastewater facilities, data centres and healthcare buildings.

Bedrock doesn’t build its own excavators. It makes an autonomy kit that retrofits onto the heavy equipment contractors already own, which keeps it out of the capital-heavy business of manufacturing machines. The team comes from autonomous driving. CEO Boris Sofman led Waymo’s self-driving trucks programme and earlier co-founded Anki, and CTO Kevin Peterson is also a Waymo veteran.

The case for autonomy here rests on labour supply. Bedrock says it takes five or more years to master excavator operation. Associated Builders and Contractors estimates that over 40% of construction workers will retire within the next five years. Bedrock says its customer base has doubled. It is expanding from the West and Central South into the Midwest and South Atlantic.

The new money will put more Bedrock-equipped machines on US sites and build out the field, commercial and operations teams that make autonomy reliable day to day.

Read more here!

Machenta raises $40M to teach robots a new task from a single demonstration! 🎓!

London-based Machenta came out of stealth on 27 October with a $40M seed round, co-led by Creandum and Union Square Ventures. Imperial College London says it is the largest seed round ever raised by a UK robotics startup. The company is building a robot foundation model that can watch one demonstration of a new physical task and then do it straight away, with no further training.

The founders have been working on this question for a while. Machenta is an Imperial spinout led by Professor Edward Johns (CEO), who holds the chair in Robot Learning in Imperial’s Department of Computing, and Vitalis Vosylius (CTO), his former PhD student. The work started six years ago in Johns’ Robot Learning Lab, built around one question: can a robot pick up a new skill as fast as a person does?

The training data stands out. Machenta’s model is pre-trained entirely in simulation, in computer-generated environments, with no real-world data or human demonstration datasets. Most robot learning today needs many repeated demonstrations per task, often collected over several days, and a developer has to update the model each time. Training in simulation avoids the cost and privacy problems of collecting human data at scale. Johns calls the end-user experience “physical vibe coding”: show the robot once, and it gets on with the task.

According to Imperial, a single model has already learned varied tasks such as packing and tidying, on different robots including humanoids. The company is open about what’s still hard. More precise, complex manipulation is still difficult, and Machenta has no near-term plans to sell the technology. For now it’s a research-first company with 11 people.

The money goes into scaling the model with more compute and more simulated data, and into growing the team at a new office and lab in King’s Cross. It’s also a strong signal for European robotics, with a UK university lab landing a seed round of this size.

Read more here!

Learn more

Sponsor of the newsletter

Tesla starts building production-quality Optimus robots in Fremont! 🦿

Tesla has started making its Optimus humanoid robot on high-volume lines at its Fremont factory in California. The space was freed up when Tesla took down the Model S and Model X assembly line, a job it finished in 46 days. Tesla’s Q3 2026 shareholder letter says the robots coming off this line are the first production-quality units it has ever made, and that the current manufacturing process is suitable for a commercial product, while hardware and process will keep evolving.

On the earnings call, Elon Musk called this “early production”. He said most of the equipment for a line designed for roughly one million robots a year is already in place in Fremont. He also warned the ramp will be slow and “somewhat bumpy” because the product is so new. Reaching Tesla’s target of 20,000 robots a week will take at least a year, “maybe a little longer”. A second, larger Optimus line is being built in Texas.

The remaining engineering work centres on durability and dexterity. Musk said the design still needs tweaks to make a robot that is safe, reliable and long-lasting. It can’t wear out after a month. He said most of the remaining work is on the hand. He expects those issues to be solved around the end of 2027, which is when Tesla plans to start delivering Optimus to outside companies. On 30 October, Musk also replied “No” when asked whether Optimus orders would open in December.

Tesla still hasn’t shared some key numbers: no production rate, unit count, price, or which Optimus generation is on the line. The milestone landed in a tough quarter, with net income down 43% year on year to $787M as spending on AI, chips and Optimus pushed operating expenses up 39%.

Even so, this is the first time Tesla has said it is building Optimus to a commercial production standard. The next questions are ramp speed, the hand, and real work inside Tesla’s own factories.

Read more here!

ANYbotics starts delivering the first legged robot built for explosive atmospheres! 🧨

ANYbotics has started customer deliveries of ANYmal X, its four-legged inspection robot built to work in oil, gas and chemical facilities where a single spark could set off an explosion. The first units are going to Equinor and DSM-Firmenich. ANYbotics first unveiled ANYmal X in 2022, so this marks the move from announcement to robots working on customer sites.

In these zones, every device has to be “ex-proof”, from radios to flashlights. Each one needs independent certification showing it can’t ignite a flammable atmosphere, even when something fails. Fixed equipment usually gets there with heavy explosion-proof enclosures or by limiting electrical energy. A walking robot can’t use either approach, because it needs exposed sensors, docking ports and a lot of power in its legs. ANYbotics built ANYmal X to the ATEX standard in Europe and IECEx globally. Its compute and inspection electronics sit in sealed, nitrogen-filled chambers, and the 12 leg motors and the battery have flameproof casings. Every core component is certified individually. Development took four years and produced more than ten patents.

On site, the robot checks pumps, compressors and valves with thermal, visual, acoustic and gas sensors. It handles steep stairs and narrow walkways without any changes to the facility. That matters in brownfield plants, which were never designed with robots in mind.

ANYbotics has tested it for a long time. Since 2023, BP, Equinor, Shell, PETRONAS, Petrobras, BASF and SLB have run ANYmal X on hundreds of onshore and offshore missions through an Early Access Program. Across its wider platform, ANYbotics reports more than 50 customers in power, metals and materials, running over 120,000 autonomous inspections a month.

Having deployed quadrupeds in energy myself, I know the explosive zones are where operators have most needed robots and had the fewest options. The full system makes its public debut at ADIPEC 2026, shown by ANYbotics, SLB and Kuwait Oil Co.

Read more here!

Physical Intelligence is in talks to raise $1B+ at a ~$20B valuation! 🧠

Physical Intelligence, the San Francisco lab building general-purpose AI models for robots, is in talks to raise at least $1B at a valuation of about $20B, Bloomberg reported on 27 October, citing people familiar with the matter. Existing backers Avenir and Lux Capital are in talks to lead the round. Nothing is signed yet, so treat the figure as a valuation under negotiation, not a closed round.

The valuation has climbed fast. Physical Intelligence raised $400M at a $2.4B valuation in November 2024, then $600M at $5.6B in November 2025, led by CapitalG. Bloomberg reported an $11.5B valuation earlier this year. A $20B round would nearly double that again. Investors across its rounds include Jeff Bezos, Thrive Capital, Lux Capital, CapitalG and OpenAI.

The company was founded in 2024 by former Google DeepMind researchers and academics from Stanford and UC Berkeley, including CEO Karol Hausman, Sergey Levine, Chelsea Finn and Lachy Groom. Its thesis is a single foundation model that controls many kinds of robots, instead of one model per machine. Its π0 model family combines vision, language and action, and is trained on data from real robots doing real tasks.

According to Bloomberg’s sources, one development in particular raised investor interest. On 13 October, Physical Intelligence opened its latest models to outside developers through a beta API. Teams can upload their own data, fine-tune the models and run them on their own hardware. In the launch post, the company acknowledged that real-world robotic applications are not easy.

Bloomberg’s sources say Physical Intelligence doesn’t have revenue yet. That contrasts with Skild AI, valued at $14B in January, which reported passing $100M in annual run-rate revenue in September. The new round is a bet that a cross-embodiment robot model, offered as a platform, can become the default layer that many robot makers build on.

Read more here! (paywalled)

Apply here! 🦾

What’s inside (version for the busiest people)

NEURA opens €17M TUM RoboGym near Munich, launches 4NE1 Gen 4
→ 2,300m² facility combining research, robot training, industrial application; NEURA contributes €11M over 3-year partnership with TUM’s MIRMI
→ Why vertical integration of research + data generation + deployment matters: knowledge capture (human expertise in materials/assembly/processes) → humanoid training → real-world validation in single facility

Agile Robots partners with Audi on mobile humanoid parts sequencing
→ Proof-of-concept at NVIDIA GTC Berlin: mobile humanoids pick and sequence parts at Ingolstadt plant for final assembly
→ Why dual-arm + mobile base solves persistent blockers: human-height reach + dexterity for high-variant/semi-rigid parts, AI perception handles mixed/unsorted containers

Eka Robotics hits $1.45B valuation with vision-force-action model
→ $150M Series A (a16z, F-Prime leading) for MIT professor Agrawal + Google DeepMind’s Haarmoja’s force-sensitive gripper + deep-learning architecture
→ Why vision-force-action beats vision-language-action: force/touch close the millimeter problem vision alone cannot, gripper adapted for flexibility + force sensors enabling real-time adjustment

Dream Labs raises $90M, launches Dream API for hardware-agnostic robot intelligence
→ Publicly available interface turns camera feed + plain-English instructions into motor commands, zero-shot across hardware
→ Why video-pretraining → gripper-agnostic transfer matters: 5M industrial robots programmed for single tasks, Dream-1 trained on 1M+ hours human video predicts next action enabling task transfer

Reliable Robotics raises $115M at $1B valuation, targets autonomous cargo flights → ex-SpaceX engineers retrofitting Cessna 208B with Reliable Autonomy System (automated flight controls + navigation + planning)
→ Why retrofit > new aircraft: existing fleet + cost reduction + regulatory path clearer than novel designs

Robotiq launches 2FC adaptive gripper with embedded tactile intelligence
→ Production-ready gripper combining force-sensing + embedded AI, train-and-deploy same hardware
→ Why touch as blind spot matters: industry pours compute into vision/sim/data but manipulation depends on force/slip/pressure most robots cannot feel

NEURA opens a robot training centre with TUM and launches 4NE1 Gen 4! 🏋️

NEURA Robotics and the Technical University of Munich have opened the TUM RoboGym at LabCampus near Munich Airport, a facility of more than 2,300 square metres bringing together research, robot training, and industrial application. Built with TUM’s Munich Institute of Robotics and Machine Intelligence (MIRMI), the three-year initiative represents around €17 million in investment, with NEURA contributing roughly €11 million. To mark the opening, NEURA also launched 4NE1 Gen 4, the new generation of its humanoid — slimmer, lighter, and powered by Neuron AI.

The Gym has been running since September, with companies from more than ten industries involved including automotive, pharmaceuticals, medical technology, packaging, logistics, agriculture, and care. Tasks range from handling cardboard boxes and filling bottles to loading and unloading wooden components. Companies can test applications, identify training needs, and assess which robot systems suit them before committing. NEURA says capacity in Munich is already in strong demand.

Data generation runs on three tracks: real-world robot data from sensors capturing movements, contacts, and outcomes; human demonstrations captured through sensor-equipped data suits, motion capture, and teleoperation; and simulation through the Data Amplifier, which expands real experience with additional variations before validating on real robots.

“The TUM RoboGym robotics center is unique in Germany and Europe,” said TUM President Prof. Thomas Hofmann. “By collaborating with a robot manufacturer based in Germany, we are also sending a clear signal for Germany as an industrial location in geopolitically uncertain times.”

Read more here!

Agile Robots and Audi bring mobile humanoids into automotive logistics! 🚙

Agile Robots, Audi, Dassault Systèmes, and NVIDIA have announced a proof-of-concept at NVIDIA GTC Berlin to automate parts sequencing at Audi’s Ingolstadt plant using mobile humanoids.

The work is happening in Audi’s Production Lab, its innovation space for testing future production and logistics technologies, where Agile Robots’ mobile humanoids pick parts from material containers and place them in sequence for final assembly. Audi’s supply chain experts contribute the sequencing expertise and define operational requirements for real-world testing. Before robots are deployed, a virtual twin of the logistics environment is being built using Dassault Systèmes’ DELMIA application, with feasibility, timings, ergonomics, and safety tested on the 3DEXPERIENCE platform using Industry World Models and NVIDIA Omniverse libraries.

Two automation problems have been persistent blockers here: handling high-variant or semi-rigid parts, and reliably picking unsorted components from containers. Agile Robots’ mobile humanoids combine humanoid reach and dexterity with the robustness of a mobile base, working at human height so they fit into Audi’s existing environment. Dual-arm manipulation lets them handle large and flexible parts, while AI-powered perception identifies and grasps a wide variety of components even in mixed containers or varying positions.

Agile Robots, founded in Munich in 2018 by researchers from the German Aerospace Center, now has over 20,000 robotic solutions deployed worldwide and more than 3,500 employees.

Read more here!

Eka is worth $1.45B thanks to a physical AI model that has experts raving! 🍓

Eka Robotics has closed a $150 million Series A at a $1.45 billion post-money valuation, led by a16z and F-Prime. Founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, the company has built a manipulation architecture combining a force-sensitive gripper with a new kind of deep-learning model. “They are already exceeding superhuman performance for certain tasks, and so we think over the next few months there will be superhuman performance in many dexterous manipulation tasks,” said a16z general partner Raghu Raghuram.

The reaction from other roboticists is unusual. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, said it had a “wow factor” compared to other recent demos, noting the robot “was able to adapt to a large number of different objects without retraining” with an aspect of “humanness, creepy humanness that it would try over and over again if it failed the first time.” A roboticist at a frontier lab, speaking anonymously, compared seeing it to when robots first achieved reliable legged locomotion. Handed a raspberry, they described “a big, heavy robot rushing towards my arm and then slowing down just in time and very gently touching me.”

The performance numbers: roughly 1,500 objects up to 1 kg picked per hour, against about 600 for a human, with reliable pickup over 95% of the time. The arm moves fast enough to damage itself, which the frontier lab observer read as real confidence in the algorithm.

The architectural choice is the story. Rather than a vision-language-action model, Eka built a vision-force-action model. “Vision is not the primary way in which you close this last millimeter problem,” Agrawal said. “Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.” Critics like Rodney Brooks have argued generative robotics models will fail without representing force and touch, and Eka’s results suggest they had a point. The company designed a custom hand with more flexibility and more force sensors, then used a physics simulator to model the forces involved and built a reinforcement learning environment around it, generating thousands of hours of robot experience in minutes rather than relying on human demonstrations.

The positioning is bottom-up. “Most of the field has been focusing on the planning,” Haarnoja said, “so they approach from the top: high-level intelligence, pushing that down towards the physical. Whereas we think about it from the bottom up.” A general model might know plates go in the cupboard; Eka’s model picks the plate up without breaking it.

The proving ground is still ahead. Eka is running secretive pilots with industrial customers on pick-and-place and wire manipulation. “What’s going to be a big challenge is when they work with integrators on real assembly lines, and it needs to run without breaking for an entire year,” Hauser said, also questioning how far the techniques generalise beyond pick-and-place. The gripper hardware costs under $1,000 and is rated for 5 million duty cycles, against the 20 million set by standard industrial arms.

Read more here!

Learn more

Sponsor of the newsletter

Dream Labs raises $90M and launches the Dream API! 🧠

Dream Labs has raised $90 million in seed funding and launched the Dream API, a publicly available interface that turns a robot’s camera feed and plain-English instructions into motor commands, even for robots it has never seen before. B Capital and Kindred Ventures led the round, with A.Capital, Unusual Ventures, SV Angel, Wing VC, Scribble VC, and Adrian Aoun also participating.

The problem the San Francisco startup describes is a familiar one. Factories worldwide run a record 5 million industrial robots, yet nearly every one is programmed for a single task and can’t transfer what it has learned to other machines. Low-cost robots now sell for a few thousand dollars but ship with no built-in intelligence. Teaching one robot one task can take specialists months, and new hardware often means starting from zero.

Dream-1, the flagship model, is trained on video of people performing everyday tasks and predicts what will happen next to determine the robot’s resulting action. Because it learns from video rather than from a single machine, the company says its capabilities transfer across hardware — controlling any robot with grippers out of the box, zero-shot. “The hardware is ready. What’s been missing is a brain any robot can use,” said co-founder and CEO Joel Jang. “When you get into a rental car, you don’t relearn how to drive. You find the controls, and your past driving experience does the rest. Dream-1 approaches a new robot the same way, so a new robot can do useful work in minutes, not months.”

Developers can sign up, connect any robot with grippers, and run tasks through models hosted in data centres worldwide. The company also offers fine-tuning as a service: submit a small batch of demonstration data for unsupported hardware and get a tailored model back within hours. There’s also a browser-based playground where anyone can type commands to a robot in Dream Labs’ San Francisco office and watch it respond live and unedited.

Dream Labs supports robots costing roughly $2,000 to $10,000, with over 20 hardware partners getting a model refined for their specific machine. Dream-1 currently handles tabletop tasks like picking, placing, and folding laundry, with wheeled mobile robots and dexterous hands to follow. “Our customers buy our robots to build things, and the hardest part has always been getting them to do actual work,” said Ryan Chan, co-founder and CEO of MakerMods. “With Dream-1, our robot was up and running in a few minutes and attempting tasks we never programmed.” The funding goes primarily toward compute as the company expands what Dream-1 can handle.

Read more here!

Reliable Robotics raises $115M and tops a $1B valuation! ✈️

Reliable Robotics has raised $115 million in fresh capital, pushing its valuation past $1 billion. The financing extends its previously announced $160 million Series D to $275 million, and includes $90 million in equity plus a $25 million term loan from Silicon Valley Bank. Total equity raised now exceeds $380 million. Returning investor Nimble Partners led, joined by Gideon Capital, RockCreek, Geek Ventures, Gold Hill Capital, Sparkwave Capital, and Island Green Partners.

The Mountain View company isn’t building a new aircraft. Its bet is that existing planes can be retrofitted with autonomous flight systems to cut operating costs and address persistent safety problems. Founded by former SpaceX engineers Robert Rose and Juerg Frefel, Reliable is automating the entire flight — taxiing, takeoff, navigation, and landing.

At the centre is the Reliable Autonomy System, combining automated flight controls, advanced navigation, and flight planning, developed for the Cessna 208B Grand Caravan, a workhorse of regional cargo transport. It targets loss of control and controlled flight into terrain, two major causes of aircraft accidents, and the company is building its own radar to detect collisions with other aircraft and obstacles.

Certification, not funding, is the real hurdle. Unlike road vehicles, aircraft must meet strict federal requirements before routine commercial flights. CEO Robert Rose expects a first certification covering continuous autopilot operation with a pilot aboard within 18 to 24 months, with removing the pilot entirely taking at least three years. Those are company estimates, not regulatory approvals.

Reliable has contracts with multiple branches of the US military and was selected for the FAA’s Advanced Air Mobility Integration Pilot Program, with plans for autonomous cargo operations serving smaller and rural communities in New Mexico and Colorado. The new capital funds fleet expansion, more production capacity, and hiring, with new offices opening in Seattle and San Diego alongside a workforce already past 250 people.

Read more here!

Robotiq launches the 2FC physical AI gripper! 🤏

Robotiq has launched the 2FC, a production-ready adaptive gripper combining tactile sensing with embedded intelligence, designed so teams can train and deploy on the same hardware. Pre-orders open today, with shipping in Q1 2027.

The argument behind it is that physical AI has a blind spot. The industry is pouring money into vision, demonstrations, simulation, and data, but manipulation ultimately depends on touch. Most robots can’t feel what they grasp, so contact data like pressure, slip, force, and position goes uncaptured. Tactile sensing itself isn’t new — the hard part is turning raw sensor data into useful training data and real-time behaviour.

“We’ve spent decades teaching robots to see the world better. Now we’re starting to understand how much intelligence comes from what happens when they actually touch it,” said CEO Samuel Bouchard. “When we built our first tactile sensors 12 years ago, the problem wasn’t sensing. The problem was what to do with all that information. This time, the intelligence needed to make use of touch data is finally catching up.”

The 2FC builds on Robotiq’s 2F adaptive gripper, which has more than 30,000 units deployed worldwide, adding tactile sensing, compute at the wrist via the AI Hub, and Contact Core, the company’s physical AI software. It captures pressure, vibration, proprioception, force, position, and current, synchronised at millisecond level through a unified API, with a ROS 2 package and C++ SDK coming to Robotiq’s GitHub. Because compute sits in the gripper, it runs grasp skills locally and in real time: grasp quality assessment, incipient slip detection, object pose estimation, object centering, and advanced grip check — emitting labelled events rather than just raw signals.

The gripper is rated for 2 million cycles and is underactuated for both parallel and encompassing grasps. Existing 2F owners can add TSF tactile sensors and the AI Hub to upgrade rather than replacing their gripper. The 2FC comes in 85 mm and 140 mm stroke versions, with a launch webinar on October 28.

Read more here!

What’s inside (version for the busiest people)

Google DeepMind’s Gemini 4 Argon turns robot video into sim and control code
→ Model watches human cornhole demo, builds MuJoCo simulation, generates whole-body joint movements, Apollo 2 sinks 7 consecutive shots on hardware
→ Why video→simulation→hardware pipeline matters: weeks of manual modeling compressed to hours, reasoning as generalization mechanism for low-level control + semantic planning

Physical Intelligence opens Pi API beta for frontier models
→ Developers upload data, fine-tune on PI infrastructure, access models through hosted gateway → Why modular training loop (collect demos
→ Early users (Almond/Axol, Chef Robotics, Deft, Ultra, Weave, Yondu) validate zero-shot task transfer; cloud-hosted inference surfaces network challenges but manages the hard part (ML infrastructure, compute, pipelines)

NEURA launches Neuron AI, multimodal perception-language-action model
→ 5-input architecture (3D vision via point clouds, tactile force-torque, audio, robot state, language) unified in single stream with memory
→ Why adding senses VLA cannot see matters: touch guides grip when object out of view, audio reveals weld quality, proprioception tracks hidden parts

NeodraDynamics launches Neo1, hybrid hydraulic-electric humanoid for heavy industrial work
→ 13-foot class with 41 DOF (39 miniature digital hydraulic actuators + 2 electric forearms), 1,102 lb combined payload, ±0.0002" repeatability
→ Why hybrid actuation thesis: hydraulics deliver force density electric motors cannot match at scale, but add fluid pressure/sealing/heat/control complexity electric-only avoids

Nature Will releases Natura-1, 50B-parameter hand foundation model
→ Zero-shot instruction following on egocentric video (1M hours human hand footage), 100ms control loop enabling real-time visuomotor execution
→ Why hand-only model fits squeeze layer critics identify: frontier vs robot foundation (whole-body missing), hand-only leaves locomotion/balance/coordination unaddressed

Origin emerges from stealth with $20M for interior construction robots
→ OG-1 drywall finishing robot deployed with US drywall contractors, pilots expanding to Germany, Sweden, France, Saudi Arabia; 150 robots targeted by end 2027 → Why job-site readiness from start (not lab-to-field transfer) matters: 3x human speed, human-level quality beats moonshot humanoids

Google’s new model turns robot video and sketches into simulation and control code! 👾

Google DeepMind showed Gemini 4 Argon a video of a human playing cornhole, and Apptronik’s Apollo 2 humanoid went on to sink seven consecutive shots on real hardware. The model mapped the environment, calculated whole-body joint movements, and autonomously built the simulation code to practise in before anything touched the physical robot.

The workflow is the interesting part. Researchers showed Gemini the human cornhole video and prompted it to practise in MuJoCo. The model mapped the human arm swing and release onto the robot, located the board with its cameras, and built the simulation. It then widened the stance and adjusted the swing to clear the hip. On hardware it made further release adjustments based on where the bags actually landed. Google says the pipeline can cut weeks of manual simulation modelling down to hours.

That last step is worth flagging: the demonstrated workflow includes feedback from physical attempts, so the seven-shot streak doesn’t establish a first-attempt success rate. Google hasn’t given a total attempt count or a breakdown of the human engineering behind the experiment.

Russell Mendonca, a Google DeepMind robotics researcher who previously worked on Tesla’s Optimus AI team, said Argon had changed his expectations for language models in robotics. “Reasoning can be a powerful mechanism for generalization,” he wrote, extending the point to low-level control as well as semantic planning. The published examples still rely on robotics infrastructure though — the arm-control demo gives Argon a calibrated visual grid and predefined motion commands.

Other examples include Spot practising foot-pedal bin disposal in MuJoCo, Argon reconstructing simulation environments from camera recordings or sketches, and an iterative planning loop using the Gemini Omni world model where Argon writes a plan, inspects the generated outcome, and revises. Google also reports generating more than 4,500 simulated episodes for VLA fine-tuning, raising real-world task success from 56% to 75% and cutting completion time from 35 to 24 seconds.

One number needs care: the reported 87% on ERQA is a set of 400 multiple-choice questions about spatial reasoning and real-world scenarios. It’s a reasoning test, not an 87% robot task-completion rate.

Read more here!

Physical Intelligence opens its API beta, putting frontier models on every robot! 🔌

Physical Intelligence has opened a beta for the Pi API, making its latest robotic foundation models available for training and inference. The pitch addresses a specific gap: models like π0 and π0.5 already cut the data and effort needed to adapt to new tasks, environments, or robots, but actually adapting them requires ML infrastructure, compute, and complex data pipelines, each a hard problem on its own. With the API, developers upload data, launch fine-tuning on compute Physical Intelligence manages, and access the fine-tuned model through a hosted gateway for inference.

The company worked with companies deploying robots to validate how broadly the models adapt across tasks, hardware, and operating environments. One example is Actor, which deploys autonomous heavy machinery: its wheel loader scoops material and loads a dump truck during a real commercial deployment with no operator inside the machine. Other collaborators include AgRobotics, CloudChef, a dairy partner, Ultra, and Weave.

A common pattern emerged across those collaborations. The specific problems differ, an inside-out garment, an ingredient that won’t stay in a scoop, a manufacturing part landing in the wrong position, but the fix is always the same loop: collect demonstrations, train a model, test on the robot, use the failures to decide what data to collect next. The API exists to let developers run that loop themselves.

Early access went to a small group to test whether they could improve autonomy independently. Almond builds Axol, a dual-arm robot, with the goal of having Axol build itself. In the demo, Axol deburrs spacers used in its own assembly, running a model the team got after four days of iterating with the API. Other early users include Chef Robotics, Deft, Formic Robotics, Foundry, General Instinct, Omakase, robot.com, Ultra, and Yondu. Most of these applications, spanning Japan, Switzerland, and the US, ran on cloud-hosted inference, which surfaced some network connectivity challenges.

The strategic logic is a feedback loop of its own. Real-world robotic applications are hard enough that construction, surgery, or warehouse automation each requires a whole company. If every one of those has to solve physical intelligence from scratch, the problem becomes unmanageable. By opening the models, Physical Intelligence accelerates the ecosystem while learning where its current models fall short.

Read more here!

Learn more

Sponsor of the newsletter

NEURA unveils Neuron AI, their foundation model! 🧠

NEURA Robotics has introduced Neuron AI, a Perception-Language-Action (PLA) model that combines 3D vision, touch, audio, robot state, and language with memory. The argument behind it is quite obvious, humans don’t work with their eyes alone. They feel a screw catch, hear a latch click, and know a part is still there when they look away. Vision-led VLA models judge by what the camera sees and reason mostly on the current frame. Neuron AI does what a VLA does and adds the senses for what a robot cannot see.

The model takes five inputs. 3D vision reads the scene as space rather than a flat image, through point clouds and depth. Tactile brings force-torque and contact data from hands and grippers. Audio captures contact sounds and spoken instructions. Robot state covers joint positions, velocities, and end-effector pose. Language handles task prompts and physical concepts. All of it is encoded, aligned in one shared space over time, and conditioned into a single action stream, with memory tracking what the robot has already seen so a part hidden behind another is still a part.

The use cases are where the case gets concrete. A camera can’t tell whether a lock has opened, but sound and touch can. Touch guides the grip when an object is out of view. In welding, the sound of the process reveals the quality of the seam.

Neuron AI runs locally on the robot rather than in the cloud. That means decisions happen in real time with no round trip to a server, the system keeps working without a connection, and production data stays with the manufacturer, a point NEURA flags as mattering specifically for European customers. One model is trained across the whole NEURA robot family and its hands and grippers, deployed on every embodiment including 4NE1 and MAiRA.

Training draws on three sources. Real-world robot data comes from robots performing tasks in the NEURA Gyms and real environments, generating synchronised vision, touch, force, audio, and movement data. Simulation-amplified data expands that through the company’s Data Amplifier and simulation stack. Human demonstration data comes from people wearing sensor-equipped data suits. Data partnerships with Schaeffler and Bosch complement the Gym training.

Read more here!

China unveils a 13-foot hydraulic humanoid that lifts over 1,100 pounds! 🤯

NeodraDynamics Robotics has introduced Neo1, a 13-foot-class humanoid built for industrial workloads beyond what conventional electric machines handle. The robot pairs hydraulic actuators with electric drives, and was unveiled alongside JD Industrial at the 2026 China International Industry Fair.

Neo1 has 41 active degrees of freedom. Thirty-nine joints use miniature digital hydraulic actuators developed in-house, while electric drives handle the two rotational joints in the forearms. Each hydraulic actuator weighs roughly 2.2 pounds and reportedly generates continuous thrust measured in tons, supporting a combined two-arm payload of 1,102 pounds. Despite the focus on raw force, the company claims repeatable positioning accuracy of ±0.0002 inches, which would make it viable for assembly work requiring controlled movement under heavy load.

Hydraulics give high force output in compact packages, but bring engineering headaches around fluid pressure, sealing, heat, and control complexity. Neo1’s hybrid approach uses electric servo control to balance load capacity against positioning accuracy. Most humanoids run entirely on electric motors, which offer precise control but struggle with the force demands of heavy industrial work. NeodraDynamics is betting hybrid actuation widens the range of tasks a humanoid platform can take on.

The intended applications sit well outside normal factory floors: heavy-component assembly and cargo movement at ports, nuclear fuel handling and equipment maintenance using radiation-resistant hydraulic modules, subsea maintenance under high water pressure, and demolition or debris removal in mines and disaster zones.

Those are intended capabilities rather than confirmed deployments, and the published specifications don’t establish how the complete robot performs under sustained industrial use. Field testing would need to show endurance and reliability outside controlled demonstrations. The company was founded by technical experts with backgrounds at Tsinghua University. Founder Yang Tao said the goal goes beyond copying human anatomy, robots should exceed human physical capability and do work people cannot safely undertake.

Read more here!

Nature Will releases Natura-1, a 50-billion-parameter foundation model for robot hands! 🪬

A physical AI lab called Nature Will has announced Natura-1, a 50-billion-parameter mixture-of-experts model built for dexterous hand control. It was trained jointly on language, images, video, and action, pre-trained on over one million hours of egocentric human video — footage shot from a person’s point of view while they use their hands — and runs a 100-millisecond control loop. The pitch is zero-shot instruction following: tell a robot what you want in plain language and it should do it, including with objects it has never seen.

“The human world offers an extraordinary range of things to do with a pair of hands,” said CEO Ning Ding. CTO Haozhan Li framed the data bet: “Hands already work with all the tools humans have developed, and are closely aligned with actions recorded in human experience.”

The release lands in the middle of an argument that’s been building in robotics for months. General Robotics CEO Ashish Kapoor describes three layers of robot intelligence: giant frontier models at the bottom, robot foundation models in the middle, post-training for specific jobs on top. He thinks the middle layer is in trouble.

The sums at stake are large. Figure has raised about $1.9 billion and signed a deal with Nscale for up to 100,000 Nvidia Vera Rubin GPUs, an initial $3.5 billion commitment with intent to scale past $6 billion, all to train Helix. Skild AI raised $1.4 billion for Skild Brain. Physical Intelligence has raised hundreds of millions for its π models.

Some caveats worth keeping. Natura-1 is a robot foundation model, which puts it squarely in the layer Kapoor thinks is getting squeezed. It’s also a model for hands only, with nothing said about locomotion, balance, or whole-body control. And the 100-millisecond control loop, while fast for a model this size, refines decisions only 10 times a second — Figure’s System 1 runs visuomotor control at roughly 200 Hz. Techniques like VLASH from MIT, Caltech, Nvidia, UC San Diego, and Tsinghua researchers could close that gap through prediction, the same way human cerebellums compensate for 100ms signal delays.

Read more here!

Origin emerges from stealth with $20M to build interior construction robots! 🔨

Origin, a physical AI company founded in early 2025 by Yogesh Ghaturle, Ashish Daga, and Tushar Makkar, has come out of stealth with $20 million in total funding co-led by Accel and Zacua Ventures. Unilazer Ventures, Schema Ventures, and Z21 also participated, alongside angel investors from Anduril, Skild AI, 1X, and Tesla.

The money goes toward deploying and building out the fleet of OG-1, Origin’s flagship robot, and supporting pilots with customers in the US and Europe, including two of the largest US drywall finishing contractors. More pilots are planned in Germany, Sweden, France, and Saudi Arabia. The company also wants to extend the platform to taping, drywall installation, drilling, and conduit installation, and is targeting 150 robots deployed by the end of 2027. Origin says the robots work at three times human speed with human-level quality, and OG-1 is being deployed in commercial spaces, multi-family apartments, data centres, hotels, and warehouses.

The market numbers back the thesis. A 2026 survey found worker shortages delayed projects for 42% of responding US construction firms. Zacua’s Construction Robotics Report 2026 found the sector drew $1.36 billion in venture funding in the first three quarters of 2025, up 125% from $612 million across all of 2024 — while robots still make up under 0.03% of global construction spend.

“Too many construction robots perform well in a lab and fail within a week on a live job site,” said Vivin Hegde, Founding Partner at Zacua Ventures. “Origin built for the job site from the start.”

Read more here!

What’s inside (version for the busiest people)

XPENG launches Iron humanoid H2 2027, tests in stores and HQ
→ Chinese EV maker selling general-purpose robot after mass production 2026, store assistants and office deployments starting early 2027
→ Why in-house Turing chip (matching NVIDIA Thor performance) matters: supply chain sovereignty + software control + cost optimization vs locked-in Nvidia dependency

Minerva Humanoids emerges from stealth at $10M
→ Germany-based Roger humanoid targets dangerous work (bomb/mine disposal, CASEVAC, offshore chemical units) with Western supply chain (Stuttgart hardware, US software/AI)
→ Why shared autonomy (VR teleoperation + onboard AI for balance/precision) beats full autonomy claims: humans decide, robot executes safely

Ultra raises $50M Series A for warehouse RaaS
→ Brooklyn-based startup scaling non-humanoid robots for goods packing at third-party logistics sites; 500K+ orders packed across US warehouses
→ Why RaaS model (integration fee + monthly support) de-risks customer CapEx vs ownership; strong demand enables price increases

Boston Dynamics names Rohit Prasad as CEO
→ Amazon veteran (12 years building Alexa, Nova foundation models) + Raytheon experience turning AI breakthroughs into commercial products
→ Why appointment frames around industrialization: productizing intelligent systems for real-world safety, productivity, efficiency vs research roadmaps

Microagi hits $100M ARR six months post-launch
→ Shift platform for collecting human activity data (paired with Atlas) enabling robot post-training self-improvement, model adaptation
→ Animesh Garg hired as CRO (ex-Usable Robots) covering post-training, self-improvement across customer deployments; goal: 1M robot deployments via data loops + compute scale

Mujin raises $224M Series D, SK Square leads
→ Japanese physical AI company scaling with major South Korean backing
→ Why geography signaxls: Asia’s robotics funding concentration deepens (China dominance + Japanese/Korean challengers) vs fragmented European/US capital

XPENG to launch its humanoid in the second half of 2027!

Chinese EV maker Xpeng plans to start selling its humanoid robot Iron in the second half of 2027, after first testing it in the company’s own stores and offices. “By early next year, you’ll see robots working alongside sales agents as assistants, or maybe when you visit our headquarters in Guangzhou, a robot receptionist will greet you,” vice chairman and president Brian Gu told Forbes Asia at the Forbes Global CEO Conference in Singapore. The general-purpose robot is scheduled for mass production this year.

Xpeng was cofounded in 2014 by billionaire chairman He Xiaopeng and first moved into robotics a decade ago, strengthening that effort by acquiring a robotics startup now called Dogotix in 2021. Iron runs on Xpeng’s in-house Turing chip, the latest version of which Gu says closely matches the performance of Nvidia’s Thor chip used in autonomous vehicles and humanoids.

The chip story goes back six years. Xpeng was using Nvidia chips for its EVs, but they didn’t fit well with its software stack. “If we want to be a leader in physical AI, we need to have our own physical core, so we can control the cost, control software compatibility and supply chain stability,” Gu said. Turing chips are now integrated across all Xpeng products and are used by partners including Volkswagen, which has worked with Xpeng since 2023 on vehicle platforms and licensed its electrical and electronic architecture for VW cars built in China.

The financial backdrop is tough. Xpeng has yet to turn a full-year profit, and in the first half of 2026 its loss widened to 3 billion yuan ($449 million) from 1 billion a year earlier. Revenue fell 4% to 32.8 billion yuan, though technology licensing and services jumped 67% to 4.7 billion yuan. Overseas sales made up 25% of revenue, about half of that in Europe, and the company wants to push that to 50% within five years while expanding into Mexico and Japan.

Xpeng is also aiming to deploy autonomous driving in Europe next year and is testing robotaxis in Guangzhou. Gu sees the two sides of the business feeding each other across manufacturing, supply chain, and especially AI. As he put it: “Cars are just robots on wheels.”

Read more here!

Minerva Humanoids emerges from stealth with $10M! ☣️

Germany-based Minerva Humanoids has come out of stealth with $10 million in pre-seed funding led by General Catalyst, with Long Journey Ventures, Credo Ventures, and angels including Hugging Face’s Thomas Wolf, OpenAI’s Laura Modiano, Lukas M. Ziegler also joining. The company’s robot, called Roger, is designed to replace humans in “the world’s most dangerous work,” CEO Sandor Felber told Tectonic: bomb disposal, mine and explosive clearance, and frontline CASEVAC.

“China right now is becoming the center of gravity of robots,” said GC Partner Robin Dechant, who led the deal. “If we don’t build our own robotics champions, we will just depend on China or other countries, which we think don’t want to… We need a sovereign robot.” That’s Minerva’s core pitch: a humanoid built on an entirely Western supply chain, with software and AI developed in the US and hardware engineered and assembled in Stuttgart. Felber says they use no single-source Chinese parts.

Roger is aimed at oil and gas refineries, offshore chemical units, biological hazards, labs, and bomb and drone disposal, anywhere a human is put in what Felber calls a moral dilemma to get work done. The company built a fully mechanical prototype earlier this year and has now completed a fully functional electromagnetic one, with three prototypes in total named Roger 1, 2, and 3. It has worked with the FBI and NATO DIANA on testing, and will soon pilot with the Ukrainian MoD.

The approach differs from most humanoid companies: Minerva isn’t chasing full autonomy. Its software, Minerva Intelligence, enables what Felber calls shared autonomy. Roger is operated remotely by a human wearing a VR set, with haptic gloves coming, while AI informed by onboard sensors corrects for weight, balance, and precise movement that are hard to judge from a distance.

The use cases he describes are concrete. A downed drone in a populated area like Kyiv or Leipzig Airport, where the robot approaches, analyses whether unused explosive is on board, and defuses it. Mine detection and clearance in a tight area. Two robots carrying a wounded operator on a stretcher out of the kill zone without exposing medics. On cost, Felber was blunt: Western-sourced humanoids cost more than consumer-grade ones, “however, they still cost a lot less than human life.”

The funding goes toward finalising the Roger prototype and building out semi-autonomous capabilities, with another round planned soon to start scaling production next year.

Read more here!

Learn more

Sponsor of the newsletter

Ultra raises $50M Series A to scale its RaaS warehouse business! 📦

Brooklyn-based Ultra has raised a $50 million Series A led by Framework Ventures with participation from Y Combinator. Combined with an earlier $12 million seed led by Y Combinator and Next View, that brings total funding to $62 million. The company builds and installs robots that pack goods for transit at third-party logistics sites, and says its machines have packed more than 500,000 orders for shipping across US warehouses.

The robots-as-a-service model lets clients skip large upfront capital outlays. Customers pay an integration fee for installation, then an ongoing monthly fee covering hardware and software support. CEO and co-founder Jon Miller Schwartz says demand has been strong enough that Ultra has been able to raise prices.

Alongside the raise, Ultra is deepening its partnership with Physical Intelligence. The split is body and brains: Ultra builds the robots, Physical Intelligence supplies the AI that lets them learn and improve. In return, the AI firm founded by a team out of Google DeepMind, gets real-world data to refine its models.

Schwartz’s argument is that humanoid robots grab attention but struggle to actually deploy, while other form factors are doing the real work. Ultra is entering a crowded field that includes France’s Exotec, Amazon, and Travis Kalanick’s logistics venture Atoms. Its bet is that non-humanoid robots paired with outside AI will win the warehouse.

Read more here!

Boston Dynamics appoints Rohit Prasad as CEO! 👔

Boston Dynamics has named Rohit Prasad as Chief Executive Officer, effective October 7. Prasad comes from Amazon, where he served as Senior Vice President and Head Scientist for Alexa and Artificial General Intelligence. Over 12 years there he was instrumental in building Alexa from its earliest days into a consumer AI service used in hundreds of millions of households, and more recently led development of the Amazon Nova foundation model family. Before Amazon he spent nearly 14 years at Raytheon BBN Technologies leading machine-learning research and its application for US government and commercial use.

The appointment is framed around commercialisation. Boston Dynamics says Prasad brings decades of experience turning breakthrough AI into category-defining products, and that his role will be productising intelligent robotic systems that improve safety, productivity, and operational efficiency in real-world industrial and commercial environments. He has been based in the Boston area for over 25 years, with ties across the region’s robotics and AI ecosystem that the company expects will help with hiring and technology partnerships.

Subject to the relevant approval process, Prasad is also expected to join the company’s Board of Directors. The move comes alongside a busy stretch for Boston Dynamics, which recently opened its Robotics Metaplant Application Center at Hyundai’s Georgia plant and unveiled a new four-finger hand for Atlas. Its portfolio currently spans Spot for industrial inspection, Stretch for box moving in logistics and retail, and Atlas, the electric humanoid still in development.

Read more here!

Microagi says it reached a $100 million revenue run rate in six months! 📈

Microagi says it has hit a $100 million annualized revenue run rate six months after launching Shift, its platform for collecting human activity data for robot training.

The graphic labels it “ARR,” but the footnote describes it as an annualized run rate, a revenue pace expressed over a year, not $100 million actually earned during the six months. The graphic also ranks Microagi as the fastest company to reach the milestone, without enough methodology to show the comparisons use equivalent measures.

Skild AI, which appears in Microagi’s chart, announced passing $100 million ARR ten months after its first commercial deployment.

The announcement follows Microagi’s hiring of Animesh Garg as chief research officer alongside the Usable Robots team. His remit covers post-training and self-improvement, adapting robot models to perform specific jobs reliably. The company pairs Shift’s human activity data with Atlas, its industrial robotics deployment platform, spanning both the gathering of training material and putting robot intelligence to work on customers’ existing hardware.

Responding to the revenue post, Garg reiterated the goal of deploying one million robots, naming compounding data loops, top-tier talent, compute scale, and execution velocity as the four primitives for winning physical AI. What the revenue claim doesn’t establish is how many robots Microagi currently has working at customer sites, or how much of the run rate comes from those deployments.

Read more here!

Japan’s Mujin raises $224M Series D as SK Square bets on physical AI! 🇯🇵

Mujin has raised a Series D of ¥251.5M USD-equivalent, roughly $224 million, with SK Square’s overseas investment arm, tgc square, among the backers. Founded in 2011, the Japanese robotics firm builds Mujin OS, a general-purpose platform that controls and coordinates fleets of robots and automation equipment across warehouses and factories.

The strategic logic for SK Square is about moving beyond semiconductors into physical AI, including robot control and factory automation, with scope to link Mujin’s technology to SK Group’s semiconductor and logistics operations. The holding company named semiconductor packaging, logistics, and plant construction as key use cases, with the longer-term idea being that integrating logistics robots and automation into a single operating system could push toward genuinely autonomous fabs.

Mujin has now raised more than $503 million in total. Its late-2025 round, backed by NTT Group, the Qatar Investment Authority, and Salesforce Ventures, valued the company at over $1 billion. The new raise sits in the top 1% of Series D robotics rounds in Japan over the past 48 months, based on a sample of 80 deals.

An SK Square official said the company will “continue to identify new investment opportunities in the AI and semiconductor value chain,” pursuing new investments “according to a strategy of selection and focus.”

The signal is capital flowing toward physical AI as chipmakers chase automated production. For SK Square, backing Mujin is a bet that robot-run factories, not just the chips inside them, are where the next layer of value sits.

Read more here!

Check out Encord's event!

What’s inside (version for the busiest people)

RobCo becomes Munich unicorn at $1B valuation
→ Doubled valuation in 9 months with Sequoia, Lightspeed, Greenfield, Kindred leading; Cherry Ventures and European Tech Collective joining
→ Why modular RaaS arms pivot to Alfie signals ambition beyond rentals: autonomous perception, planning, execution on unstructured industrial tasks

Dyna Robotics launches Dyna 2.1 for full-shift autonomous work
→ Semi-humanoid upper torso on wheeled base completing commercial laundry workflows without intervention (wash, dry, fold, stack, delivery)
→ What Mean Time Between Interventions metric reveals: optimization for real-world shift reliability vs task completion speed

Runway releases Praxis-1, world action model from video pretraining
→ Open-weight model trained on abundance of third-party video (not teleoperation demos), simulates robot policies with 0.95 real-world correlation
→ Why video-scale ceiling vs teleoperation data bottleneck matters: policy improves with more footage not more hardware

Boston Dynamics releases 4-finger hand for Atlas
→ 13 DOF (4 thumb, 3 per finger), pressure-based tactile sensing on fingertips/palm enabling grip-strength adaptation to shape, material, contact conditions
→ Why industrial requirements shaped design: complexity vs power/cost/durability/maintenance tradeoff; sim-to-real trained with hardware friction, motor torque variance, disturbance randomization

Figure retires F.02 fleet by jumping into molten steel
→ Robots trained to autonomously leap into 75-ton electric arc furnace at Finnish foundry, avoiding disassembly delays that would derail F.04 launch
→ Why custom actuators and proprietary IP made traditional recycling too labor-intensive vs destructive disposal at scale

ANYbotics launches Shift, fleet orchestration for inspections
→ Operations engine connecting plant models, robot fleets, inspection data, maintenance systems into continuous workflows
→ What multi-modal sensor fusion (thermal, acoustic, visual, gas) + contextual operating conditions reveal: anomalies only count when context says so, not raw sensor noise

Munich robotics startup RobCo becomes a unicorn! 🦄

RobCo is now valued at more than $1 billion (around €855 million), doubling its valuation in just nine months. Fresh capital is flowing in as part of the transaction, and long-serving employees are being allowed to sell part of their shares. Existing backers Sequoia, Lightspeed, Greenfield, Kindred, Lingotto, and Promus Ventures took part, with Cherry Ventures and the European Tech Collective joining as new investors. RobCo didn’t disclose the total raised. In February it closed roughly €85 million in a Series C led by Lightspeed and Lingotto.

Founded in Munich in 2020 by Hölzl, Paul Maroldt, and Constantin Dresel, RobCo made its name with modular robot arms that snap together like building blocks for different tasks, sold through a robotics-as-a-service model so customers rent rather than buy. The next phase centres on Alfie, a robot designed for autonomous industrial use: it’s meant to perceive its surroundings, work out the steps required, and carry them out on its own, taking on unstructured and safety-critical tasks conventional industrial robots haven’t been able to automate. Alfie will be officially unveiled in March 2027 at RobCoN in Munich, the company’s first summit.

The US is now RobCo’s fastest-growing market, with customers in more than a dozen states, production and assembly in Austin, and a lab in San Francisco. Hölzl has moved to the US himself to drive the expansion. “RobCo is working toward a future where AI doesn’t just generate content but acts independently in the physical world,” said Sequoia partner Luciana Lixandru.

RobCo is part of a broader European wave. Neura Robotics raised up to $1.4 billion in June from Tether, Nvidia, and Amazon, the largest round in German history. London’s Humanoid became a unicorn this summer with a €133 million Series A at €1.1 billion. Rome-based security startup Exein is valued at $1.7 billion after a $270 million round, and Agile Robots brought Google DeepMind on as an AI partner. The caveat is humanoids, where China holds roughly 90% of the global market. RobCo, by contrast, is aiming its robots squarely at industry and the factory floor.

Read more here!

Dyna Robotics launchesa semi-humanoid! 🦾

Dyna Robotics has launched the Dyna 2.1 physical agent, a semi-humanoid robot that can autonomously complete entire physical workflows, like a commercial laundry shift, without human intervention. It’s being deployed in hotels, laundromats, and restaurants.

The hardware is an upper torso with two arms, compatible with either parallel-jaw grippers or dexterous hands, mounted on four steerable wheels so it can move around without risk of toppling. To handle hours-long shifts it runs a vision-language orchestrator that reasons through the workflow alongside a whole-body controller that coordinates driving, reaching, bending, and lifting into one fluid motion. Underneath sits Dyna 2, the company’s world action model trained on a million hours of human and robot data.

“Our goal is to make general purpose robots commercially viable for real-world scenarios,” said co-founder and CEO Lindon Gao. “Our customers need robots that complete an entire workflow for a full shift, without human babysitters, not just specific tasks.”

In a commercial laundry, that means loading the washer and dryer and turning the machines on, bending to reach the back of a dryer to unload bulk towels, snapping towels flat and folding them into stacks sorted by size, then crouching to place finished stacks on bottom shelves or reaching up for the top ones. It also corrects its own mistakes: picking up a dropped towel, or abandoning a bad grasp to try again, without asking for help.

The metric Dyna optimises for is Mean Time Between Interventions, how long the machine runs before a human has to step in. The robot captures execution data as it works, from the reasoner’s decision traces to the controller’s joint loads, feeding it back into the system to steadily increase minutes of work per intervention. Beyond laundry, the company lists food preparation and data center management as workflows the agent can handle end to end.

Read more here!

Learn more

Sponsor of the newsletter

Runway introduces a world action model for robotics! 🎥

Runway has introduced Praxis-1, an open-weight world action model that turns the company’s video pretraining into robot control. It’s built on the same large-scale video pretraining behind Runway’s general world models.

Praxis-1 learns mostly from third-person video, built on the same large-scale pretraining behind Runway’s world models, so it already understands how objects behave and how tasks unfold. Also, performance improves as we scale video, so its ceiling is set by how much video it can learn from, not how many robot demonstrations exist.

The bet is on abundance. People upload more footage of everyday life each day than any robot lab could capture through teleoperated demonstrations. Runway says simulating robot policies inside its world model predicts real-world results with 0.95 correlation, which it claims compares favourably with more expensive 3D reconstruction techniques. The company has also extended its video pretraining into interactive real-time models like Solaris and GWM Worlds 2, teaching them accurate physics, how hands move, and what a task looks like partway through.

Praxis-1 has been trained on manipulation tasks ranging from straightforward pick-and-place, like lifting soda cans, to harder work with deformable objects, like packing gift bags. Early partners including Noble Machines, Standard Bots, and Ultra are running it on their own hardware.

When Runway releases Praxis-1 publicly in the coming months, it will ship with open weights rather than as a closed model. Founded in 2018 and based in Brooklyn, Runway is better known for generative video models like Gen-4.5 and Act-Two, and already offers GWM-1, a general world model.

Read more here!

Boston Dynamics gives Atlas a new 4-finger hand! 🪬

Boston Dynamics has released a four-finger robot hand built to help Atlas perform continuous, complex work. The hand has 13 degrees of freedom across four fingers, four in the thumb and three in each of the others. Pressure-based tactile sensors on the fingertips and palm detect subtle changes in contact force, letting the hand adjust finger movement and grip strength based on an object’s shape, material, and contact conditions.

In the demonstration video, Atlas inserted a drill bit into a drill, pressed the trigger, and drilled a hole in wood. It also tightened a small nut by rotating it between its fingers, spun a thin drumstick, and manipulated two golf balls in one hand simultaneously. The four-finger design came out of industrial requirements, weighing design complexity, power consumption, cost efficiency, durability, and maintenance. The hand is roughly human-sized with modular components and a simplified structure.

Training used a sim-to-real approach, with the robot learning in virtual environments before the results transferred to physical hardware. The simulations incorporated hardware characteristics and friction while varying motor torque, object shapes, and external disturbances for reinforcement learning.

The company recently opened the Robotics Metaplant Application Center, which replicates Hyundai Motor Group’s manufacturing environment. That center will collect Atlas’ manufacturing data and validate its learning through a loop of data collection, training, field application, feedback, and retraining.

Read more here!

Figure decommissions the F.02 fleet by having the robots jump into molten steel! 🔥

Figure has retired its F.02 fleet, and the method is unusual enough to be the story. The robots were trained to autonomously jump into a vat of molten steel at a foundry in Finland.

F.02 marked a lot of firsts for the company: the first deployment at BMW, the birth of Helix, the first robot doing household chores, and the first logistics deployment. But as the F.03 fleet grows, maintaining F.02 stopped making sense. The disposal problem was real, though. The robots are full of custom-built actuators and other IP, and disassembling each one by hand would have taken enough technical staff time to delay the F.04 launch.

So Figure asked the internet what to do. The answer came from Arnold Schwarzenegger, who told them to melt the robots. Figure reached out, he wanted to be involved, and things moved from there. No metal foundry in the US or Mexico would let them drop a robot full of lithium-ion batteries into expensive equipment. The company even contacted former MythBusters to help with the search. The only place in the world willing to take it on was a foundry in Imatra, Finland.

Training happened at Figure’s San Jose campus, with large airbags set up so robots could practise jumping off the second floor of a building. Using stunt artist movements as reference, the team trained a new AI model in simulation that would let the robots jump accurately into a bucket of molten steel at a foundry they had never visited.

At the foundry, three massive graphite electrodes power a 75-ton electric arc furnace. The team had 24 hours and six melts to get it right, and each melt had only a 20-minute window before the steel cooled and formed a crust. Despite the heat and electromagnetic fields, the robots ran their AI policies without issue, even as camera equipment and other electronics failed in the environment.

The resulting metal bars went back to the US and are being machined into a limited series of artifacts commemorating F.02. Most of the fleet is gone, with a few left in storage at HQ.

Read more here!

ANYbotics launches Shift to transform robot inspection operations! 🔎

ANYbotics has launched ANYbotics Shift, the operational engine behind its autonomous inspection solutions. Shift connects a facility’s plant model, robot fleet, inspection intelligence data, and maintenance systems into one continuous workflow, letting teams plan and run multi-robot missions while organising captured field data around existing asset management infrastructure.

ANYbotics robots collect hundreds of thousands of inspection data points monthly, creating a comparable inspection baseline across every asset over time, so reliability engineers can see an asset’s degradation curve and intervene before a minor anomaly becomes a multi-million-dollar shutdown.

As sites scale from single robots to full fleets, the challenge shifts from what one machine can inspect to how whole fleets operate. Shift is built around four modules. Fleet schedules, monitors, and controls the robots from one place, with missions triggerable directly from a distributed control system.

Insight combines thermal, acoustic, visual, and gas readings with operating conditions like motor speed and load, so a reading only counts as an anomaly when the context says it should. Connect turns confirmed anomalies into work orders inside core software, integrating with Yokogawa, SAP, IBM Maximo, GE Vernova APM, Cognite, and Oracle. Maps provides a live digital twin using the existing asset hierarchy. Shift is already deployed across more than 200 robot deployments running hundreds of thousands of inspections monthly.

Read more here!

Check out Encord's event!

What’s inside (version for the busiest people)

AMD acquires World Labs, Fei-Fei Li joins as Chief Scientist
→ Less than a month after Atlas launch, World Labs consolidates research team within AMD to scale hardware/software ecosystem
→ Why joining backer company matters: GPU optimization partnership becomes vertical integration, open ecosystem positioning vs Nvidia's proprietary approach

Skild AI uses self-play to teach humanoids superhuman soccer
→ 140 years simulated play, taught itself dribbling/shielding/tackling without hand-crafted rewards
→ Why self-play (AlphaGo precedent) matters more than imitation learning alone: caps robots at human capability

Cognex acquires RealSense for $500M, expanding vision portfolio
→ 1M+ units shipped, 20K active developers, 3D depth sensing for autonomous navigation vs Cognex’s fixed-arm inspection
→ What $600M→$1.6B market growth (2030) demonstrates: robotics perception becoming standalone business

Vesoma emerges from stealth with humanoid Kyle
→ Why learning through interaction thesis remains unproven: locomotion ≠ contact-rich manipulation or shift reliability
→ Early 2027 Vesoma 1 unveiling targeted, no customer delivery date; design renders labeled as visualizations, not photos

Niantic Spatial launches Places Library for embodied AI
→ 100 high-fidelity 3D environments (photorealistic Gaussian splats + aligned collision meshes) covering industrial, logistics, commercial, residential
→ What single-source reconstruction (no manual cross-registration) prevents: drift between visual scene and physics layer

World Labs is joining AMD! 🤝

World Labs has signed a definitive agreement to join AMD, less than a month after unveiling its Atlas world model. The company says the research and technical breakthroughs it has made since founding in 2024 gave it a clear vision for AI in the spatial and physical world, and that accelerating into that future requires scaling its efforts, scaling its reach, and getting closer to the hardware.

The two companies started a deep technical partnership last year, beginning with model training and inference optimisation on AMD GPUs. As the teams worked together, World Labs says it became a natural fit to bring together AMD’s software and hardware ecosystem with its own foundation models and applications.

Fei-Fei Li will join AMD as Executive Vice President and Chief Scientist, working directly with CEO Lisa Su. Justin Johnson and Ben Mildenhall will work with Li to continue leading the World Labs team as it joins AMD to form what the companies describe as a world-leading frontier research organisation. The stated commitment is an end-to-end open AI ecosystem spanning hardware, software, platforms, and widely accessible open models.

The timing is notable. AMD was already among World Labs’ backers alongside Nvidia and Autodesk, part of the $1.2 billion the company had raised. Atlas launched at the start of September and was positioned squarely at robotics through its real-to-sim workflows, going head to head with Google’s Gemini Omni and video models from ByteDance and Google.

The transaction is expected to close by the end of 2026, subject to regulatory approvals and other customary closing conditions.

Read more here!

Skild AI is teaching humanoids to play soccer against itself! ⚽️

Skild AI has announced a post-training result showing that a strong base model like S1 can learn extremely dexterous and dynamic tasks by competing against itself in simulation. The reasoning behind it: S1-class models are pre-trained on human data, which caps them at human capability. Skild believes physical self-play is how robots exceed it.

Self-play predates language models as a route to superhuman performance. AlphaGo beat Lee Sedol in 2016 with the famously “inhuman” move 37, and AlphaGo Zero beat the original AlphaGo 100-0 just three days after starting self-play. The approach extended to StarCraft II with AlphaStar and Dota 2 with OpenAI Five, which beat world champions in 2019. Then self-play lost momentum as reinforcement learning from verifiable rewards turned out to be simpler and more immediately viable. Skild wants to bring it back, this time for physical AI.

The policy learned by playing against recent versions of itself, so every gain in capability produced a harder opponent. In its first few simulated months inside NVIDIA’s Isaac Sim it could barely walk. By what Skild calls college age, it could stand back up after falling. Over time it taught itself to dribble past defenders, shield the ball, and tackle opponents, none of which were hand-crafted rewards. They emerged because they helped it score.

After 140 years of simulated play, Skild transferred the policy onto a real robot and played a match against it. The company treats the successful sim-to-real transfer, combined with how complex the task is, as a strong signal for self-play as a path toward physical AGI. Soccer works as a testbed because it demands both physical and strategic ability, and existing robot soccer competitions still perform well below human level. But Skild says the approach is general and is already extending it to everyday robotics tasks.

Read more here!

Learn more

Sponsor of the newsletter

Cognex acquires RealSense for $500M to expand its vision platform! 👁️

Cognex has entered a definitive agreement to acquire RealSense for $500 million in cash, moving its machine-vision portfolio into 3D depth sensing, autonomous navigation, and robotic perception. RealSense brings more than 1 million units shipped, roughly 160 employees, over 70 patents, and a developer ecosystem of around 20,000 active users. Its platform is used in vision-guided fixed-arm robots, autonomous mobile robots, quadrupeds, and humanoids.

The two portfolios cover different problems. Cognex’s existing systems let manufacturers identify and inspect products, take measurements, and guide fixed-arm robots in controlled industrial settings, optimised for precision. RealSense’s technology helps robots recognise surroundings, calculate distances, avoid obstacles, and navigate changing environments, with higher-frame-rate depth sensing built for real-time motion, localisation, and collision avoidance.

Cognex puts the robotic perception market at roughly $600 million today, growing more than 25% annually to around $1.6 billion by 2030. That outlook rests mainly on established fixed-arm and autonomous mobile robot deployments, with quadrupeds and humanoids treated as potential longer-term upside rather than a requirement for the investment case.

RealSense is expected to generate $80 million to $90 million in 2026 revenue with growth above 50%. The deal is expected to close in the fourth quarter of 2026 and isn’t expected to materially affect Cognex’s 2026 results.

Read more here!

Vesoma emerges from stealth with a humanoid robot! 🇪🇺

The startup behind Munich’s boxy Kyle humanoid prototype now has a name: Vesoma. The company has come out publicly with more than 60 people across Munich and Limassol, a first product called Vesoma 1, and a bet that learning through interaction can carry humanoids past impressive demonstrations into reliable work. It began work in December 2025 and has fitted out more than 3,500 square metres of lab and office space across the two sites. First Vesoma 1 units are being built.

The leadership resolves a question left open by earlier coverage of Kyle. Martin Riedmiller, previously a research director and Control Team lead at Google DeepMind, is chief AI officer. Nikolai Ensslen, co-founder of motion-control and functional-safety specialist Synapticon, is co-founder and CEO. Peter Skoromnyi, who co-founded games company Easybrain, is chief strategy officer. Operations chief Jonas Reinecke came from precision-manufacturing company Daedalus, and head of design Marcelo Gutierrez previously worked on Tesla’s Optimus, Apptronik’s Apollo, and Cartwheel’s Yogi.

Vesoma says it designed, built, and walked its first humanoid prototypes within six months using a learned locomotion policy. That’s an engineering milestone rather than evidence the machines can do useful work across a shift. Kyle was described by Ensslen as an alpha development platform distinct from the first-generation product, with an early 2027 unveiling targeted. The new announcement doesn’t give a customer delivery date. Worth noting: the polished robot images in the media kit are labelled design visualisations, not photographs of a finished machine.

The company’s central argument is that demonstrations alone won’t deliver the reliability customers need, and that an agent should improve through its own interactions with the physical world, with hardware designed around that process. Locomotion is the first example; Vesoma says the harder work is extending it to contact-rich manipulation, coordinated whole-body tasks, and environments where mistakes damage things. No manipulation success rates or intervention figures have been published, so its criticism of imitation-led approaches reads as a thesis rather than a settled verdict.

Initial markets are manufacturing, logistics, and warehousing, and the company explicitly rules out weapons, military, and surveillance uses. The next real test is what Vesoma 1 does beyond walking: which tasks it learns, how much experience that takes, and how it holds up when conditions change.

Read more here!

Niantic Spatial introduces the Places Library for training embodied AI! 🗺️

Niantic Spatial has launched the Places Library, a curated catalog of 100 high-fidelity real-world 3D environments delivered as simulation-ready assets. The argument behind it: a policy that performs well in one clean, familiar setting can still fail the moment floor plans, lighting, or operating conditions change, because it never learned to handle that variation. The initial catalog covers four common deployment type, industrial environments, last-mile and logistics settings, commercial spaces, and residential environments.

The technical piece worth noting is how the assets are built. Each environment ships as a single USDZ package pairing a photorealistic Gaussian splat with an aligned collision mesh, both derived from the same reconstruction. The splat carries the visual detail, the mesh provides physical structure for interaction, and the asset arrives gravity-aligned, metric-scale, and collider-ready. Because both layers come from one source, the visual scene and the physics layer stay in register. There’s no manual cross-registration between an appearance model and a separate collision environment, and no drift between what a policy sees and the surfaces it can actually hit.

Depth is where Niantic says its reconstruction stands apart, and what customers value most. Photometric reconstruction alone struggles with white walls, glass, and other low-texture surfaces, exactly the places where a policy can least afford phantom obstacles. Their MVSAnywhere zero-shot multi-view depth fills those gaps, producing cleaner geometry at true metric scale from an ordinary 360 camera with no LiDAR. The workflow runs in NVIDIA Isaac Sim, Isaac Lab, and compatible OpenUSD simulators.

Access comes through Embodied AI plans, with the number of places scaling by plan, and each plan pairs library access with capture minutes teams can spend reconstructing their own sites. Selected scenes can also be relit through a Gaussian Splat Relighting beta, generating new lighting and weather variants from the same capture. Niantic is also running a voting system so teams can prioritise what gets captured next, factory, warehouse, construction, data center, and refinery environments are all on the list.

The bigger idea is using real places as the foundation for domain randomisation. A single capture shouldn’t have to represent a single training condition, and a growing catalog plus custom captures gives teams variation across geometry, materials, lighting, and layouts without building every environment from scratch.

Read more here!

Check out Encord's event!

What’s inside (version for the busiest people)

Agility unveils Digit 5 for cooperatively safe work
→ $300M+ multi-year orders, 50 lb payload (+40%), 90-min battery charging in 9 minutes (10:1 run-to-charge)
→ Why OSHA safety compliance + NVIDIA Halos partnership + independent safety controller matters: first deployment outside North America planned H1 2027

SoftBank acquires RAI Institute from Hyundai
→ Marc Raibert’s robotics research legacy returning to SoftBank after Boston Dynamics exit
→ Why timing matters: $5.3B ABB robotics acquisition closing simultaneously, RAI brings AI depth to industrial automation portfolio

UBTECH commissions 10K-unit super factory in Liuzhou
→ 10-minute takt time, flexible mixed-model assembly (bipedal Walker S + wheeled Cruzr on same line) using Cruzr robots for material handling
→ What 1,445% H1 2026 revenue surge (921 bipedal deliveries) vs Figure’s 1K per hour reveals: scale doesn’t solve software generalization

Figure’s Helix 2.5 completes tasks in 30 unfamiliar homes
→ Index-pretrained model 56% success on unseen objects/layouts (bed, towels, toys) vs 9% from scratch, using 50% less data
→ What recovery behaviors (gripping failures, stance shifts, re-looking) demonstrate: whole-body work in unfamiliar environments without prior training

Boston Dynamics opens Atlas training center at Hyundai Georgia
→ Learning parts sequencing and assembly in test environment, scaling to 10x larger facility 2027 with 25K deployments targeted
→ Why evidence to watch: skill transfer speed between tasks/locations, human intervention necessity, reliability gains with training

Agility unveils Digit 5, built for cooperatively safe work at scale! 🦿

Agility Robotics has unveiled Digit 5, its first humanoid engineered to work in close proximity to people without the physical safety barriers traditional automation requires. “We built it to the exact requirements our customers gave us after three years of Digit 4 working on their production floors,” said CEO Peggy Johnson. “The market has responded with more than $300 million in multi-year orders and a growing pipeline of customers across manufacturing, warehousing and logistics.”

The safety architecture is the headline. Building on Digit becoming the first humanoid to pass an independent field evaluation on a customer production line for OSHA-administered industrial safety standards, Digit 5 adds safe human detection using proprietary AI and multiple sensors, which autonomously avoids, stops, or assumes a seated position; visual and auditory cues that convey motion intent to nearby people; and an independent safety controller that triggers responses when someone comes within an unsafe distance. Agility is the first launch partner for NVIDIA’s Halos for Robotics platform, pairing its systems with NVIDIA IGX Thor and Halos Core. The company is also a project leader on ANSI/A3 TR R15.108 and a contributor to ISO 25785-1, the first international safety standard for humanoids.

The hardware upgrades came from three years of customer feedback. A new leg design with proprietary cycloidal actuators lifts up to 50 lb (22.7 kg), 40% more payload, covering the full range of single-person lift tasks in OSHA-regulated facilities. A new 90-minute battery charges in 9 minutes, giving a 10:1 run-to-charge ratio compared to Digit 4’s 2:1, which works out to more than 20 productive hours in a day. Swappable grippers with ISO-standard mounting flanges let tools change per job. Digit 5 weighs 284 lb (129 kg), stands 5’11” (1.81 m), and reaches up to 7.2 feet, well above Digit 4’s 5.5 ft, while still fitting the aisles and doorways built for people.

Where Digit 4 focused mostly on tote-based material handling, Digit 5 is meant to cover a facility’s full workflow: depalletizing on arrival, machine tending, kitting and sequencing, quality inspection, and palletizing for shipment. Agility’s robot management platform coordinates handoffs with existing AMRs, conveyors, and warehouse systems.

The track record behind it: Digit 4 has logged more than 65,000 hours across customer sites including GXO, Schaeffler, Amazon, and Toyota Motor Manufacturing Canada. At GXO’s Flowery Branch facility near Atlanta, it hit a cumulative 100,000-tote milestone at roughly 98% accuracy. Digit is assembled at RoboFab in Salem, Oregon, designed at full capacity for up to 10,000 robots a year. For the first time, Agility plans commercial deployment outside North America, starting with the EU and UK, with early access in the first half of 2027 and general availability by the end of 2027.

Read more here!

Learn more

Sponsor of the newsletter

SoftBank agrees to acquire the RAI Institute! 😮‍💨

SoftBank has agreed to acquire the Robotics and AI Institute (RAI) from Hyundai Motor Group, according to multiple sources. Financial terms aren’t known. The deal has been rumoured for a couple of months and is now under review by the Committee on Foreign Investment in the United States (CFIUS), which screens foreign investments in US businesses for national security impact. How far along the review has progressed is unclear.

There’s history here. Marc Raibert founded both Boston Dynamics and the RAI Institute, and SoftBank owned Boston Dynamics from 2017 to 2021. The RAI was founded in 2022 as part of Hyundai’s acquisition of Boston Dynamics, spun out with an initial investment of more than $400 million. That initial funding has now ended, according to multiple sources. RAI’s research covers robot control, perception, manipulation, navigation, and AI, including the whole-body learning framework behind the acrobatic movements of Boston Dynamics’ new Atlas. Neither party is saying much for now.

The timing is important because SoftBank is also moving toward closing its $5.3 billion acquisition of ABB’s robotics business, originally expected to complete by mid-to-late 2026. ABB brings decades of industrial automation experience and a large installed base; RAI brings AI research depth. Meanwhile, Hyundai said in July that its shareholders were pursuing SoftBank’s remaining 9.65% stake in Boston Dynamics after SoftBank exercised a put option.

Raibert founded Boston Dynamics in 1992 after pioneering dynamic-legged robotics at Carnegie Mellon and MIT. Google acquired the company in 2013 and sold it to SoftBank four years later for a reported $100 million. Hyundai took a controlling stake in 2021 at a $1.1 billion valuation, with SoftBank initially keeping 20%. Raibert stepped down as CEO in January 2020 and launched RAI in 2022. What his role would be if this deal closes is unknown. If it does, SoftBank will once again own a piece of Raibert’s robotics legacy, this time alongside a portfolio spanning industrial robots, warehouse automation, and AI.

Read more here!

Inside UBTECH’s 10,000-unit super factory in Liuzhou! 🏭

UBTECH has officially commissioned its Industrial Humanoid Robot Super Smart Factory in Liuzhou, Guangxi, billed as the world’s first plant engineered specifically for 10,000-unit annual capacity of industrial bipedal and wheeled platforms. The 14,000 square metre facility is built around one benchmark: a completed industrial humanoid rolling off the line every 10 minutes.

Instead of rigid automotive conveyor chains, UBTECH opted for a flexible mixed-model assembly floor that builds the bipedal Walker S series and wheeled Cruzr S2 and Cruzr Y1 systems on the same line. Assembling a single chassis means driving more than 2,000 screws across 50 distinct specifications, handled by automated torque stations that monitor driving depth and stream torque-angle telemetry in real time. Workpieces sit on self-docking rotating dollies with collaborative arms, each chassis tagged with RFID linking its serial number to component batches. Finished units go into a 65 square metre high-bay ASRS buffering up to 112 robots. Every machine gets more than four hours of end-of-line stress testing before passing through an automotive-grade lighting tunnel for visual inspection.

The most striking part is who does the material handling. Wheeled Cruzr Y1 and Cruzr S2 mobile manipulators with 3D vision handle depalletizing, bin picking, tote replenishment, and transit between stations, adapting to shifted bins and uneven pallet orientations. They work alongside AGVs, Wali autonomous forklifts, and Chitu unmanned haulers. The fleet runs on UBTECH’s Yanshee MOM platform integrated with Siemens’ Intelligence Center X. Before breaking ground, UBTECH and Siemens Digital Industries Software built a 1:1 digital twin in Siemens Plant Simulation to check whether a 10-minute takt time was mathematically achievable without bottlenecks.

The economics drive everything. UBTECH is banking on scale to cut manufacturing costs 20-30% annually, targeting sub-$20,000 per unit by 2030, widely seen as the threshold where humanoids stop being enterprise experiments and become standard factory infrastructure. The company says its supply chain is over 90% localised in China, and the Liuzhou site sits near automotive partners SAIC-GM-Wuling and Dongfeng Liuzhou Motor, where Walker S units have already run trials. Its full-size humanoid revenue surged 1,445% in H1 2026 on 921 bipedal deliveries. For contrast, Figure recently celebrated its 1,000th Figure 03 at a takt time of one robot per hour.

The caveat is absorption. Enterprise industrial clients accounted for over 96% of that revenue, and those deployments depend on long validation cycles, custom line integration, and cautious capex. Building hardware also doesn’t solve the software problem: foundation models like UBTECH’s Thinker-VLA have progressed on structured tasks, but real-world generalisability is still open. The question isn’t whether 10,000 humanoids can be built systematically. It’s whether factory floors are ready to put 10,000 of them to work.

Read more here!

Helix 2.5 goes to work in 30 homes it has never entered! 🏡!

Figure rented thirty Bay Area houses, reset the rooms, and sent in a humanoid that had never walked those floors. Helix 2.5, the company’s newest neural network, arrived with no extra practice on those couches, beds, or towels. Staff scattered living rooms with toys, piled up towels, and left beds undone, then asked the robot to finish three jobs on its own: gather every toy into a basket, fold towels and put them away, and make the bed.

Helix 2.5 is a software update rather than new hardware, running on Figure 03 and drawing from Index, the company’s growing collection of movements captured from real humans. The task data came from a different set of toys, towels, and bedding than the ones used in testing, so the robot worked with whatever sofa, bed, or folding surface each house happened to have.

Index did most of the heavy lifting. With training, model changes, and scoring held constant, a copy of the model built from scratch completed roughly 9% of full trials. The Index-pretrained version finished 56%, about six times more often. Across 420 scored attempts (140 per task), the robot completed 237 jobs: 94 with the bed, 87 with the towels, 56 with the toys.

Living rooms held 13 to 15 toys, all of which had to end up in the basket. Towels had to be picked up, folded, and stored in the linen closet, with unfolded corners marked down. Beds needed pillows and comforter corners smoothed or the trial failed. Time limits were strict too: one minute per toy, three minutes per towel, one minute per piece of bedding. Miss the clock and the trial ended. Any safety shutdown also counted as a failure.

Helix 02 could unload a dishwasher and run for long stretches, but everything had been trained in the same rooms. Helix 2.5 matched that success rate using roughly half the data, in locations it had never seen. Recovery behaviour saved several runs: when the robot lost its grip it would back off, shift stance to reach an awkward spot, look again when something went out of view, or walk around the bed to smooth the blanket. Figure says this is the first time it has shown whole-body work in unfamiliar homes with unfamiliar objects.

Read more here!

Boston Dynamics opens an Atlas training center at Hyundai's Georgia factory! 🏗️

Boston Dynamics has announced that its Robotics Metaplant Application Center (RMAC) is operational at Hyundai Motor Group Metaplant America, outside Savannah, Georgia. Atlas robots there are learning to arrange automotive parts for assembly. The company plans to move to a building roughly ten times larger in 2027, and is discussing training and data collection with existing Spot and Stretch customers in aerospace, semiconductors, and logistics.

The release also restates plans for 25,000 Atlas deployments across Hyundai and Kia plants over the next few years, plus a US factory capable of producing 30,000 robots annually. Those are ambitions rather than current fleet or output numbers. Hyundai’s July statement on its robotics strategy put the start of Atlas deployment at the Georgia Metaplant in 2028, beginning with parts sequencing and targeting component assembly by 2030, all subject to technical validation and business requirements.

Sequencing makes sense as a starting point: a production line needs the right components presented in the right order, so the robot has to identify and handle parts while keeping up with the pace of the surrounding process. But training and deployment are different things. Working through a task in a test environment is one step. Sustaining throughput, recovering from mistakes, and operating safely alongside other activity are what determine whether it’s worth anything to a factory.

Boston Dynamics’ Atlas leaders have described a learning approach combining human teleoperation, reinforcement learning in simulation, and eventually observation of human activity. Their argument is that the variety of automotive tasks makes individually engineered automation expensive to extend, so the commercial case for a general-purpose robot rests on being able to teach it new work without rebuilding the automation each time. A center inside Hyundai’s campus is a place to test that against real workflows. The evidence worth watching: how quickly skills transfer between tasks and locations, how much human intervention remains necessary, and whether reliability climbs with more training.

That question also connects to Hyundai’s discussions about bringing Atlas to its Czech factory, though no European deployment date has been announced. Skills developed in Georgia could inform a wider rollout, but success at one site still has to translate to different equipment, processes, and workplace requirements elsewhere.

Read more here about it!

Check out Encord's event!

What’s inside (version for the busiest people)

World Labs unveils Atlas, $1.2B spatial intelligence world model
→ Fei-Fei Li's omni model generates 1-minute 1440p video from single image with pixel-perfect camera control
→ Why spatial reasoning (3D objects, environments, interactions) matters more than text/video generation for AGI and robotics

Maven Robotics emerges from stealth with $100M
→ Mixed palletizing at distribution centers: 8 robots at 99%+ uptime running 16 hours/day
→ Why solving one customer problem at a time beats racing for models: started with “cartoon of robot,” won $80B palletizing market

Dynamic Creatures brings robots to entertainment
→ Marc Theermann (BD CSO) + Farbod Farshidian (RAI) build SnowJay platform for natural movement and human-robot interaction
→ Why interaction as output matters more than dexterous manipulation alone: theme parks, casinos, cruise ships targeting custom IP characters

Skild AI crosses $100M ARR, 10 months post-deployment
→ 60+ customers across mobility (10%), manipulation (90%) spanning NVIDIA Blackwell assembly, wire harness manufacturing, 1.4M daily meals at Mitsui
→ What two-month egg reliability curve reveals: 5% accuracy demo identical to 99%, but cycle-time matching and supplier change management separate deployable from impressive

World Labs introduces Atlas, a world model for spatial intelligence! 🌐

World Labs has raised $1.2 billion from backers including Nvidia, AMD, and Autodesk, and has now unveiled Atlas, a world model that generates, reconstructs, and simulates 3D environments from text, image, and video prompts. The company calls it an “omni model,” pretrained from scratch and built as a multimodal autoregressive diffusion transformer, letting users create a scene and view it from any camera position they specify.

Atlas generates up to a minute of video at 1440p from a single image with what World Labs describes as pixel-perfect camera control, and reconstructs real scenes from as few as one image while keeping geometric consistency in its 3D output. The company points to visual effects and game design as applications, but robotics is the real target. The model supports real-to-sim workflows where developers capture a physical space with an ordinary smartphone and rebuild it as a simulation robots can navigate and interact with.

The release advances Li’s theory of spatial intelligence: that AI cannot replicate the physical world without reasoning natively about 3D objects, environments, and how they interact. Li, former director of Stanford’s AI Lab, launched World Labs in February 2024 on the argument that artificial general intelligence is impossible without those capabilities.

The competition is stacking up. Atlas goes directly against Google’s Gemini Omni and overlaps with video generation models including ByteDance’s Seedance 2.5 and Google’s Veo. There’s also pressure from China, where a number of startups founded this year are racing to build world models of their own. In a blind human evaluation of camera-path adherence, World Labs says Atlas was preferred over competitors including Gemini Omni Flash and FLUX.

World models have become their own front in the AI race, with the biggest players targeting the same ground. World Labs is betting that spatial reasoning, rather than text or video generation, is the missing layer for both robotics and creative tools. Atlas will power future versions of Marble and other World Labs products.

Read more here!

Learn more

Sponsor of the newsletter

Maven Robotics comes out of stealth with $100M! 📦

Maven Robotics has emerged from stealth with $100 million from RoboStrategy, LocalGlobe, Vine Ventures, and XTX Markets Ventures, and plans to build 250 third-generation robots while starting design on a fourth-generation platform.

The origin story explains the approach. In 2024, Maven had nothing but what CEO and co-founder Hamza Derbas calls “a cartoon of a robot and a team of people.” A large consumer goods company was in town meeting four rival robot companies about automation. Derbas talked his way into a meeting and, rather than pitch his views on robots, asked to visit their factories and warehouses. They won the deal against companies with working robots. Two years later Maven says it has as many as eight robots running 16 hours a day at 99%+ uptime.

The robots sit on wheeled bases that move 10 miles an hour, with two arms lifting up to 30 kilograms. Their main job is mixed palletizing: pallets of boxed goods arrive at a distribution centre from different factories, and the robot builds new pallets containing a mix of products for individual stores. “Within 48 hours of them putting the stuff on the shelves, they want to change the mix based on real-time demand,” Derbas said. Today that work is done by people running around the warehouse picking one of this, one of that.

Derbas came from automotive engineering and spent nine years at Apple’s special projects group, widely believed to have been the self-driving car effort disbanded in 2024. He founded Maven with his brother Khalid, now CFO. Like other physical AI companies, the team leans on self-driving veterans and their data pipelines, which return information from operating robots within minutes or hours for retraining and redeployment.

Palletizing alone is an $80 billion market, but the next tasks Maven wants require manipulation capabilities that don’t exist yet. The company is collecting more data, tapping third-party providers, and has built pincer-like gloves so humans can emulate the gripper form factor they want. “We’re grounded in solving one customer problem at a time,” Derbas said. “We’re not in the race for models — we’re in the race to solve industrial labor.”

Read more here!

Boston Dynamics spinout to bring characters to life with robotics! 🎡

ex-Boston Dynamics CSO, and ex-RAI veteran have launched Dynamic Creatures, a company building robotic characters for entertainment. It was co-founded by CEO Marc Theermann, former chief strategy officer at Boston Dynamics, and CTO Farbod Farshidian, former research leader at the RAI Institute. The name nods to both Boston Dynamics and Jim Henson’s Creature Shop.

Theermann traces the idea back to those Boston Dynamics videos of Spot and Atlas dancing and flipping. “We always put out those videos, A, because it’s super fun, and B, because we wanted to bring the general public along,” he said. Along the way he started noticing real demand. “Humans love to see robots dancing and love to see robots in these fun environments.” The company is targeting theme parks, casinos, cruise ships, and hospitality, and is Boston Dynamics’ official entertainment and hospitality partner.

Investors include Eniac Ventures, Kindred Ventures, Heliad, Sunshine Lake, and BlueGrass Ventures, plus strategic angels Marc Raibert, founder of Boston Dynamics, and Lukas Ziegler (yes, it’s me!!!).

The company’s platform is called SnowJay, and it handles two things: natural movement and human-robot interaction. “One of the cores of our platform is that we can make any robot walk in these very natural ways,” Theermann said, either through motion capture or animation, using NVIDIA tools and reinforcement learning. Farshidian leads that work, having helped give Atlas its more natural gait at CES 2026. The interaction side is where Theermann thinks the industry is looking the wrong way. “If you look at many robots out there, they’re deaf and mute. Everybody is talking about vision language action models... Lots of people are focused on dexterous manipulation as the next big hurdle, which is certainly super important, but that’s not what we’re focused on. We’re focused on human-robot interaction as the output.”

Right now Dynamic Creatures builds on third-party hardware, mainly Spot and Atlas, with Unitree compatibility too, though Theermann says they want to build their own humanoid eventually. Two in-house prototype characters exist: a purple poodle named Danielle and a Tyrolean mountain troll. But the first two partners want something else. “Partners coming to us with their IP and saying, can you bring my IP to life? I have an IP already from a movie or from a TV show.”

First customer deployments are targeted for early 2027, starting as managed pilots with a mix of teleoperation and autonomy. The company has hired puppeteers to work out which emotions land. Safety is the other research thrust. “Right now, in factory environments, deployed humanoid robots typically are behind a light curtain,” Theermann said. “We want to get to a point where you can hug the robot.”

Read more here!

🚨 ICYMI!

Next week there’s a great event in Menlo Park hosted by Fellows Fund called Fellows Forum, with a lot of the conversation centered on deploying physical AI into real-world environments.

If you’re in the Bay Area on September 23–24, don’t miss it!

Check it out here!

Skild AI crosses $100M ARR ten months after its first deployment! 💼

Skild AI has crossed $100 million in annual recurring revenue ten months after its first commercial deployment, with $50 million already recognised. The company now has 60+ paying customers across moving goods, deliveries, site inspection, security, food preparation, and work inside warehouses, factories, and data centers. Mobility accounts for 10% of revenue, including the AMR business from the Zebra/Fetch Robotics acquisition at 4%. The remaining 90% comes mainly from manipulation.

Some of the named deployments are substantial. With NVIDIA and Foxconn, Skild Brain runs on dual-arm manipulators doing high-precision assembly of NVIDIA Blackwell systems, work that changes every product cycle and used to mean reprogramming every robot on the line. At Sumitomo Wiring Systems, the company is working to automate wire harness manufacturing processes long considered impossible to automate. With Mitsui & Co., whose supply chains serve 1.4 million meals a day across Japan, it’s piloting general-purpose robots in commercial kitchens.

Demo videos are a bad measure because a successful clip from a robot with 5% accuracy looks identical to one with 99% accuracy. Skild taught its model to make eggs last year; the first egg took a week, and two months more to work reliably across different eggs and setups. Accuracy alone also misleads: on a line where five stations are robots and five are people, every station must hit roughly the same cycle time, so a robot that’s 99.9% accurate but ten times too slow isn’t almost deployable, it’s not deployable. And change is constant, since suppliers swap components and customers reorder assembly sequences. That last lesson drove S1, the in-context learning model released two weeks ago, where an operator records a demonstration and hands it to the robot as a prompt without touching the weights.

The longer-term idea they call physical RSI: deploy a general model, let each deployment specialise, then fold data from across those deployments back into the base model. Experience that adds little to a specialist who has already mastered one task can still help a generalist learning many. The analogy they use is doing separate PhDs in chemistry, physics, and maths in parallel universes, then distilling all of it back into your high-school self.

Read more here about it!

Hugging Face launches Microduck open-source robot at $399
→ 25cm duck with camera, lidar, IMUs learns new tricks via RL in simulation then deployed
→ What acquisition of Pollen Robotics signals: democratizing physical AI through affordable open-source hardware

Gobano’s Toutatis RL engine achieves 99%+ reliability from 10-200 demos
→ Task-specific world model + off-policy RL on frozen embeddings reaches 99/100 on towel folding, 31/31 on zip insertion
→ Why splitting Reliability (repeated success) vs Performance (speed) modes matters: conventional RL blends both, masking rare failures

Gatik raises $200M following PepsiCo multiyear deal
→ 41 driverless box trucks moving Frito-Lay products in Dallas, Phoenix, Arkansas; $600M contracted revenue locked
→ Why middle-mile niche (dynamic routes up to 400 miles) succeeds where robotaxis/long-haul competed

Teradyne sues JAKA German subsidiary for patent infringement
→ Multiple patents covering hardware and software in force/power-limited arms; Unified Patent Court ruling covers 17 EU states
→ Why takedown signal matters more than compensation: protecting safety standards around cobots vs copycat products

Nori launches A3 household robot at $1,688
→ Wheeled platform with 2-claw arms, 8-hour runtime, owner-trained skills via desktop app or code
→ What no teleoperation + community skill marketplace reveal: developer-first approach vs fixed capabilities

Hugging Face is selling a $399 open source duck robot! 🦆

Hugging Face has unveiled the Microduck, a duck-like robot that sells for $399 and ships before Christmas. CEO Clem Delangue described it as an open-source robot you can teach new tricks with reinforcement learning. The 25-centimetre-tall duck can waddle, pick things up with its beak, get back up when it falls, crouch, and roller skate. “Welcome to the era of open-source affordable robots to democratize physical AI and world models,” Delangue said.

The Microduck perceives the world through a camera, lidar sensors, and two IMUs. Pollen Robotics notes that behaviours can be trained in simulation and deployed directly onto the robot, after which developers can fine-tune, retrain, and redeploy. The SDK, simulation, and full RL training stack are on GitHub.

Hugging Face is best known as the platform where developers go for open model weights, but it acquired French startup Pollen Robotics in April 2025 to build affordable open source AI hardware. The two launched Reachy Mini a couple of months later. They currently sell the $499 Reachy Mini, powered by a Raspberry Pi, and the $399 Reachy Mini Lite, which runs off a Mac or PC.

On the privacy question of putting a camera-equipped robot in your bedroom, Delangue has previously argued that bots running open source models are better than a black box system controlled by a few organisations. Open source does give developers auditability and control, though it doesn’t guarantee sensitive data stays private once consumers install third-party apps on top, since those apps can access cameras and microphones and send data elsewhere depending on how they’re built.

The launch lands while Hugging Face is reportedly set to be acquired by Nvidia at a $13 billion valuation. The two have partnered for years, with Nvidia providing infrastructure since at least 2023, and both have publicly pushed for open source AI.

Read more here!

Gobano introduces Toutatis, an RL engine for getting robots to 99%+ reliability! 🎯

Gobano Robotics has released Toutatis v1, a reinforcement learning engine that takes 10 to 200 human demonstrations and turns them into reliability-optimised controllers in days. The framing is about what separates a lab result from a production system: a robot that does a task once proves capability, but one that does it shift after shift creates value. At 95% success you get 50 failures per 1,000 operations. At 99% you get 10. Production wants 99.9%, or roughly one.

The core argument is that demonstration data teaches a robot what people did, not what maximises success or eliminates the rare failures that stop a system running unsupervised. Gobano draws the analogy to language models: pretraining creates capability, RL converts it into performance. They point to DeepSeek-R1-Zero, where pure RL from a pretrained base produced reflection and verification behaviours without supervised reasoning examples. Their thesis is that RL is both necessary and sufficient for robot deployment, because deployment is defined by outcomes rather than demonstrations.

Technically, Toutatis trains a task-specific world model from the same demonstrations used for the imitation policy. The model builds a compact latent representation of what’s in the scene, how it changes under an action, which transitions create progress, and where failures emerge. Gobano then freezes the visual encoder and runs off-policy RL directly on those cached embeddings, which made training roughly 10x faster than their equivalent end-to-end visual RL setup. Two timescales run in parallel: fast policy updates on the frozen representation, and periodic world-model refreshes as real-world coverage expands.

One design choice worth noting is the split between “Reliability” and “Performance” training modes. Conventional RL maximises average task reward, which blends speed and success, so the average can rise while rare failures persist. Reliability mode prioritises repeated successful completion until the acceptance threshold is hit; Performance mode then rewards faster completion while protecting that reliability level.

On the Robomimic benchmark, Toutatis hit 1,000/1,000 on Square and 998/1,000 on Tool Hang over trailing windows of 1,000 rollouts. On real hardware, ball picking went from 11 demonstrations and an initial 24/100 to 108/108 across 854 rollouts. Towel folding started from 135 demonstrations, needed 215 more targeted ones to cover harder configurations, and finished at 99/100 after 2,633 rollouts. Zip-tie insertion reached 31/31 after 3,375 insertion-focused rollouts. The models are small enough to train on a gaming PC and run locally on a gaming laptop, with inference around 3 ms for ball picking and 10 ms for towel folding. Gobano is already using Toutatis in deployments and plans to open the system publicly.

Read more here!

Self-driving truck startup Gatik raises $200M following PepsiCo deal! 🚚

Gatik has raised $200 million, two months after signing a multiyear commercial agreement with PepsiCo. The round was led by Qatar Investment Authority and Koch Disruptive Technologies, with Millennium Management, ARK Invest, and Intact Private Capital participating. The Santa Clara company has now raised about $500 million since coming out of stealth in 2019 and didn’t disclose its valuation.

“These are all long-term investors,” said co-founder and CEO Gautam Narang. “All the scale-up plans that we have and the growth that we have in mind for the next few years, these are the kind of financial partners that can help us.”

Gatik picked a lane nobody else wanted. While dozens of companies chased robotaxis, sidewalk robots, long-haul big rigs, and mining vehicles, Gatik built driverless box trucks with Isuzu Motors for short-haul delivery. It started with fixed routes under 10 miles and grew to dynamic routes with dozens of pick-up and drop-off points covering up to 400 miles. That middle-mile niche is now largely its own, with customers including Loblaws, Kroger, Tyson Foods, and Walmart, its first customer disclosed six years ago.

Last year the company pulled safety drivers from its commercial routes. It now runs dozens of fully driverless trucks commercially across several markets, with third-generation vehicles operating around the clock on surface streets and highways in light rain and snow. The largest public partnership is with PepsiCo: 41 driverless box trucks moving Cheetos, Doritos, and other Frito-Lay products from distribution centres to stores in Dallas, Phoenix, and Northwest Arkansas.

Narang says the company has locked in commercial deals worth $600 million in contracted revenue. The new funding goes toward hiring engineers and operational staff beyond its current 350-person workforce, and expanding into new markets, potentially outside North America eventually.

Read more here!

Teradyne Robotics ramps up its fight against cobot copycats! ⚖️

Teradyne Robotics, parent company of Universal Robots, has filed a patent infringement lawsuit against the German subsidiary of Chinese robotics company JAKA. The case alleges JAKA infringed patents covering both hardware and software used in Universal Robots’ cobots. It’s Teradyne’s second major case against a Chinese robotics company in Europe this year, after suing Elite Robots’ German subsidiary over copyright infringement in February. In April, the Regional Court of Hamburg issued a preliminary injunction barring Elite Robots Germany from offering the disputed software in Germany.

The JAKA case works differently. “This is a patent case where the other one was about copyright,” David Brandt, VP of R&D and CTO at Universal Robots, told The Robot Report. “It’s multiple patents this time, and they cover both significant hardware and software aspects of our products.” Teradyne became interested after noticing similarities between the two companies’ force- and power-limited arms. “Our robots and the robots of JAKA look quite alike,” Brandt said. “We then chose to acquire one of the robots and take a deeper look at it. Then it became quite obvious that there’s possible patent infringement in this.”

The case was filed with the Unified Patent Court Local Division in Copenhagen, since the alleged infringement occurred in Denmark and other EU member states. A ruling could take effect across 17 of the 18 EU member states participating in the UPC, plus certain countries outside the agreement including the UK and Spain. JAKA did not respond to a request for comment.

Brandt was clear the decision wasn’t about JAKA’s Chinese ownership. “To us, it’s not about China. It’s about other companies copying stuff that we designed, whether they’re Chinese or not.” He also acknowledged the basic difficulty of protecting a robot once it’s shipped: a competitor can buy one, take it apart, and study the design. The goal isn’t compensation. “Ideally, we would like them to stop copying and infringing our IP. This is more to send a very clear signal.”

There’s a safety argument too. Universal Robots helped establish the market for robots that work near people under proper risk assessment, and Brandt worries that copycat products not meeting those standards could hurt someone and reflect badly on the whole industry, pushing regulators to treat cobots as a safety problem.

Teradyne Robotics posted $100 million in Q2 2026 revenue, up from $91 million in Q1 and 33% higher year over year, marking its fifth consecutive quarter of growth. Universal Robots has sold more than 100,000 cobots. Brandt suggested JAKA won’t be the last target: “We do not see this as a one-time thing. JAKA is not the only robot out there that looks suspiciously like ours.”

Read more here about it!

Nori Robotics launches a $1,688 household robot! 🔧

Nori Robotics, a Y Combinator-backed San Francisco startup, is selling the A3, a wheeled household robot for $1,688. The company says it can help cook, fetch items from the fridge, put things away, load the dishwasher, fold clothes, and pour ingredients. It has two arms with claws that lift a couple of pounds each, uses LiDAR and cameras for navigation, and runs up to 8 hours on a full charge. The first batch, roughly a hundred units, has already sold out.

The catch is who it’s for. Nori describes the A3 as a robot “built for people who want to actually build with one.” Unlike a robot vacuum with a fixed set of capabilities, the A3 is meant to be trained by its owners, who can then package that training into skills and share them with the Nori community through an online marketplace. Owners can either write code or train it via a desktop app that simulates the robot’s movements on screen. That opens it up beyond housework: one demo shows it flipping pages of sheet music while a guitarist plays.

That approach cuts both ways. You probably won’t get much functionality out of the box, or any guarantee the robot performs tasks without mistakes. But it may also mean the A3 doesn’t lean on teleoperation, where a human watches through the robot’s cameras and remotely drives it through the task, which is both less impressive and more intrusive than most buyers expect from a home robot.

Price-wise it undercuts everything else in the category by a wide margin. Weave Robotics’ Isaac 1 costs $8,000 and ships this fall. The bipedal 1X NEO is $20,000 and arrives next year. Nori also says most of the A3’s parts can be 3D printed and replaced by owners.

“We’re excited to see what developers, researchers, and teams build when capable humanoid hardware becomes this accessible,” said founder Antonio Li. A second batch of 100 units is open for order now, shipping this fall.

Read more here about it!

NVIDIA acquires Hugging Face for $12.9B
→ Open-source AI platform consolidation as chip dominance faces erosion (OpenAI, Google, Amazon, Anthropic building own chips)
→ Why open-source ecosystem on NVIDIA hardware matters more than closed labs for customer optionality

Figure launches Index for crowdsourced robot data
→ 264K downloads, 44K weekly active users, 16M videos in 4 months; 30 mins/second upload pace (4.9 years daily)
→ What 373 unique tasks, 1,146 objects, 116 environments per 1K hours reveal: diversity beats vendor throughput

SoftBank in talks for 1X majority stake at $6B valuation
→ Control purchase after selling Boston Dynamics stake, anchoring Skild AI $1.4B, acquiring ABB robotics $5.4B
→ Why balance sheet stability matters: funding production lines, World Model Lab compute, NEO delivery infrastructure

Skild AI releases S1 with video in-context learning
→ One set of weights executes unseen 10-minute tasks (plant potting, pancake cooking) from single demonstration, no fine-tuning
→ What 66% unseen task success at 100K hours vs 9% for language-prompted VLA reveals: video prompting worth 380 post-training episodes

Waymo planning Munich robotaxi service, public launch end 2027
→ Follows 11-city expansion playbook: manual mapping → testing with safety drivers → employee/guest access → public launch
→ Why Germany’s 5-year-old Level 4 framework attracts competitors (Mobileye, VW testing) vs London battleground with Wayve, Uber, Baidu

NVIDIA closes in on Hugging Face acquisition for $12.9B! 🤗

NVIDIA has agreed to buy Hugging Face for $12.9 billion, according to The Information. Business Insider, which first reported the takeover interest over the weekend, says the talks value the company at more than $13 billion but haven’t produced a signed agreement yet and could still fall apart. Neither company has commented, which is notable in NVIDIA’s case since it usually moves fast to correct reports it thinks are wrong.

Hugging Face, founded in 2016, is the main place developers share and download open-source AI models. Owning it would give NVIDIA a foothold in open-source AI at a moment when Nvidia’s chip dominance looks less secure than it did. OpenAI, Google, Amazon, and Anthropic are all building their own AI chips to reduce their dependence on Nvidia. A healthy open-source ecosystem gives customers alternatives to those closed labs, and those alternatives mostly run on Nvidia hardware. That’s the same logic behind NVIDIA already pouring tens of billions into its own open-source models.

Hugging Face CEO Clem Delangue has been publicly aligned with Nvidia’s open-source push all year. On CBS’s “Face the Nation” this month he described using an NVIDIA-modified version of a Chinese open-source model to defend against a cyberattack, and pointed to a letter signed by Jensen Huang and 24 other companies urging Washington to support open models rather than restrict them. In a CNBC interview in late July he warned that China is clearly dominating open-source AI.

There are other angles. Nvidia scaled back its DGX Cloud business about a year ago, and Hugging Face already helps developers run models on rented compute, offering a way back into that market. Nvidia has also promised to cover the cost of tens of billions in cloud deals for customers, and could resell unused capacity to Hugging Face’s users.

The price is a big jump. Hugging Face last raised $235 million in 2023 at a $4.5 billion valuation, in a round led by Salesforce Ventures with participation from GV, IBM Ventures, and NVIDIA. It also turned down a $500 million investment from NVIDIA late last year at a $7 billion valuation, saying it didn’t want a dominant investor swaying its decisions. Revenue is around $150 million a year, up from roughly $100 million two months earlier, and Delangue told TechCrunch last month the company is close to profitability. A $13 billion price would be a steep multiple on that, but a hard one to walk away from.

Read more here!

Figure comes out of stealth with Index, its physical data collection network! 📱

Figure has unveiled Index, an app-based pipeline for collecting robot training data from people around the world. The argument behind it is that the data needed to scale a general purpose robot doesn’t exist on the internet. It has to come from the real world, sampled across as many environments as possible.

The numbers from four months in stealth are substantial. Index has crossed 264,000 downloads across 108 countries with over 44,000 weekly active users. Creators, as Figure calls them, have uploaded more than 16 million videos. The app processes 30 minutes of video uploads every second, which works out to 4.9 years of human work uploaded daily. Figure has paid out $15 million to Creators so far and says it will spend over $1 billion on data and compute over the next 12 months.

Figure tried buying data first. Vendors couldn’t hit the throughput, diversity, or quality bar its Helix AI stack needed, so the company built the pipeline itself. Per 1,000 hours collected, Index contains 373 unique tasks, 1,146 unique manipulated objects, and 116 unique environments. The diversity comes from the people generating it: every new Creator brings an unfamiliar environment, unusual objects, and their own way of doing things. Figure says it has collected tasks as obscure as cleaning kitty litter, changing oil, and busing restaurant tables, alongside cooking, laundry, and work inside logistics centres, restaurants, factories, and offices.

Ingesting that much video meant rebuilding the data infrastructure around consumer app constraints: 24/7 availability, continuous large-scale compute, and real-time feedback to users. The pipeline runs five stages. Automated filters screen for technical, visual, and semantic quality. Human analysts audit samples at the user level for deliberate evasion attempts. Videos are embedded and discarded if they’re too similar to existing data. What remains gets rebalanced using task quotas and embedding-based clusters. Finally, hierarchical text captions are generated for every episode.

Anyone can sign up as a Creator and record tasks in their own home or workplace, or book a Creator through the app to come do chores at their house or business. Figure frames this as groundwork for ordering robots as a service: today a person comes to clean your house, eventually a robot does it.

Read more here!

SoftBank in talks to buy majority stake in 1X at $6 billion valuation! 💸

SoftBank is in talks to take a majority stake in 1X Technologies at a valuation of roughly $6 billion, according to The Information. A completed deal would hand Masayoshi Son controlling interest in one of the most prominent bipedal platforms in the West, and would mark a defining moment for 1X as it tries to move from research prototypes to hardware in people’s homes. Neither company has commented.

Last September, 1X was reported to be seeking up to $1 billion at a valuation of $10 billion or more. Against that, $6 billion reads as a markdown, but the comparison needs a caveat: the $10 billion was an ask that was never confirmed to have closed, and a majority purchase is priced differently from a minority round. Control usually trades at a premium to the venture mark, not a discount. What surrendering a controlling interest buys 1X is balance sheet stability, which matters when you’re funding low-volume production lines in Hayward, compute for the 1X World Model Lab, and support and logistics infrastructure ahead of NEO deliveries.

The reporting leaves the most consequential terms unaddressed. How large is “majority”? A 51% position and an 80% position mean very different things for 1X’s independence. Is $6 billion pre- or post-money? Is the money primary capital funding the company, or secondary cashing out existing holders? Is OpenAI, 1X’s earliest institutional backer, selling? And does founder Bernt Børnich keep operational control? None of it is public.

For SoftBank, the deal fits a broader consolidation of the robotics supply chain around what it treats as a split between brains and bodies. After selling its remaining Boston Dynamics stake to Hyundai for $325 million this summer, SoftBank anchored a $1.4 billion Series C in Skild AI for foundation models, bought ABB’s robotics division for $5.4 billion, and is reportedly in talks to anchor an $800 million round for Agile Robots. NEO covers the one category ABB’s arms can’t reach: compliant, quiet, tendon-driven hardware built to operate around people in unstructured homes.

The strongest argument that $6 billion undervalues 1X is demand. When preorders opened in October 2025 at $20,000 outright or $499 a month, the company reportedly sold out its entire first year of production in five days. The strongest argument the other way is that none of it has shipped. 1X promised initial NEO deliveries in 2026, and it’s now late August with no customer deliveries publicly confirmed. Closing that gap is exactly what SoftBank’s money would pay for, and exactly what makes the timing of a control deal worth watching.

Read more here!

Skild AI introduced S1: show it a video, and the robot does the task! 🎬

Skild AI has released S1, a robot foundation model built around in-context learning. Show it a video of a task, short or long, seen or unseen, and it executes. No fine-tuning, no post-training. One set of weights produced every result in the company’s announcement.

The framing is a comparison to language models. Skild argues robotics has been stuck in the BERT era, where each new application demands more data collection and another fine-tuning run. The shift from those early models to ChatGPT came from in-context learning, or prompting: a user introduces a novel concept through the prompt and the model handles it without touching the weights. Skild thinks that’s the whole point of pre-training in robotics too. Notably, S1 uses video demonstrations rather than language to specify tasks. Language works fine for atomic actions like “hand me the mug,” but nobody learns to fold a fitted sheet or whisk egg whites to stiff peaks from a sentence.

The headline results are on long-horizon tasks the model never saw during training: plant potting, pancake cooking, pour-over coffee, and kit assembly. These run up to ten minutes, span dozens of steps, and are driven by a single visual demonstration. S1 dug into soil to make room for a plant, pressed a coffee filter into a funnel, and flipped a pancake. Skild says this is the first time a robotics foundation model has shown in-context learning on tasks this long that weren’t in the pre-training set. For the plant potting task, the supplies arrived at the office at 8:54 PM and S1 was executing autonomously by 9:27 PM. The gap between recording the demonstration and autonomous execution was 11 minutes.

The scaling study is where the case gets made. Skild trained both in-context and language-conditioned policies on identical data and compute, from 1k to 100k hours. On tasks the models had seen, the two end up close: 96% for ICL against 89% for the language-prompted VLA. On unseen tasks the gap is enormous. At 100k hours, language prompting reaches 9% while ICL hits 66%. Expressed differently, one video demonstration is worth roughly 380 post-training episodes, which for tasks over four minutes means 50 to 100 hours of teleoperation.

Several behaviors emerged that weren’t explicitly trained. S1 kept working when objects were slid away mid-approach, swapped out, or the lighting changed. It retried after failures rather than blindly proceeding. When a prompt showed watering a plant with a watering can but only a cup was available, it used the cup. And it sometimes improves on flawed demonstrations: in one prompt the demonstrator dropped an egg and made a mess, while S1 did the same step with a controlled motion. It treats the demonstration as a specification of the goal, not a trajectory to copy.

Read more here about it!

Waymo robotaxis are headed to Munich! 🥨

Waymo plans to launch a robotaxi service in Munich, months after setting up an entity in Germany. It still has mapping, testing, and regulatory hurdles to clear before commercial operations can begin.

The rollout follows the same sequence Waymo has used in the other 11 cities where it operates commercially. First it manually drives its vehicles to map the streets. Then it tests autonomous vehicles with human safety drivers, before removing them. At that stage it usually opens rides to employees, media, and invited guests, then to limited public service, and finally to all riders once regulators sign off. Waymo says it’s working with Germany’s Federal Motor Transport Authority (KBA), state authorities, and local officials, and expects commercial ride-hailing to open to the public toward the end of 2027.

Germany was the first EU country to create a legal framework for Level 4 autonomous driving, adopting the legislation more than five years ago. That has made it a hotspot for AV testing, with Mobileye and Volkswagen among the companies holding permits. Most permit holders, including Wayve and Autobrains, are at Level 3, where the system handles driving in specific conditions but a driver must be ready to take over. Waymo currently holds no permits in the country.

Germany isn’t the only European battleground. London is shaping up as another one, with Waymo planning a commercial service there in 2026 that would put it against Wayve and Uber. Baidu has also begun testing in the city through its partnership with Lyft and Freenow.

Read more here about it!

Unitree shares close up 460% on Shanghai debut
→ 150.80 yuan offer to 845 yuan close, first mainland-listed humanoid maker
→ Why profitability (278M yuan net profit 2025) + volume (5,500 units) separates Unitree from hype

Veeda AI raises $90M seed from Radical, Khosla
→ Ex-NVIDIA researchers (Fidler, Ling, Gojcic) building world models for physical AI with global footprint
→ What “The Matrix for physical AI” reveals: simulating trial-and-error avoids costly real-world failures

Gravis Robotics lands $200M SoftBank Series A
→ ETH Zurich spinout retrofitting excavators across 8 manufacturers (Caterpillar, Volvo, John Deere, Hitachi) with same core software
→ Why mixed-fleet compatibility matters: 2/3 global heavy equipment outside top 3 makers, driven by existing fleet investments

Uber partners Zipline for one million daily drone deliveries by 2029
→ First Uber Eats drone drops by end 2026, expanding from Zipline markets to dozens of US cities
→ Why platform aggregation strategy (Zipline + Flytrex, no in-house) prioritizes breadth over control

Matic launches Cues voice and gesture control
→ Point at spill, say “clean this,” robot localizes and cleans autonomously across 75 languages
→ What 8x airflow vs competitors + 9-year development + WIRED 10/10 reveal: specialization in execution over novelty
→ 13K families using after years building, $115M raised with 6-month refund guarantee

Unitree shares close up 460% on Shanghai debut! 📈

Unitree Robotics listed on Shanghai’s STAR Market on Wednesday. Shares were offered at 150.80 yuan ($22.36) and closed the first day at 845 yuan, up more than 460%. Other Chinese humanoid makers have gone public before, but this is the first listing on the mainland.

The company was founded in 2016 and is officially called Yushu Technology. It shipped over 5,500 humanoid robots last year and turned a net profit of 278 million yuan in 2025, which almost nobody else in the sector can say. Price is its main weapon against US competitors. Unitree’s robot dogs start at $2,700. Boston Dynamics’ Spot costs around $70,000. Its child-sized G1 humanoid launched in 2024 at $13,500, and Tesla still hasn’t delivered a rival product.

Unitree is based in Hangzhou, part of a cluster of robotics firms that have taken in heavy government investment. The number of Chinese robotics companies more than tripled between 2020 and 2024. Fei Qin of the University of Bath said Beijing treats the sector as a strategic priority, partly because an ageing population points to a worker shortage. Leju Robotics and AgiBot are both expected to list in the coming months.

Not everyone is convinced. Harold Soh of the National University of Singapore said home robots are still years away, with battery life, privacy, and reliability all unsolved. The near-term market is factories and hospitals. There’s also the Trump administration’s July announcement that it will ban new Chinese-made humanoid and quadruped robots on national security grounds, which Beijing called politicising trade. Robotics researcher David Hsu was blunter about where things stand: American robotics companies barely make the news anymore, and Chinese firms have taken their place as the cool kids on the block.

Read more here!

Ex-NVIDIA researchers raise over $90M for world model startup Veeda! 🌍

Sanja Fidler has left NVIDIA after eight years and started Veeda AI, a Toronto company building world models for physical AI. It has raised over US$90 million in seed financing from Radical Ventures and Khosla Ventures. That would make it one of the three largest seed rounds in Canadian history.

Sanja is the CEO, and her co-founders are both former NVIDIA colleagues: chief scientist Huan Ling, who worked with her at NVIDIA’s Toronto Spatial Intelligence Lab, and chief technology officer Zan Gojcic, a Zurich-based former research director. The three incorporated in early June and issued 60.6 million seed shares at US$1 each in late July, adding Radical partner Tomi Poutanen and Khosla partner Sven Strohband as directors on the same day.

Fidler described the company as building The Matrix for physical AI. The idea is that robot intelligence can interact with an entirely virtual world in a million possible ways and learn from its own mistakes, because trial and error in real life is neither safe nor economical. World models are typically trained on image and video data and simulate the consequences of actions, giving systems a better grasp of real-life mechanics. Veeda is hiring scientists and machine learning engineers, plus staff to collect training data and manage chip clusters. Its site lists offices in Toronto, Mountain View, Singapore, and Zurich, with a Swiss subsidiary set up in July.

Fidler is a significant figure in Canadian AI. She joined the University of Toronto’s computer science faculty in 2014, was a founding member of the Vector Institute, and holds a federal research chair. At Nvidia she helped build the company’s flagship world model alongside Ling, whose doctorate she supervised. She’s also the latest researcher to leave a US tech giant to start a foundation model company. Others in the world model space include Yann LeCun’s AMI Labs, Fei-Fei Li’s World Labs, and Waabi, led by fellow U of T professor Raquel Urtasun, which raised US$750 million in January co-led by Khosla.

Read more here!

Gravis Robotics raises $200M from SoftBank for autonomous construction! 🚜

SoftBank is putting $200 million into Gravis Robotics’ Series A, which the company says is the largest round in construction robotics history. Gravis spun out of ETH Zurich in 2022 and retrofits existing excavators and construction equipment with its Gravis Rack autonomous control kit and software.

The technical problem is that most physical AI works in static worlds. A self-driving car steers around obstacles on fixed pavement without disturbing anything. An excavator does the opposite: it deliberately reshapes the ground it’s working on. Gravis trains on simulated experience, moving billions of cubic yards of virtual material from soft clay to rock-filled soil. “Skilled operators read the earth through subtle physical feedback, listening to the engine strain, sensing the machine vibration and reacting to hydraulic resistance,” said co-founder and CTO Dominic Jud. “Our AI takes that same physical input and grounds it in machine telemetry, responding to varying subterranean forces and soil mechanics at microsecond speeds.”

The mixed-fleet angle matters commercially. Roughly two-thirds of global heavy equipment demand sits outside the top three manufacturers, and contractors pick machines based on regional service relationships and existing fleet investments. The Gravis Rack has been installed on machinery from Caterpillar, Case, Develon, John Deere, JCB, Hitachi, Sumitomo, Yanmar, and Volvo, and the same core software can drive a compact excavator or a machine five times its size without custom reprogramming. Gravis claims productivity gains of up to 30% over peak manual operations.

Operators can stay in the cab with Gravis Copilot, using real-time 3D guidance and hazard detection, or step away entirely and supervise robotic fleets when full autonomy is engaged. Either way, every machine with the kit acts as a mobile sensor, automating surveying and hazard mapping while other work happens. Gravis has deployed across four continents and was recently chosen to lead an $8 million UK government-backed CAM Pathfinder project with Flannery Plant Hire, the country’s largest equipment rental provider.

Read more here!

Uber adds Zipline drones to its Eats delivery network! 🚁

Uber is investing in and partnering with Zipline, with the goal of making one million deliveries per day using the startup’s drones by the end of 2029. The first Zipline deliveries on Uber Eats will happen by the end of this year, starting in Zipline’s existing markets, with plans to expand into dozens of US cities. Neither company disclosed the investment amount.

Uber thinks Zipline’s drones can fulfill Eats orders within five to 10 minutes. “Truly quick commerce is proving to be an even bigger market than the original food market was,” CEO Dara Khosrowshahi told The Wall Street Journal. “We think this can be an enormous tailwind for the next leg of growth for Eats.”

This follows the playbook Uber has used for robotaxis: bring as many companies onto the platform as possible rather than building the technology in-house. The company sold off Uber Elevate and its Autonomous Technologies Group, then committed more than $10 billion to dozens of autonomous vehicle providers instead. It’s already working with Israeli drone startup Flytrex after a partnership announced late last year. The approach has limits, though. Uber recently clashed with Waymo, and the two are expected to part ways when their contracts expire in 2028, having also landed on opposite sides of a growing fight over autonomous vehicle regulation.

Zipline, based in San Francisco, recently closed an extended $800 million Series H that pushed its valuation to $7.6 billion. “Every great transportation revolution has changed where people live, how businesses operate, and how economies grow,” said co-founder Keller Cliffton. “Together with Uber, we’re taking the next step toward building a world where getting what you need is as fast and effortless as sending a text, no matter where you are.”

Read more here about it!

Matic launches Cues, voice & gesture control for home robots! 🏠

Matic has launched Cues, adding voice and gesture control to its home cleaning robot. The pitch is simple: spill coffee, point at it, say “Hey Matic, clean this,” and the robot hears you, sees the gesture, locates the spot in 3D, and goes to clean it on its own.

The other commands work similarly. Say “Hey Matic” and it locates your voice, turns, and looks at you. Say “Hey Matic, follow me” and it walks behind you. Say “Hey Matic, go clean the living room” and it navigates there using the house map it already has. Matic says the system understands 75 languages and is simple enough for a five-year-old or an eighty-year-old to use.

On the cleaning side, Matic claims 8x the airflow of competitors, with specialised algorithms for rugs, corners, toekicks, and mopping, and says performance keeps improving through software updates. The company raised $115 million and spent nine years building the robot. Around 13,000 families use it, and WIRED gave it a 10/10, which the magazine hasn’t handed to a piece of hardware in a decade.

Matic is offering a full refund if buyers don’t love it after six months. To mark the launch, the company is cleaning 300 homes with Matic across San Francisco and New York City.

Read more here about it!

Unitree prices IPO at $9B, first mainland-listed humanoid maker
→ 2,760x oversubscribed, 150.8 yuan/share raising $900M for R&D and manufacturing → Why humanoid sales (867.8M yuan) overtaking quadrupeds signals market maturity: real revenue, real profits, not speculation

Dyna Robotics’ Dyna-2 proves human-to-robot scaling law
→ 1M+ hours egocentric human video pre-training, zero-shot performs 39 robot tasks across embodiments
→ What monotonic scaling from 1K to 1M hours demonstrates: lockbox tasks unsolvable at 100K hours, 90% success at 1M

HII commits $900M to Path Robotics and GrayMatter over 7 years
→ Autonomous welding, sanding, grinding across Navy shipbuilding (carriers, subs, destroyers, unmanned vessels)
→ Why performance-based agreements with tech readiness gates signal procurement maturity vs R&D contracts

Nucleus Robotics deploys in operating factory in <90 days
→ Hardware-independent control software enabling multi-vendor humanoid coordination layer
→ Why intelligence layer (perception, planning, real-time adaptation, fleet coordination) matters more than locking to single vendor

Niantic/NVIDIA/Flexion compress months of adaptation to days
→ Photorealistic office scan enables RGB-only policy training with zero real-world fine-tuning
→ What real2sim photorealistic 3D Gaussian splats matching collision meshes reveal: visual and physics layers must agree or miscalibration transfers to hardware

Unitree prices IPO at $9 billion valuation in China! 🇨🇳

China’s Unitree has priced its Shanghai initial public offering at 150.8 yuan ($22.34) per share, valuing the company at around 61 billion yuan ($9.04 billion), in a deal that would make it the country’s first mainland-listed maker of humanoid robots. In a filing to the Shanghai Stock Exchange, Unitree, also known as Yushu Technology, said it was seeking to raise 6.1 billion yuan (roughly $900 million). The Hangzhou-based company is selling 40.45 million new shares, or 10% of its enlarged share capital, on Shanghai’s STAR Market, with Chinese AI company DeepSeek among the strategic investors. IPO subscriptions open August 10, with proceeds going toward robot software and hardware, new products, and building a manufacturing base.

The numbers in the prospectus are extraordinary for a hardware company. Overall revenue more than quadrupled to 1.7 billion yuan in 2025, with humanoid robots generating 867.8 million yuan in sales and overtaking four-legged robots as Unitree’s largest business. The company is profitable, and from 2023 to 2025 it sold 33,294 quadruped robots and 5,632 humanoids. Founded in 2016 by Wang Xingxing with cheap robot dogs, Unitree has completed an extraordinary 10-year arc to become the world’s first publicly listed humanoid robot manufacturer. Investor appetite has been staggering: the offering was reportedly oversubscribed by 2,760 times, with demand during the offline subscription process equivalent to 71.46 billion shares.

The same week as the IPO, Unitree quietly dropped a new low-cost modular humanoid starting at $4,290, with 31 degrees of freedom and flexible deployment. For context, Tesla’s Optimus is estimated at $20,000-$30,000. That pricing underscores how aggressively Chinese manufacturers are driving down the cost of humanoid hardware, even as the market runs hot on real revenue, real profits, and real products rather than pure speculation.

But the geopolitical backdrop is impossible to ignore. The IPO comes as the United States and China ratchet up trade and technology tensions, with Washington tightening Chinese access to U.S. technology and markets, including new restrictions on foreign-made humanoid and four-legged robots—and Beijing responding with export curbs and sanctions. U.S. sales accounted for 13.3% of Unitree’s revenue last year, and while its existing robots have U.S. approvals, future models could be barred from sale there entirely.

Read more here!

Dyna Robotics releases a world-action model, Dyna-2! 📈

Dyna Robotics has introduced Dyna-2, a new flagship world-action model (WAM) pre-trained on more than one million hours of egocentric human video—roughly 170 years of continuous waking experience. The model establishes several novel scaling laws, most notably proving for the first time that a human-to-robot transfer scaling law exists: more human data in pre-training improves prediction on robot data the model has never seen, with zero robot data used during pre-training. The core bet is that the right source of pre-training data is sensorized video of humans performing everyday tasks, which already exists at effectively unbounded scale and carries exactly what a manipulation policy needs to learn, how scenes evolve, how objects respond to contact, and how hands interact with them.

Dyna-2 is a single generative model that can denoise future video and future actions jointly or separately, built on a video-diffusion backbone. Trained on a corpus of over one million hours (43.8 million clips, 97,160 unique task instructions, 9,917 distinct objects), the model shows that held-out prediction improves monotonically and follows a clean power law all the way to the million-hour scale. Crucially, that same trend transfers across the embodiment gap: evaluated zero-shot on 39 robot tasks across two embodiments, performance predictably improves as purely human data grows, with an inflection point emerging between 10k and 100k hours of pre-training.

The scaling law carries through to real robots in post-training. With just a few hours of robot data—and no robot data during pre-training—Dyna-2 performs tasks across bi-manual parallel-jaw arms, semi-humanoid, and dexterous hand platforms, with mean normalized performance rising from 20% to 53% as pre-training scales from 1k to 1M hours. Some tasks require a threshold amount of pre-training before becoming solvable at all: Lockbox Key Turning was never solved up to 100,000 hours, but at one million hours it succeeded 90% of the time. In one striking example, just 13 minutes of teleoperation data was enough to fine-tune the model to open bottle caps using two five-fingered robot hands.

Read more here!

HII signs up to $900M agreement with Path Robotics and GrayMatter Robotics! ⚓

Huntington Ingalls Industries (HII) has announced long-term performance-based production agreements with GrayMatter Robotics and Path Robotics, intending to award up to $900 million in total shipbuilding work across seven years. Path Robotics provides robotic welding technology, while GrayMatter offers robotic sanding systems.

Both companies are part of HII’s High-Yield Production Robotics (HYPR) program, launched in April, which aims to accelerate advanced, adaptive automation in the fabrication of both crewed and uncrewed naval platforms. The funding is contingent on the two companies meeting clearly defined technology and manufacturing readiness, as well as performance milestones. With the agreement, HII plans to accelerate development and deployment of physical AI across U.S. Navy shipbuilding programs, including aircraft carriers, submarines, destroyers, amphibious ships, future frigates, and unmanned surface vessels.

The agreements include two stages. In the development stage, both companies will partner with HII to develop, validate, and qualify high-precision production techniques applicable across autonomous welding, grinding, blasting, painting, assembly, and inspection, then integrate them into an autonomous production line. For instance, Rove pairs Path Robotics’ Obsidian physical AI model with a quadruped robot to bring autonomous welding beyond fixed cells, deploying directly onto large assemblies and immovable structures.

In the delivery stage, HII will begin sourcing shipbuilding work from both companies through the new line, contingent on favorable cost, schedule, and quality performance, starting with small steel structures and growing to include units and modules. In 2026, HII plans to outsource more than 2.5 million hours of shipbuilding work, a 30% increase from 2025, while expanding its structural assembly network to enable more work to be completed outside the shipyards before final assembly.

Read more here!

Robotics Maps - Odense 🇩🇰

Danish city of robotics! 🇩🇰

When I arrived in Odense in 2024, the first thing I saw was robots. Lots of them.

Cleaning robots greeted me at the station, an autonomous one trailing advertising banners rolled past, and a few minutes later I spotted UR robots playing tic-tac-toe in a shop window.

Odense is one of Europe’s strongest robotics ecosystems, especially for collaborative and industrial robots.

Nucleus Robotics exits stealth with 90-day factory deployment platform! 🥷🏻

Nucleus Robotics has emerged from stealth, with founder Melvin Schwarz announcing via X that the German company deployed humanoid robots inside an operating factory in under 90 days. The announcement serves as the first public test of Nucleus Robotics’ core pitch: software that helps humanoid robots adapt to existing industrial sites without rebuilding production lines.

Nucleus Robotics describes its core technology as an intelligence layer spanning perception, task planning, real-time adaptation, and fleet coordination for industrial humanoid operations. Registered in Gütersloh in February 2026, the company’s framework covers imitation-based training, real-time monitoring, and teleoperation capabilities for handling complex exception cases. By focusing on hardware-independent control software, Nucleus Robotics aims to enable manufacturers to add new robotic systems without locking factory workflows into a single hardware vendor.

Schwarz previously co-founded VisionAI before assembling a European robotics team with experience spanning major automation enterprises, research institutions, and aerospace organizations. The startup is now positioning its operator software layer to capture deployment management as manufacturers increasingly test humanoids from various global hardware suppliers, a bet that the future of industrial robotics will be multi-vendor, and that whoever controls the coordination layer will hold a valuable position in the stack.

Read more here about it!

A humanoid learns to walk a real office by training entirely inside a photorealistic scan of it! 🏢

A collaboration between Niantic Spatial, NVIDIA, and Flexion Robotics has demonstrated something striking: a humanoid can learn to navigate a real office by training entirely inside a photorealistic scan of that office, then walk it with no real-world fine-tuning, using RGB camera vision instead of a depth sensor. The weight of the story is in the economics of deployment. Reinforcement learning is the scalable way to teach a humanoid to navigate, but a robot doesn’t get many second chances in the real world,a misjudged gap or a collision with a glass door is expensive and can damage hardware. So training happens in simulation, which raises the question: how do you know a policy will hold up at a specific site before the robot arrives?

Simulated worlds are usually randomized, untextured geometry, because depth is easy to simulate. But that leaves the robot blind to materials, semantics, and what things actually are. RGB is what humanoids already carry and encodes both geometry and meaning, but you can only train an RGB policy if the simulated imagery genuinely looks real. The unlock: if the training world is a faithful reconstruction of the actual deployment site, RGB becomes trainable—and the policy specializes in the building it will work in. Deploying to a new environment has traditionally taken months of on-site adaptation; this pipeline compresses that to days.

Each company owns a distinct piece. Niantic Spatial handles “real2sim”, turning a real place into a high-fidelity, simulatable copy from a few minutes of video off an off-the-shelf 360° camera, no LiDAR or specialist workflow. Crucially, the visual layer (a photorealistic 3D Gaussian splat) and the physics layer (the collision mesh) come from one reconstruction, so they agree by construction. Most pipelines build geometry separately and try to register the two—and if a rendered wall and a collision wall disagree by a few centimeters, the robot carries that miscalibration onto real hardware. NVIDIA handles training: the splat and mesh export as a single USDZ matching NVIDIA’s NuRec spec, loading directly into Isaac Sim and Isaac Lab, gravity-aligned and collider-ready with no manual conversion. Flexion handles “sim2real,” training an RGB-only navigation policy inside the reconstruction using domain randomization plus large-scale offline-trained image encoders that produce robust features on real imagery.

Read more here about it!

Google DeepMind launches Gemini Robotics 2 for whole-body control
→ Vision-language-action model now controls full humanoids (feet to fingertips) vs previous upper-body table-top only
→ What same checkpoint controlling three different embodiments (Apollo 2 SharpaWave, Apollo 2 Inspire, Franka Duo Robotiq) reveals about generalization across platforms

Humanoid releases KinetIQ Ascend with real-world reinforcement learning
→ Why scaling trend suggests path to 100% reliability: robot performance improves predictably as training time added, similar to LLM compute scaling
→ Fleet learning loop: every deployed robot becomes part of continuous improvement system through supervisor interventions

Generative Bionics unveils Gene.01 with full-body tactile smart skin
→ Distributed multimodal sensing detecting touch, proximity, force, temperature enabling force-aware task teaching
→ Why physics-native AI optimizing hardware/mechanics/sensing together outperforms vision-only systems: infers physical forces for complex manipulation

Procore acquires DroneDeploy for $845M, creating construction AI
→ Combines 400M photos + 126M drawings from Procore with DroneDeploy’s 20 trillion sq ft visual data + 100K labeled safety issues
→ Why multi-modal perception (drones, ground robots, cameras) enabling autonomous action replaces manual jobsite inspections

Reimagine Robotics emerges from stealth with "learn on the job" robots
→ Google DeepMind's Scholz + colleagues launch London/Sydney startup reducing behavior prototyping from 1 day to 10 minutes
→ Why "monkey-see, monkey-do" approach (show, correct, repeat) eliminates specialist programmers vs fixed task programming

Gemini Robotics 2 brings whole body intelligence to robots! 🧠

Google DeepMind has introduced Gemini Robotics 2, the intelligence layer powering the next generation of adaptable robots. The major advance unlocks intelligent whole-body control, advanced dexterity, and multi-robot collaboration. Where previous models controlled only a humanoid’s upper body for table-top tasks, Gemini Robotics 2 can now control entire humanoid robots from feet to fingertips, enabling a robot to walk, crouch, stretch, and manipulate objects to clean up a cluttered room, and even team up with other robots to finish the job faster.

The release comes as three models. Gemini Robotics 2 is the most advanced vision-language-action model (VLA), converting vision and language input into motor control, capable of controlling full humanoids and other bi-arm robots. Gemini Robotics ER 2 is the most capable embodied reasoning model—a vision language model that acts as the robot’s high-level brain, enabling robots to communicate with humans, understand the physical world, and plan multi-step tasks lasting several minutes. Gemini Robotics On-Device 2 is the most efficient VLA, optimized to run locally on robotic devices with fast adaptation to new embodiments.

Gemini Robotics 2 unlocks a new level of physical dexterity across different end effectors. The model can control the five-fingered, 22 degree-of-freedom SharpaWave hand on the Apollo 2 robot to complete delicate actions like tying knots or sealing a ziplock bag, and can also operate standard two-fingered parallel grippers on a Franka Duo platform for complex tasks like tight packing. The same model checkpoint controls three different embodiments, the Apptronik Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and the Franka Duo with a Robotiq gripper, though DeepMind notes multi-finger dexterous manipulation remains challenging.

Safety remains foundational to the research. DeepMind is introducing ASIMOV-Agentic, a new benchmark for agentic safety orchestration and uncertainty resolution that measures the reasoning agent’s ability to refuse unsafe tool calls and proactively request human intervention when uncertain. With enhanced embodied reasoning, Gemini Robotics ER 2 is the company’s safest robotics model to date in safety constraint following and human proximity benchmarks, better detecting when humans are nearby and bringing the robot to a safe stop if someone approaches too closely.

Read more here!

Humanoid released its new platform KinetIQ Ascend 🦿

Humanoid has revealed how it uses reinforcement learning (RL) to improve its robots’ performance through experience, rather than relying solely on human demonstrations. Currently, imitation learning is the primary way to teach robots manipulation skills, an approach that works well for impressive demos but doesn’t create robots that can outperform people or operate reliably in real-world environments.

The company’s new approach, KinetIQ Ascend, extends KinetIQ, the proprietary AI framework behind its robots, with the ability to learn from trial and error. Humanoid believes this is the first published demonstration of end-to-end, vision-based RL running on production humanoid manipulation systems in real-world deployment conditions.

One of the key findings is that robot performance improves in a predictable way as more robot training time is added, similar to how large language models improve as more compute and data become available. The observed scaling trend, supported by simulation experiments, suggests the method scales all the way to 100% reliability. The results were demonstrated on several industrial tasks. In a machine-feeding application where a robot picks steel bearing rings from a bin and places them onto a conveyor, RL increased throughput by 42%, with the robot performing at 1.5x the speed of the human demos it originally learned from. In a second task picking items from a cluttered tote and handing them to a person, throughput increased by 85% and success rates improved from 80% to 98%.

The robots achieved these improvements after only a few days of learning. The company also uncovered two findings that exceeded expectations. First, improving only the most difficult part of a workflow can improve the performance of the entire task. Second, the robots appeared to generalize what they learned. After training on a single object type, performance improved on other objects the robots had never seen during training. Humanoid describes its approach as building a “capability factory,” a system designed to rapidly create, improve, and scale robot skills across industrial environments.

The long-term vision goes even further. Because the same training framework can run on deployed robots, every robot could eventually become part of the learning system, meaning fleets would continuously improve in a feedback loop: better robots generate better data, leading to even better robots over time. When a supervisor steps in to correct or assist a robot on-site, that intervention automatically becomes a training signal—turning routine operation into continuous improvement without any additional data collection effort. The humanoid race is becoming a question of scale, and Humanoid’s research suggests real-world RL may be one of the technologies that helps robots perform thousands of tasks every day with speed, reliability, and consistency.

Read more here!

Italian humanoid with smart skin for safe collaboration! 🫆

Italian deep-tech company Generative Bionics has unveiled significant updates to Gene.01, a humanoid robot designed for industrial applications. Developed in just six months, the robot combines full-body tactile sensing with a physics-native AI system that integrates its body, mechanics, and intelligence to better perceive, interpret, and respond to the physical world.

Unlike conventional humanoids that primarily rely on vision, Gene.01 features a distributed multimodal “smart skin” covering its body. The tactile system detects touch, proximity, force, and temperature, allowing the robot to sense people before and during physical contact, enabling more natural interactions while improving safety in shared workspaces. The force-sensing capability also allows workers to teach tasks by physically guiding the robot and demonstrating the required force for operations such as gripping or handling objects. This addresses limitations of vision-only learning systems, which often struggle to infer the physical forces needed for complex manipulation.

At the core of the platform is Generative Bionics’ proprietary physics-native AI, which jointly optimizes the robot’s hardware, mechanics, sensing, and motion intelligence rather than treating them as separate systems. The AI uses digital models of both the robot and its human partner to optimize movement, ergonomics, balance, and walking stability. The architecture is based on research published in Nature Machine Intelligence describing a human-aware embodied intelligence framework originally developed through the ergoCub humanoid project.

Generative Bionics is also releasing Gene.01’s digital twin as open source, making it available through standard software distribution ecosystems including PyPI, conda-forge, and the official ROS build farm, giving Physical AI developers a common foundation for simulation and testing before field deployment. Rather than marketing Gene.01 as a one-size-fits-all humanoid, the company has designed it as a scalable platform customizable for manufacturing, logistics, inspection, healthcare, and public safety. Its first industrial deployment is being developed with Italian shipbuilder Fincantieri, where Gene.01 is being adapted into a welding humanoid for shipyard operations, while a partnership with German actuator specialist Synapticon will build out a European supply chain for future production.

Read more here!

Robotics Maps - Austin 🤠

Austin has quickly become one of the fastest-growing robotics hubs in the US.

Thanks to its mix of AI talent, advanced manufacturing, and a strong startup culture.

The ecosystem is supported by The University of Texas at Austin, which produces top engineers in robotics, AI, and computer science. Around the university, a growing number of startups are developing autonomous systems, drones, and humanoid robots.

Austin has also attracted major technology companies. Tesla operates its Gigafactory near the city and is developing the Optimus humanoid robot, creating a large pool of robotics, AI, manufacturing, and automation talent.

Procore Technologies acquires DroneDeploy for $845M! 🏗️

Procore Technologies, a construction management software provider, has announced it will acquire DroneDeploy for approximately $845 million in cash. DroneDeploy offers a robotics and visual intelligence platform for the built world that has been used on over 3 million job sites across more than 180 countries.

DroneDeploy’s visual intelligence technology bridges the physical world—including active construction sites and operational assets, with the digital world, providing real-time visibility into daily operations. The company provides perception through three-dimensional ground and aerial imaging, spanning drones, ground-deployed robots, and mobile, fixed, and wearable cameras. When combined with Procore AI, that captured visual data can be translated into observations that drive autonomous action within the Procore platform.

The companies said the combination of their technologies will enable users to see, understand, and act within Procore, replacing manual jobsite inspections with multi-modal perception capabilities. A range of cameras, drones, and robots can regularly evaluate construction sites and automatically initiate appropriate responses.

Once completed, the transaction will bring together two extensive, domain-specific datasets: Procore’s record of construction decision-making, encompassing nearly 400 million photos, over 126 million drawings, and over 10 million RFIs, submittals, and inspections in the past year alone—and DroneDeploy’s visual record of physical builds, spanning approximately 20 trillion square feet of visual data, tens of millions of user-generated annotations, and more than 100,000 labeled safety issues. Procore plans to integrate capabilities including real-time perception and advanced reasoning to deliver “digital co-workers” for the field and back office. The transaction is expected to close later this year, subject to customary closing conditions and regulatory approvals.

Read more here about it!

Reimagine Robotics emerges from stealth with robots that ‘learn on the job’! 🐒

Reimagine Robotics has emerged from stealth, developing technology it says anyone can train and use, enabling robots to learn on the job. “A useful robot should be able to learn from the person doing the work,” stated Jonathan Scholz, co-founder and CEO. “They should be able to show it a task, put it right when it makes a mistake, and move on to the next problem. That is what it means for a robot to learn on the job. It’s a process we call ‘monkey-see, monkey-do.’” Scholz founded Google DeepMind’s Applied Robotics team in London and led it for seven years, later co-founding Reimagine Robotics in April 2025 alongside former colleagues Oleg Sushkov, Akhil Raju, and Misha Denil, with headquarters in London and Sydney.

Instead of requiring specialist programmers whenever a task or production process changes, Reimagine Robotics lets workers show the robot what to do, watch it attempt the task, and then correct it on the spot. “A robot should arrive with the attitude of a new colleague: ‘How can I help? What do you want me to do?’” said Scholz. People remain central to the model throughout deployments. “For us, this is not about taking people out of the process. A robot that learns on the job depends on people. The worker identifies the bottleneck, shows the robot how to help, and corrects it until it is useful. I think of it more as a tool to amplify human labor.”

The company’s first phase was funded by pre-seed financing from Fly Ventures, firstminute capital, and a number of angel investors, and it has already deployed its robots in advanced manufacturing and electronics disassembly facilities. At a made-to-order plastics business, Reimagine Robotics trained its robots to tend 3D printers overnight by removing print beds, operating latches, and pressing controls—and the customer’s own team used the platform to automate additional stages including washing, curing, and drying. In a separate deployment recovering valuable critical materials from used hard drives, the company worked with process engineers to develop a three-robot disassembly cell combining robots and people to optimize the workflow in real time.

During that project, Reimagine Robotics said it reduced the time required to prototype and test a new robot behavior from approximately one day to about 10 minutes, allowing teams to propose a new use for a robot and test their ideas almost immediately. “This last year was about building a core product and a team, and working with customers to test the platform,” said Scholz.

Read more here about it!

Humanoid raises $152M Series A at $1.35B valuation
→ Why Schaeffler partnership involving thousands of robot deployments signals commercial validation beyond funding
→ Q4 beta deployments across manufacturing, logistics, retail with Bosch as contract manufacturer

Walden Robotics launches from stealth at $1.1B valuation
→ Toyota and Deviation Capital co-lead $300M, general-purpose robots in production at Toyota plant since February
→ What Diffusion Policy and Large Behavior Models (LBMs) demonstrate: continuous learning through real-world practice

mimic robotics unveils M1 hand and U1 wearable exoskeleton
→ 15 active DOF matched to human hand, tendon-driven, 25kg payload, <0.05 Nm backdrive torque enabling 50g weight sensing
→ Why keeping morphology constant across learning pipeline matters: no cross-embodiment gap between human demos and robot execution
→ Data pyramid approach: video (base) → wearable (middle) → robot teleoperation (top quality)

Sunday Robotics previews ACT-2, achieves 99.1% zero-shot laundry success
→ 785 autonomous attempts across 9 garment types, unseen homes, no per-home adaptation using single fixed model
→ What single-demonstration learning generalizing to unseen environments reveals: post-training gains transfer to real-world performance

Humanoid raises $152 million Series A at $1.35 billion valuation! 💰

UK-based robotics startup Humanoid has raised $152 million in a Series A funding round at a post-money valuation of $1.35 billion, as investors pour money into companies developing robots that can augment human labor. The round was led by Prime Movers Lab, with participation from Schaeffler, Bosch, Fubon Financial Holding Venture Capital, and Aglaé Ventures, bringing Humanoid’s total funding to $270 million.

The proceeds will fund development of the company’s next-generation platform, expansion of its proprietary AI software platform KinetIQ, commercial deployments with industrial customers, and the start of mass manufacturing of wheeled robots. Humanoid plans to begin deploying beta robots at customer sites in the fourth quarter of this year across manufacturing, logistics, retail, and other industrial sectors.

The company said it has partnerships with SAP, Nvidia, Bosch, and Siemens, and recently signed an agreement with Schaeffler involving the planned deployment of thousands of robots at manufacturing facilities. Bosch said it would serve as Humanoid’s contract-manufacturing partner while providing technical expertise in hardware design, production, and supply chain operations.

The financing signals growing investor appetite for “physical AI” companies that combine artificial intelligence with robotics, as manufacturers and logistics companies increasingly look to humanoid robots to address labor shortages and automate repetitive industrial tasks.

Read more here!

Walden Robotics launches with $300 million to put general-purpose robots to work today! ⚙️

Walden Robotics, a full-stack Physical AI company building and deploying general-purpose robots that continuously learn and improve while performing real work, has launched out of stealth with $300 million in funding at a $1.1 billion valuation. The round is co-led by Toyota (Toyota Motor Corp, Toyota Invention Partners, and Toyota Ventures) and Deviation Capital, with participation from NVIDIA, Boeing, AE Ventures, Samsung Ventures, Prologis Ventures, CoreWeave Ventures, and more.

The Cambridge, Massachusetts-based company deploys robots into real production environments in manufacturing and logistics, doing useful work side-by-side with people from day one. Walden enables skilled team members to delegate difficult-to-automate tasks so they can focus on problem-solving, judgment, and craft. Its robots are useful across industries, with strategic partners in automotive, aerospace, semiconductors, electronics, logistics, and life sciences, meeting fast-growing needs from industries facing labor shortages, demographic shifts, competitive pressure, and increased demand.

Walden’s product builds on a decade of foundational AI and robotics research the team helped pioneer, including Diffusion Policy and Large Behavior Models (LBMs), the frontier model class that powers Walden’s robots and lets them quickly learn new tasks and continuously improve through real-world practice. The company launched out of Toyota Research Institute in January 2026 and began working with customers across multiple industries immediately. Since February, Walden’s general-purpose robots have been doing useful work in production at a Toyota plant in North America, moving from first pilot to real work in under two months.

Read more here!

mimic robotics unveils their hand M1 and wearable U1! 🤌🏼

Swiss company mimic robotics has announced the mimic hand M1, a highly backdrivable, tendon-driven robotic hand designed and built in-house for industrial automation. Alongside it comes the mimic wearable U1 (”umimic”), an exoskeleton that captures human demonstrations directly matched to the kinematics of the M1. Together with a fully in-house software stack, the systems form an integrated platform for general-purpose dexterous manipulation.

The core bet is keeping morphology constant across every phase of learning by running the entire pipeline on hands. Traditional two-finger grippers create a cross-embodiment gap: if you pre-train on human video and then deploy a gripper, the end effector is different and the steps a human takes to manipulate an object aren’t the steps a gripper needs. mimic never introduces that gap. The company organizes its data around a “data pyramid”, human video at the base (lower quality, massive scale), wearable device data in the middle, and robot teleoperation and deployment data at the top (highest quality, collected on its own robots).

The M1 has 15 active degrees of freedom across 21 joints, matched to the most functional degrees of freedom of the human hand including abduction and an opposable thumb. It delivers over 25 kg steady-state payload in a cylindrical power grasp, 25 N fingertip force, backdrive torque under 0.05 Nm (low enough to sense weights as light as 50 g through motor current), force estimation sensitivity under 0.1 N, and fingertip position accuracy of ±0.18 mm. Rather than packing motors into the palm, mimic relocates actuators to the forearm and drives the fingers through tendons routed over bearings and pulleys, replacing sliding friction with rolling friction for predictable, consistent behavior across units and cycle counts — essential for learned policies to transfer between hands.

The U1 wearable matches both the kinematics and contact geometry of the M1, with a rigid linkage mechanism enforcing 1:1 correspondence on every degree of freedom. Tactile sensors, joint encoders, and wrist cameras are placed in one-to-one correspondence with the robot, so demonstrators experience the same sensing modalities the robot will. The tradeoff: geometry is fixed to the robot rather than adapting to every user’s hand, but in return demonstrators feel contact forces through their own fingers and work at nearly natural human speed without control latency or retargeting.

Read more here!

Experiencing the Sound of Science live at Siemens Transform UK! 🎼

I had the chance to attend Siemens Transform UK in Manchester this week, and it was amazing to experience Siemens’ Sound of Science acoustic digital twin live, meeting Cedrik Neike and Stephan Frucht in person and hearing, quite literally, what the future of acoustic simulation sounds like.

For the first time in the UK, Siemens brought its Sound of Science application to a British venue: Hallé St Peter’s in Manchester, home to the Hallé Orchestra’s rehearsals and recordings. What makes this special is that Siemens has built a full acoustic digital twin of the space. It draws on the building’s layout and architecture, combined with a dry studio recording of a musical piece, to create a life-like, physics-based simulation of what a concert will sound like before a single note is played in the real room. Standing there and hearing a performance “rendered” through the actual acoustics of the hall, before any rehearsal, was genuinely striking.

Conductors and venue operators can explore different orchestra layouts and evaluate their acoustic impact without staging full rehearsals, cutting rehearsal times and logistics costs. The application even lets you experience how sound varies throughout the venue, and can transport you back to the building’s original church acoustics through an immersive VR experience. Hallé St Peter’s, a former church in the heart of Ancoats, was a fitting choice: it represents the “church” category among the four key venue types Siemens works with (concert halls, churches, opera houses, and festival halls).

What I found most compelling is the bigger story behind it. This is the same digital twin technology Siemens deploys across industry, modeling products, factories, buildings, and processes virtually before making changes in the physical world, applied here to a cultural setting. And Siemens’ connection to music runs deep: Ernst von Siemens acquired Deutsche Grammophon and built it into the world’s leading classical label, and the company remains closely tied to the arts through the Siemens Arts Program, which empowers young people in the arts, not just engineering.

Talking with Stephan Frucht, artistic director of the Siemens Arts Program, that “arts + tech” bridge really came through. As he put it, the digital twin technologies already helping industry test ideas and optimize resources now bring those same advantages to the concert hall.

Technology changes fast, but when you can anchor it to something as beautiful and human as music, it becomes something people can trust and understand, and that’s exactly what made experiencing it live so memorable.

Read more here about it!

Sunday previews ACT-2, hitting 99% zero-shot success! 👕

Sunday Robotics has previewed ACT-2, which it calls the first robotics model to achieve reliability by unifying broad generalization with high performance. The headline result: 99.1% zero-shot success on laundry folding across diverse, unseen home environments, measured across 785 autonomous attempts spanning 9 garment types. ACT-2 builds on ACT-1, introduced in November 2025, which demonstrated long-horizon mobile manipulation and generalization to unseen homes. Where ACT-1 expanded what robot foundation models could do, ACT-2 focuses on how reliably those capabilities hold as conditions change.

The central breakthrough is that reliability gains from rapid post-training on limited in-house data generalize to real, unseen homes. Sunday found that scaling pretraining sharply narrows the generalization gap between in-domain and out-of-domain performance, meaning in-house improvements become a reliable predictor of real-world performance. The company also reports being the first to demonstrate that an end-to-end model can learn a new behavior from a single demonstration and generalize it to unseen environments, they post-trained four copies of the same pretrained model, each on one demonstration of a different folding technique, and all four successfully executed their new techniques on a held-out garment.

The evaluation covers T-shirts, long-sleeved tops, polos, sleeveless tops, blouses, pants, leggings, and shorts across sizes from XXS to 8XL, with garments starting in baskets, piles, on beds, or on the ground, arbitrarily oriented and naturally crumpled, all with zero per-home adaptation and the same fixed model checkpoint. Fold quality averaged 4.72 out of 5 stars across 778 completed folds, with 98.3% rated four stars or better and 73.8% receiving a perfect score. Median completion time was 2 minutes 13 seconds. Emergent behaviors included retrieving clothes from the floor, replanning under disturbance from children and adversarial interference, and repositioning its mobile body to handle everything from baby clothes to 53-inch towels.

Read more here about it!

1X equips NEO with human-level dexterous hands
→ 25 actuated DOF, 5-to-15:1 gear ratios, tendon-driven enabling force transparency and tactile sensing
→ What backdrivable joints revealing grip pressure, contact location, shear demonstrates: hands simultaneously motors and sensors

LimX Dynamics prepares IPO at $2.21B valuation
→ Chinese humanoid startups rushing public listings as investment hits $6.95B in Q2 alone (6x YoY)
→ Why “0 to 1” innovation crossed, next barrier is product-market fit not technology → 100+ Chinese humanoid companies competing, Unitree fast-tracked Shanghai listing, 50K humanoid shipments forecast

Holiday Robotics raises $105M Series A, largest Korean humanoid round
→ Friday robot built around 64 DOF (40 in hands) with tactile sensing, force-aware control, wheel-based mobility
→ What Vision-Language-Skill (VLS) stack reveals: vision/language define goals, reusable skills execute with full deployment infrastructure in-house

NVIDIA and Hugging Face integrate Isaac GR00T 1.7 into LeRobot
→ Open reasoning VLA model + Isaac Teleop framework for human demonstrations now in open source library
→ What 350K trajectories, 57M grasps, 15M dataset downloads demonstrate about democratizing robot development

Human-level hands? 1X just gave humanoid robot NEO something close! 🪬

1X has released the hands for its NEO humanoid robot, showcasing unprecedented speed and precision. The headline spec: 25 degrees of freedom, all of them actuated, and tendon-driven motion, resulting in what 1X calls near human-level dexterity, strength, and reliability. The hands have enough strength to lift a 20-pound kettlebell yet enough delicacy to pick grapes off their stems, install a light bulb, and pick up a screw off the floor.

Most robot hands are essentially “write-only”, you command the hand to move to a position, it goes there, and no data comes back. It doesn’t feel anything the way a human feels things, thanks to aggressive gearing (often 100-to-1 or 200-to-1) that swallows any contact force feedback before it reaches the motor. Neo’s hand runs at dramatically lower gear ratios, roughly 5-to-1 to 15-to-1, using tendons to pull the fingers. The result is that every single joint is simultaneously a motor and a sensor—push on a finger and it gives (it’s backdrivable), and it reports exactly how hard you pushed. 1X calls this “force transparency”: force flows out, information flows back, and the robot knows what’s happening at its fingertips.

There’s also a skin layer doing real work: high-resolution tactile sensing across the fingertips that measures pressure, contact location, and shear, the sideways force that tells the hand something is about to slip. So when an object starts sliding out of its grip, Neo can theoretically feel it and re-grip in real time, just like humans.

Uniquely, these hands are washabl, IP68 sealed, just like a phone, so Neo can immerse them in water without any problems. That means Neo can fry an egg or chop vegetables and then clean up after itself, rather than spreading mess around the house. 1X says the hands have been through millions of test cycles, with wrist joints proven past two million cycles under load.

As 1X points out, while legs are for getting a robot somewhere and the body houses the battery and compute, the hands are what do the actual work—folding clothes, putting away laundry, cooking, whatever the chore is. Notably, Neo will ship with more hardware capability than the AI currently knows how to use, meaning over-the-air updates should make each Neo progressively smarter over time.

Read more here!

Chinese humanoid startups are rushing to launch IPOs! 📊

Humanoid startup LimX Dynamics is getting ready to go public, just over four years after it was founded during the pandemic. “Listing is a must,” said founder Will Zhang, emphasizing the importance of timing, ahead of the company’s announcement that it had raised $200 million in a pre-IPO round valuing the startup at 15 billion yuan ($2.21 billion). Several overseas investors participated, including UAE-based Stone Venture, Italy-based GGG, and Germany-based Redstone VC. The startup said it is already preparing for its IPO, likely in Hong Kong, and is in a confidential phase of review.

Zhang compared the situation to Chinese electric car startups Nio, Xpeng, and Li Auto, which successively listed in the U.S. from 2018 to 2020. “Once the technology is mature, if [the company] doesn’t list, then like WM Motor, it may disappear,” he said. The urgency comes as China now has well over 100 humanoid companies, which fall under the national push for “embodied AI.” Reflecting a rapid surge in interest, investment in the sector hit 47.09 billion yuan ($6.95 billion) in the second quarter, more than double that of the first quarter, and up over six times versus the same period last year, according to industry data provider Xiniu.

China has fast-tracked approval for humanoid company Unitree to list in Shanghai, while Hong Kong processes applications from more than 500 companies across sectors. “With more industrial and collaborative robot companies potentially coming to IPO, competitive pressure is likely to persist,” Morgan Stanley said in a report, noting sector players DeepRobot and Leju that are looking to list soon. The investment firm forecasts 18% growth in China’s industrial robots market this year, and shipment of 50,000 humanoids.

LimX aims to create fully autonomous commercial service robots. The company said it will kick off a multi-year plan to ship thousands of humanoids to the Middle East, and is delivering its entertainment-focused Luna humanoid to customers in South Korea. To Zhang, the technology behind humanoid robots has already crossed the “0 to 1” line of innovating from scratch—the next barrier to entry lies in making a good product that meets users’ needs. Other backers in the latest round include Chinese precision parts company Lens Technology, IDG Capital, WestSummit Capital, Nio Capital, and Hefei Binhu Industry Development Group.

Read more here!

Holiday Robotics raises $105M in Series A funding! 🇰🇷

South Korea’s Holiday Robotics has raised $105 million in Series A funding to develop dexterous humanoid robots for industrial work. South Korean business newspaper Maeil Business reported that the round is the largest Series A investment for a South Korean startup and the largest single funding round for a South Korean humanoid robot company. Existing investors Stone Bridge Ventures, Atnum Investments, Intervest, and Spring Camp participated, alongside new backers including IMM Investment, SL Investment, KB Investment, Bon Angels Venture Partners, and Goodwater Capital. Financial institutions including Korea Development Bank and Industrial Bank of Korea also participated.

Founded in 2024, the company will use the capital to grow its research and engineering teams, strengthen manufacturing readiness, and continue developing Friday, its humanoid robot. The company’s approach starts with hands. Holiday said humanoid robots create industrial value through grasping, placing, adjusting, sensing contact, and repeating precise movements, tasks difficult to automate because they involve friction, weight, alignment, and surface variation. Friday has 64 degrees of freedom, with 40 dedicated to its hands. The robot is built around high-dexterity hands, tactile sensing across the hand, and force-aware control that allows it to feel contact and adapt to it. It uses a wheel-based mobile base to move through factories, a dexterous upper body for manipulation, and hot-swappable batteries to support full-shift operation. Safety is built into the hardware, with lightweight hands, fully backdrivable hand and arm joints, and 360-degree perception using multiple cameras and LiDARs.

Holiday is also developing its software and deployment stack in-house, built around what the company calls Vision-Language-Skill (VLS). In that approach, vision and language help define what a robot should do, while reusable skills define how the robot performs the task. The full-stack system includes Friday, Holiday Sim (a simulation environment), Holiday Lab (a system for turning demonstrations and failures into reusable skills), and OASys (a deployment system).

Read more here about it!

Robotics Maps - The Bay Area 🌁

This is the capital of robotics! 🌁

Silicon Valley is home to so many physical AI companies that you could spend a month visiting them and still not see even 10% of them (trust me, I tried).

So, the Bay Area is the world's leading ecosystem for robotics startups, bringing together top AI talent, top universities, experienced founders, and unmatched access to capital.

The region is anchored by Stanford University and University of California, Berkeley, two of the world's top universities for AI and robotics. They produce a constant stream of researchers, founders, and breakthrough technologies.

Check my LinkedIn post with all companies listed: CLICK HERE

NVIDIA and Hugging Face bring new models and frameworks for the open robotics community! 🔓

NVIDIA and Hugging Face are collaborating to bring the NVIDIA Isaac GR00T 1.7 open reasoning vision-language-action (VLA) model for humanoid robots and the NVIDIA Isaac Teleop framework to LeRobot—Hugging Face’s open source library for robotics, with NVIDIA Cosmos 3, a frontier model for physical AI, planned soon. Together, these integrations give developers a more accessible and standardized path for end-to-end robot development while driving innovation and collaboration across the open robotics community.

The integrations include NVIDIA Isaac Teleop (an open source framework for robot data collection that captures high-quality human demonstrations), NVIDIA Isaac GR00T 1.7 (the first open and commercially viable robot foundation model), and the upcoming NVIDIA Cosmos 3 (a frontier world foundation model that will help developers generate and augment robotics data and simulate scenarios when real-world data is limited).

These integrations build on a broader set of NVIDIA resources already connected to LeRobot. That includes the largest open source physical AI dataset—downloaded more than 15 million times—which features more than 350,000 real and simulated trajectories and 57 million grasps. It also includes NVIDIA Isaac Sim– and Isaac Lab-based simulation frameworks to help developers set up environments, generate robot data, and validate behaviors before moving to physical robots, plus NVIDIA Isaac Lab-Arena in the LeRobot Environment Hub for prototyping complex simulation environments.

Read more here about it!

Weave Robotics launches Isaac 1 mobile home robot
→ Why hybrid autonomy (autonomous + remote ops for edge cases) matters more than full autonomy claims
→ $7,999 or $449/month, California deliveries fall 2026, expanding nationwide 2027

Agility Robotics goes public via SPAC at $2.5B valuation
→ Digit deployed across 9 customer sites (Schaeffler, GXO, Toyota, Mercado Libre) with $300M+ in multi-year orders
→ Why $620M raised for production scaling + 30-customer pipeline signals transition from pilots to commercial deployment

Apptronik launches Apollo 2 in bipedal and wheeled configs
→ Expanded 90K sq-ft Robot Park in Austin generates high-quality training data for Google DeepMind’s Gemini Robotics foundation models
→ What modular configurations reveal: bipedal for adaptability, wheeled for safety compliance in existing operations

Flexion’s Reflect v1.0 achieves 90% success on complex missions
→ Custom VLM mission controller continuously monitors, reasons, and replans across multi-step tasks vs 38% without RL
→ Why natural-language instructions enabling real-time prompt changes matter more than task-specific programming

Weave Robotics is ready for house chores! 🧺

US-based Weave Robotics has released Isaac 1, a wheeled mobile home robot designed to automate everyday household chores such as tidying rooms and doing laundry. The robot can make beds, sort clutter, fold clothes, and put items away while operating autonomously in most situations. When it encounters tasks it cannot complete independently, remote human assistance steps in to ensure the job is finished. Isaac 1 succeeds the company’s earlier Isaac 0, a stationary autonomous robot launched in February 2026 that folds laundry in 30-90 minutes.

Unlike its predecessor, which remained stationary at a table, Isaac 1 features a motorized wheeled base that allows it to move throughout a home. The robot is equipped with a telescopic body that adjusts its height from about 3 feet (0.9 m) to 5 feet 9 inches (1.75 m), enabling it to reach furniture, beds, shelves, and laundry baskets while maintaining a compact profile when idle. Its upper body houses two robotic arms designed for household manipulation tasks, using multiple onboard cameras and AI-powered perception to identify objects, understand room layouts, and determine where misplaced items belong.

The robot operates autonomously by default but can transition to remote human teleoperation whenever it encounters situations beyond its current capabilities. Through this hybrid autonomy model, a remote operator can temporarily guide the robot using its onboard camera feeds to complete difficult tasks before returning control to the autonomous system. To address privacy concerns, the company says Isaac 1 includes hardware controls that physically disable its cameras when they are not required for operation.

Weave Robotics plans to continue expanding Isaac 1’s capabilities through over-the-air firmware updates, allowing the robot to perform additional household tasks as its AI models improve. The company is offering the robot for $7,999 or through a $449-per-month subscription. Customers can reserve a unit with a $250 deposit, with deliveries scheduled to begin in California in fall 2026 before expanding across the United States during 2027.

Read more here!

Agility Robotics to go public via SPAC in a $2.5B deal! 📈

Agility Robotics, the humanoid robotics startup that spun out of Oregon State University in 2015, plans to go public through a merger with special purpose acquisition company Churchill Capital Corp XI in a deal that values the company at roughly $2.5 billion. The transaction is expected to generate more than $620 million in proceeds, including about $200 million from a group of new and existing institutional investors.

Agility is best known for Digit, a bipedal robot that is being used across nine customer sites, including with Schaeffler, GXO, Toyota Motor Manufacturing Canada, and Mercado Libre. The company has enjoyed backing from high-profile tech companies and funds like Amazon, Nvidia, SoftBank Vision Fund 2, and DCVC. Now it plans to use the capital raised in the SPAC merger to increase production capacity of its next-generation Digit v5, fulfill existing orders, and expand to new and existing customers.

The company said it has secured more than $300 million in multi-year orders for the new model and a pipeline of more than 30 potential customers evaluating large-scale deployments. “Humanoid robots are poised to become a critical driver of productivity, supply chain resilience, and American technology leadership,” Agility CEO Peggy Johnson said. “With commercially deployed humanoids already operating in customer environments today, Agility is helping enterprises address labor shortages, improve efficiency, and safely integrate AI-powered automation into their operations.”

The combined company is expected to trade under the ticker symbol AGLT on a North American stock exchange that has not yet been announced. The move makes Agility one of the first pure-play humanoid robotics companies to pursue a public listing, marking a significant milestone as the industry shifts from pilot deployments toward commercial scale.

Read more here!

Apptronik releases Apollo 2, and a data collection and training facility! 🦿

Apptronik has made two major announcements. First, the company launched Apollo 2, its updated humanoid robot. Second, it opened its newly expanded Robot Park, its flagship data collection and training facility for humanoid robots in Austin, Texas. Apollo 2 comes in both bipedal and wheeled-base configurations, designed to learn real-world work through large-scale data collection across a wide range of tasks and environments. As part of Apptronik’s research partnership with Google DeepMind, the data Apollo 2 collects helps advance Gemini Robotics, Google DeepMind’s foundation models for robotics.

Their Robot Park enables the data collection that is fuel for that, and Apollo 2 is the machine that makes it possible.” Apollo is based on nearly a decade of development on 15 previous robots, including NASA’s Valkyrie. Apptronik started out of the Human Centered Robotics Lab at the University of Texas at Austin, has nearly 300 employees, and earlier this year raised $520 million, bringing total capital raised to nearly $1 billion.

By offering Apollo in modular configurations, Apptronik said it can optimize data collection across a variety of operational environments. The wheeled-base is designed to conform with existing safety standards for industrial mobile robots, allowing it to fit easily into existing customer operations, while the bipedal configuration provides maximum adaptability for complex environments. Everything the company is proving through the Apollo 2 platform is directly powering the development of its commercial product, Apollo 3.

Inside the newly expanded nearly 90,000-sq.-ft. facility in Austin, both bipedal and wheeled Apollo 2 systems learn across an array of customer use cases in logistics, manufacturing, retail, and other activities. Through a combination of teleoperation and autonomous execution, Apollo 2 continuously generates significant quantities of high-quality training data used to train and refine the Gemini Robotics AI models.

Read more here about it!

Robotics Maps - Massachusetts 🦞

The Massachusetts ecosystem is one of the world's leading hubs for robotics, combining top research, companies, and deep-tech investment.

At its core are Massachusetts Institute of Technology (MIT) and Harvard University, which produce world-class researchers, engineers, and robotics startups.

Many of today's leading robotics companies trace their roots back to these universities.

The region is also home to companies like Boston Dynamics, whose work has inspired advances in humanoid and mobile robots.

Flexion’s new model gives humanoid robots 90 percent success in complex missions! 🎯

Flexion Robotics has introduced Reflect v1.0, a robotics intelligence platform that enables humanoid robots to complete complex, multi-step tasks autonomously, without human intervention during execution. The company showcased the system through a workplace delivery demonstration in which a humanoid robot retrieved a snack parcel, navigated stairs and an elevator, unpacked the box, and stored the items in a drawer, all from a single natural-language instruction. Reflect v1.0 unifies reasoning, perception, motion, control, and runtime into one robotics platform.

Long-horizon autonomy remains one of the biggest challenges in robotics because even highly reliable navigation, manipulation, and perception systems can collectively fail when errors accumulate across many sequential tasks. Reflect v1.0 addresses this by integrating a custom vision-language model (VLM) that acts as a mission controller, continuously monitoring progress, reasoning about the environment, and replanning whenever necessary. The platform combines this mission controller with a motion layer that uses vision-language-action models trained on real-world data alongside reinforcement learning-based skills, while a real-time whole-body controller maintains balance, stability, and precise movements throughout the mission.

A key feature of Reflect v1.0 is its use of natural-language instructions rather than task-specific programming. Users can modify missions simply by changing the prompt, allowing robots to perform different tasks or even receive updated instructions while a mission is already underway. Reinforcement learning has significantly improved the platform’s reliability: in an internal evaluation involving a 16-step mission, a supervised fine-tuned model achieved only a 38 percent end-to-end completion rate, but after reinforcement learning was applied across multiple layers of the system, completion rates increased to 90 percent.

Read more here about it!

NVIDIA launches Halos for Robotics, industry’s first unified safety system
→ Full-stack architecture extending 18,600 engineering years of AV safety to physical AI
→ What NVIDIA IGX Thor + Holoscan Sensor Bridge + Halos OS reveal: safety built into every layer from compute to certification

Genesis AI unveils Eno, debut multi-purpose robot
→ Non-humanoid form factor with human-like hands, folding mechanism for “calm intelligence” and minimal footprint
→ Why full-stack in-house AI + optional transparency screen signal trust-first design for human-robot collaboration

Hyundai acquires SoftBank’s remaining 9.65% Boston Dynamics stake for $325M
→ Consolidates 100% ownership under Hyundai Motor Group, deepening robotics commitment
→ What put option exercise signals: transition from SoftBank’s diversified portfolio to automaker’s core automation strategy

AEON begins production tasks at BMW Leipzig plant
→ Humanoid now performing actual manufacturing work while training on future applications
→ Why shift from “can these operate in manufacturing” to “how quickly productive and scalable” marks new phase

NVIDIA announces the industry’s first safety system for Physical AI! 🛡️

NVIDIA has announced NVIDIA Halos for Robotics, the industry’s first full-stack, comprehensive safety system for robotics and physical AI that unifies AI compute and safety. The system extends NVIDIA Halos’ proven autonomous vehicle safety to robotics, giving machines that sense, decide, and act in the real world a single common safety architecture. Agility, a leading humanoid robotics company, is the first to use Halos for Robotics to build safety into its humanoids working for customers including Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada.

Drawing on more than 18,600 engineering years of autonomous vehicle safety development, Halos for Robotics provides developers with a common safety architecture for building, validating, and deploying physical AI systems. Safety is built into every layer: NVIDIA IGX Thor and Holoscan Sensor Bridge provide industrial-grade AI compute and sensor connectivity, the Halos OS software stack handles safety functions and applications, and the NVIDIA Halos AI Systems Inspection Lab, the world’s first ANAB-accredited program for functional and AI safety for physical AI, helps partners prepare for third-party certification.

Agility is extending its leadership in humanoid safety by integrating NVIDIA IGX Thor and Halos Core into its proprietary safe human detection system for its humanoid robot Digit. The company will also participate in the Inspection Lab to ensure Digit’s safety-related software, AI components, and cybersecurity protections meet rigorous standards such as IEC 61508, ISO 13849, and ISO/IEC TR 5469 before final third-party certification. “For humanoids to deliver value at scale, safety has to be built into the robot and validated across the entire system,” said Peggy Johnson, CEO of Agility. “This collaboration unlocks true human-robot teamwork.”

The Halos for Robotics ecosystem brings together more than 40 companies across software, embedded systems, sensors and silicon, industrial applications, and certification bodies—including Acontis, QNX, Advantech, Infineon, NXP, STMicroelectronics, Texas Instruments, FORT Robotics, and KION Group. Certification bodies including TÜV Rheinland, TÜV SÜD, UL Solutions, exida, SGS, and CertX all recognize the Inspection Lab as part of their certification process. NVIDIA Halos Core is available in early access for registered developers, and the open source Halos Outside-In Safety Blueprint is now available in early access on GitHub.

Read more here!

Eno is the debut multi-purpose robot from Genesis! 👀

Genesis AI has unveiled Eno, its debut robot. Based between Paris and San Carlos and founded by robotics engineer Zhou Xian, the company has secured investment from former Google CEO Eric Schmidt, among others. While decidedly non-human in form, Eno has been explicitly designed to fulfill almost all of the mobility and agility requirements of a typical adult, eschewing the head-torso-legs form factor in favor of a blend of agility and mobility.

Eno’s most obviously analogous appendages are its arms, with disarmingly human-like gloved hands. Everything else has the soft, clinical sheen of a robotic arm or welding machine. Describing it as “radically minimalist,” the Genesis AI team started with a premise of “calm intelligence,” positing that future robots, whether at home or elsewhere, need the ability to slip silently from view when not in use. The robot has a folding mechanism with articulated panels that fold up and down to extend its height while minimizing size when not in use. The hands were developed in house and closely mimic the abilities of a human hand.

Described as a “general-purpose robot,” Eno’s full-stack AI was developed in house by Genesis AI and is said to be sophisticated enough to make it a true physical agent—one that can decipher, decode, and learn tasks just like agentic AI on a high-end smartphone. To build trust right out of the box, one option is an optional screen that shows a representation of Eno’s processes in real-time. The company stresses the importance of Eno being able to work alongside humans, rather than as a remote machine in a dark warehouse.

Read more here!

Hyundai to buy SoftBank’s remaining stake in Boston Dynamics for $325 million! 💸

Hyundai Motor Group plans to buy SoftBank Group’s remaining 9.65% stake in Boston Dynamics for $325 million to make the U.S. robotics firm a wholly owned subsidiary, according to a report by South Korea’s Maeil Business Newspaper.

The report said SoftBank had told Hyundai it would like to exercise its rights to sell its remaining stake in Boston Dynamics under a put option agreed when it sold Boston Dynamics to Hyundai. Hyundai Motor Group, including Group Executive Chair Euisun Chung and affiliates Hyundai Motor, Kia, Hyundai Mobis, and Hyundai Glovis, already owns just over 90% of the robot maker.

The move would consolidate full ownership of Boston Dynamics under Hyundai, deepening the automaker’s commitment to robotics as it continues to position the company at the center of its long-term automation and physical AI strategy.

Read more here about it!

Free Robotics Resources 👨🏼‍🏫

Learn how to use MATLAB and Simulink with platforms like VEX Robotics, LEGO and other educational robotics kits.

MathWorks experts share knowledge on elementary concepts in robotics for beginners.

Access it here for free!

AEON begins production tasks at BMW plant! 🏭

BMW Group and Hexagon Robotics have reached a new milestone in their collaboration as AEON, Hexagon Robotics’ humanoid robot, begins performing production tasks within BMW Group Plant Leipzig, while simultaneously being trained on future manufacturing applications. “Humanoid robotics is entering a new phase,” said Arnaud Robert, President of Hexagon Robotics. “The question is no longer whether these systems can operate in manufacturing environments, but how quickly they can become productive, versatile, and scalable.”

The next phase of the collaboration will focus on increasing autonomy, robustness, and operational performance. Battery assembly activities are also expected to move from the Innovation Garage into the factory environment as part of future pilot activities. Robert added that the partnership with BMW and the progressive rollout at the Leipzig plant is enabling Hexagon to demonstrate the value of humanoids in the real world.

The collaboration reflects a broader trend toward the adoption of humanoid robotics in industrial environments. According to Hexagon’s global Robot Generation study, people overwhelmingly favor the use of robots in factories, warehouses, and hazardous environments, while preferring human involvement in caregiving and education. The findings suggest strong public support for robots performing physically demanding industrial tasks when deployed with appropriate governance and safety measures.

Looking ahead, BMW Group and Hexagon Robotics are continuing to build the path from pilot systems to production-ready humanoid workforce solutions. AEON will be deployed in production by the end of 2026, marking an important step toward the next generation of manufacturing automation.

Read more here about it!

Neura Robotics raises up to $1.4B Series C at $7B valuation
→ Tether, Qualcomm, Amazon, NVIDIA, Bosch, Schaeffler, EIB participate with conditional milestones
→ Why European robotics reaching mega-round scale signals competitive credibility alongside US/China leaders

Theker raises €80M, Europe’s largest robotics Series A
→ Barcelona-based startup builds reconfigurable robots: swappable hands, arms, form factors vs fixed humanoid design
→ Raising 2x target, growing from dozens to 120 employees, expanding showrooms across Europe, US, Asia

Robotiq ships TSF-85 tactile sensor digital twin on NVIDIA Isaac Sim
→ 97% SSIM accuracy on hyperelastic model, 90% on elastic, trained on 53,400 real-world maps
→ What 2.3M cycle durability with zero signal drift reveals: models maintain consistency through demanding factory use

Standard Bots raises $200M Series C at $1B valuation
→ AI-native robots taught by demonstration, targeting 10% US industrial deployments by next year, 100% US manufacturing by 2027
→ Why 150% job posting increase + 15% employment growth in robot-adopting plants challenges “job loss” narrative

Neura Robotics raises up to $1.4 billion in Series C! 💰

German company Neura Robotics has raised a Series C round worth up to $1.4 billion to develop its humanoid robotics technology, with U.S. tech giants Tether, Qualcomm, Amazon, and Nvidia all participating. The round also featured European industrial companies Bosch and Schaeffler, alongside the European Investment Bank. The company hit a valuation of around $7 billion, according to a source familiar with the matter.

“The future of AI will not only live on screens,” said David Reger, founder and CEO of Neura Robotics. “It will move, interact, learn and work beside us in the real world.” The full funding is contingent on Neura hitting certain milestones based on company performance, the source added.

Investors have piled into robotics startups in recent times as attention turns to deploying AI in physical systems that can interact in real-world environments. Robotics companies have raised $55.8 billion so far in 2026, according to Dealroom, a record figure nearly double the previous record raised last year. The majority has been raised by companies in the U.S. and China, but new European robotics companies are emerging too, including SoftBank-backed Agile Robots and U.K.-based Humanoid.

“Many believed globally relevant AI infrastructure companies could only emerge from Silicon Valley,” Reger said. “We believe the next generation of AI leaders can emerge anywhere in the world where there is enough vision, engineering talent and execution speed. With this financing, Neura is firmly among the global leaders in the robotics race, alongside the best in the US and China.”

Read more here!

Theker just raised $85M to build the factory robots! 🦾

Theker, a Barcelona-based AI robotics startup, has raised $85 million in what it’s calling “Europe’s largest ever robotics Series A.” The company aims to go beyond robots trained for a single task. “If you always have to put the same cookie in the same box, that works perfectly, but most processes aren’t like that,” co-founder Carla Gómez Cano told TechCrunch. Theker is designed for that messier reality.

Unlike humanoid robots designed around a fixed form, think Boston Dynamics, Theker’s machines are built to be reconfigured. Their hands, arms, and overall form can be swapped out or resized depending on the task, whether that’s sorting packages, packing clothing, or handling bottles and cans in a warehouse. Inditex, Zara’s parent company, signed on as an early backer, signaling where Theker’s ambitions start—the broader goal is to move beyond retail into heavier industrial settings like manufacturing, where the complexity and scale of manual tasks is even greater.

The Series A came less than a year after a record seed round, led by American VC firm CRV and backed by a mix of traditional and strategic investors, including Samsung and Aglaé Ventures, the investment vehicle tied to LVMH chairman and CEO Bernard Arnault. Gómez Cano said Samsung is not a client yet but the two are in advanced discussions—a potential trifecta of customer, supplier, and investor that would give the startup both revenue and credibility in manufacturing at scale.

To demonstrate it can deliver, Theker has a showroom in central Barcelona and plans to open others as it expands across Europe, the U.S., and Asia. The company has already received 15,000 job applications and could grow from dozens to up to 120 people by the end of the year. Raising twice its target has reinforced the startup’s conviction in keeping its HQ in Barcelona, a growing robotics hub, and in Europe’s tech ecosystem more broadly.

Read more here!

Robotiq releases tactile sensor digital twin on NVIDIA Isaac Sim! 🫆

Robotiq has released the digital twin of its TSF-85 tactile sensor in NVIDIA Isaac Sim, the first industrial-grade tactile sensor digital twin shipping on a commercial collaborative gripper. Tactile sensing promises to accelerate robotics, but its adoption has been limited by the lack of data from industry-ready hardware and accurate simulations. Model builders can now train contact-rich manipulation policies in simulation, then run them on the same physical sensor designed to operate reliably on the factory floor.

Most tactile sensors rely on a deformable contact interface, the very property that gives them sensitivity is also what makes them difficult to simulate. The TSF-85 digital twin is built around that constraint. It generates synthetic tactile maps through a custom Isaac Sim UI panel, visualizes them in real time, runs data generation at the simulation refresh rate, and exports to HDF5 for downstream training pipelines.

The simulation method is documented in two peer-reviewed publications. The first, published in Frontiers in Robotics and AI (2025), draws on a dataset of 53,400 real-world tactile maps to train, validate, and test each simulation pipeline, achieving up to 97% Structural Similarity Index Measure (SSIM) for the hyperelastic model and 90% SSIM for the elastic model on 12 unseen objects. A companion paper at ICCRT 2025 releases an open dataset of 46,200 real and synthetic tactile samples.

Simulation accuracy matters, but it only helps if the real sensor remains stable over time. The TSF-85 has been tested through 2.3 million cycles at maximum gripper force, with no significant variation in tactile signal output. This means models trained on its data can continue to rely on consistent signals for edges, shapes, textures, and geometry even after demanding real-world use.

Robotiq has been the go-to components provider for both academic research labs and industrial production for more than a decade, exactly the bridge Physical AI requires. The digital twin supports NVIDIA Isaac Sim 5.1 and is available now on GitHub.

Read more here about it!

Free Robotics Resources 👨🏼‍🏫

This course gives a solid intro to modern robot learning, mixing classical methods with recent large-scale models.

Students begin with imitation learning, reinforcement learning, and policy optimization, then move on to topics like Vision-Language-Action (VLA) models and foundation models for robotics.

Access it here for free!

Standard Bots raises $200 million to bring jobs back to the US! 🇺🇸

Standard Bots has announced a $200 million Series C at a $1 billion valuation, with the company claiming to be the largest manufacturer of AI-native industrial robots in the United States. The pitch is counterintuitive: robots will make American manufacturers more competitive, and that will unlock more jobs. “We’ve gone from 20 million manufacturing workers in 1979 to only 13 million today,” their CEO said.

Standard Bots makes AI-native robot arms and industrial humanoid robots that can be taught by demonstrations rather than code—you show the robot the job, it watches, learns, and starts doing the job itself. The pitch is that this collapses the cost and complexity that have kept advanced automation out of reach for most American manufacturers, especially small and mid-sized shops. Customers range from Sunoco, Lockheed Martin, Amazon, NASA, and the U.S. Army to hundreds of smaller manufacturers. “AI-native robots are the essential power tool of the 21st century,” said CEO Evan Beard. “The quickest way to get to full autonomy is through deployments, collecting real-world data, and iterating as fast as possible.”

The company argues investment in robotics is key to closing the gap with China. According to the International Federation of Robotics, China installed 295,000 industrial robots in 2024, 54% of the global total, more than the rest of the world combined, and its operational robot stock now exceeds 2 million units, roughly 4.5 times that of second-place Japan. The Americas accounted for just 9% of new deployments. Standard Bots says it will deliver 10% of new U.S. industrial robot deployments by next year, designs almost all its own parts including its own actuators, and by 2027 plans to manufacture everything—from metal in to robots out—right here in America.

The $200 million Series C was led by robotics fund RoboStrategy alongside existing backer General Catalyst, a steep markup from the company’s previous $63 million raise in 2024. “Standard Bots stands out because they’ve solved one of the hardest problems in industrial automation: making robots that are not only powerful, but actually usable on the factory floor without specialized programming,” said Andrew Kang, CEO of RoboStrategy. The big question is whether adding more robots really does mean adding more jobs, Standard Bots cites a 2025 University of Minnesota study finding roughly 150% increase in job postings and 15% increase in employment in plants that adopt robots compared to non-adopters. Whether that brings the jobs back, or just the work, is the question the next few years will answer.

Read more here about it!

OpenAI revives robotics division after 5-year pause
→ Hiring full-stack engineers targeting infrastructure robots for data centers, power grids, factories
→ Why near-term focus signals clarity: not consumer/warehouse, but skilled worker support building AI boom’s physical infrastructure

Genesis AI releases Genesis World 1.0 simulation framework
→ 89% hardware correlation, 45% smaller reality gap, 200+ hour real-world task runs in <30 minutes simulation
→ What separation of training (real) from evaluation (simulated) reveals: avoids simulator overfitting, enables systematic variable sweeps

Robotiq launches IQ for automatic robot integration
→ 5-minute Fit Check replaces weeks of manual quotes and site visits for palletizing decisions
→ Why deployment-ready workcells with validated designs in actual factory environments change SMB economics

Helsing launches Area 9, unveils RX-1 robotics platform
→ European defense tech incubator building full robotic stack in-house: hardware + software layer for autonomous research
→ What partnerships with ETH Zurich and INRIA Paris signal: developing unpredictable terrain navigation for debris-laden environments

OpenAI just made robotics its next frontier! 🤖

OpenAI is actively hiring robotics engineers as it revives its robotics ambitions, with an initial focus on robots that can help build data centers, power grids, and other critical infrastructure. On May 31, Sam Altman posted on X that OpenAI Robotics is hiring, seeking “exceptional full-stack hardware, ops, systems, and ML engineers to help us program and manufacture robots that are useful for society.” Greg Brockman, OpenAI’s co-founder and president, followed up to confirm the division is “making rapid progress towards building AI that can help people in the physical world.”

The move marks OpenAI’s strongest robotics push since shutting down its robotics team in 2021. The near-term focus is specific: robots to support skilled workers building infrastructure, not household helpers or warehouse sorters, but the workforce that will need to wire, pour, and assemble the data centers, power grids, and factories that the AI boom demands.

This is not OpenAI’s first attempt at robotics. The company ran a robotic hand project called Dactyl from 2017, most notably solving a Rubik’s Cube one-handed in 2019, before disbanding the entire robotics team in 2021. Since then, it has maintained a foothold through minority investments, the OpenAI Startup Fund led a $23.5 million round for 1X Technologies in 2023 and joined the $675 million round for Figure AI alongside Microsoft, NVIDIA, and Jeff Bezos. However, Figure subsequently ended its collaboration with OpenAI in February 2025 to build its AI models entirely in-house.

OpenAI’s robotics hardware lead Caitlin Kalinowski, who joined from Meta in November 2024, resigned on March 7, 2026, citing concerns about the company’s Pentagon deal and the lack of guardrails around domestic surveillance and lethal autonomy.

Read more here about it!

Genesis AI releases Genesis World 1.0! 🎮

Genesis AI has released Genesis World 1.0, a comprehensive simulation framework that treats simulation not just as a data generator, but as the fundamental infrastructure layer for robotics development. The platform includes Nyx (a photorealistic rendering engine built for robotics), Genesis World (a unified physics simulator supporting rigid and deformable bodies), and Quadrants (a GPU-accelerated compiler). The key insight: simulation should accelerate the robot development cycle itself, turning model evaluation from a real-world bottleneck into a compute problem.

Genesis’s approach prioritizes evaluation first. Rather than training on simulated data, they train on real-world data but evaluate in simulation—keeping training and evaluation workflows separate. This avoids overfitting to simulator artifacts. Their simulation correlates at 89% with on-hardware rollouts, and their reality gap is 45% smaller than competing simulators. In tests, a single evaluation that would take 200+ hours of continuous robot operation in the real world runs in under 30 minutes in simulation.

The platform enables systematic evaluation across hundreds of tasks with hundreds of episodes each, revealing exactly where policies fail. They can sweep variables—object shapes, lighting angles, camera trajectories—across multiple axes simultaneously, something nearly impossible on physical hardware. This lets researchers validate design choices before committing to large-scale pretraining, saving significant compute.

Going forward, Genesis is focused on three directions: scaling post-training through millions of simulated training iterations, building hybrid simulators that merge classical physics with learned models, and enabling self-evolving physical AI where agents improve end-to-end across both simulated and real-world loops. The north star: keeping robotics development from being bottlenecked by real-world wall-clock time, moving progress at the speed of compute instead.

Read more here!

Robotiq launches IQ to automate robot integration! 🔥

Robotiq has launched IQ, an AI platform that automates the entire robot integration process, attacking one of industrial robotics’ most stubborn problems. Integration has always been manual, unpredictable, and expensive. Site visits, weeks waiting for quotes, custom engineering every single time. IQ aims to change that entirely.

The platform starts with a 5-minute Fit Check, a structured assessment that tells you whether palletizing is a fit for your floor, what deployment looks like, and whether the financial return makes sense, including in 1-shift operations. No site visit required. No engineering hours upfront. No waiting weeks for a quote that may not go anywhere. Just an honest read on your opportunity before you commit to anything.

Once fit is confirmed, IQ guides the entire project, built on lessons from over 1,000 Robotiq deployments. It captures the reality of your operation, connects your team with the right people at the right step, and generates a validated workcell design simulated in your actual factory environment. Cycle time, reach, payload, all confirmed before you commit. The result: a deployment-ready palletizing workcell with predictable performance from day one.

Robotiq calls this “Automatic Integration”, the idea that deploying a robot workcell should be as predictable and repeatable as ordering one. For the thousands of small and mid-sized manufacturers who have been priced out of robotics by integration complexity and unpredictability, this changes the equation entirely. IQ was unveiled at RUC 2026 to selected partners and opens to everyone June 4th.

Read more here about it!

Robotics maps - London 🇬🇧

The robotics startups ecosystem is growing rapidly in London. Here’s a curated map of most of London-based companies working in the physical AI space. If I missed a company you know of, let me know! :)

Helsing officially launches Area 9, pushes into robotics! 🛡️

German defense tech giant Helsing has officially unveiled Area 9, a new research division designed as a pie-in-the-sky incubator for European defense tech, think Moonshot Factory or Skunk Works, but for Europe. As part of the launch, the company also announced RX-1, a robotics research platform developed by Area 9. “Until now, robotics research in Europe meant depending on platforms built elsewhere,” said Helsing’s Chief Scientist Antoine Bordes, who is leading the project. “We built RX-1 so that doesn’t have to be true anymore. Designed here, owned here, iterated here.”

Area 9 was first announced last year, with the innovation team initially working on AI agents for the battlefield, including Centaur AI. The Centaur agent was publicly unveiled in June 2025 when Helsing and Saab announced it had flown the latter’s JAS 39 Gripen E fighter. Centaur will also serve as the AI brain for the company’s CCA analog, the CA-1 Europa, unveiled last fall. Now the company is moving into the physical realm with robotics.

RX-1, in its current form factor, looks like a robot dog. Helsing plans to build all parts of the robotic stack—the robot itself plus the software layer that researchers can develop and test different autonomy capabilities on. The platform is more research-intensive and less currently-fieldable, but Helsing says it will make RX-1 available to research institutions and robotics laboratories across Europe. The company has already teamed up with the robotics group led by Marco Hutter at ETH Zurich and INRIA Paris to further develop RX-1.

The company’s goal is to develop an autonomous robotic platform that can navigate real-world, unpredictable terrain—like debris-laden environments—built entirely in Europe. “Area 9 exists to find the problems that look impossible and break them open,” Bordes said. “That’s why we made RX-1.”

Read more here about it!

Humanoid’s 1,000+ robot deal with Schaeffler signals 100K total shipments
→ Seven-digit actuator purchase (1M+) at 50% coverage implies 100K robots across all clients over 5 years
→ Why dual role matters: Schaeffler as both aggressive adopter (thousands of units) and preferred supplier

Boston Dynamics reveals Atlas fridge-lifting via reinforcement learning
→ Atlas trained millions of simulation hours with varied weight, friction, grip, positioning conditions
→ What proprioception demonstrates beyond cameras: sensing weight, balance, resistance for real-time unstable load adaptation

Locus Robotics acquires Nexera for AI-driven grasping
→ NeuraGrasp’s patented soft membrane + AI + vision adapts dynamically to shape, texture, material, porosity, weight variations
→ Why mobile manipulation at enterprise scale is the frontier: efficiently grasping millions of SKU types with speed/precision

Rivian spinoff Mind Robotics raises $400M, total hits $1B
→ $900M raised in 2 months at $3B+ valuation as Scaringe targets factory automation with “robotics with human-like skills”
→ Why “Project Synapse” launched: existing startups not fully equipped for industrial work

Humanoid’s 1,000+ robot deal with Schaeffler! 🏭

Two-year-old British startup Humanoid has signed a deal to integrate “a four-digit number of humanoid robots into live manufacturing operations” at German industrial giant Schaeffler’s factories starting later this year. The bigger news, however, might be that Humanoid has committed to buying a “seven-digit number of actuators” from Schaeffler over the term of the agreement, suggesting Humanoid intends to ship a staggering 100,000 humanoid robots across all of its clients over the next five years.

Four months ago, Schaeffler and Humanoid announced they’d put “several hundred” humanoid robots into factories over the next five to six years. On Wednesday, the companies upgraded the number significantly, signing an agreement to integrate at least 1,000 – and potentially many more, humanoid robots into Schaeffler’s manufacturing operations by 2032. This makes it one of the largest disclosed humanoid robot rollouts to date and arguably the most aggressive escalation between a humanoid OEM and an industrial customer.

The supply agreement indicates Schaeffler will become Humanoid’s preferred supplier, covering more than 50% of the company’s demand for joint actuators for its wheeled-based platforms through 2031. This partnership is expected to translate into the supply of a “seven-digit number of actuators” – at least one million. The wheeled HMND 01 Alpha offers close to 30 active degrees of freedom, and the “Schaeffler-relevant” actuator count per robot is probably roughly 18-22 joint actuators. At 50% coverage, one million actuators means 100,000 robots over the next five years.

Schaeffler is simultaneously one of the most aggressive industrial adopters of humanoids and a preferred component supplier for the platforms it’s buying. The company is committed to buying thousands of humanoid robots from Humanoid, at least 1,000 from Switzerland’s Hexagon Robotics, plus a bigger partnership with Germany’s Neura Robotics for 4,000-6,000 units. Schaeffler CEO Klaus Rosenfeld said the company is now engaged with roughly 45 humanoid robotics players worldwide and expects its humanoid robotics division to secure an order book in the hundreds of millions of euros by 2030.

Read more here about it!

Boston Dynamics reveals how Atlas robot lifts a.. fridge! 🧊

Boston Dynamics has revealed how its Atlas humanoid robot learned to lift and carry heavy industrial objects using reinforcement learning and large-scale simulation training. In a newly released technical blog, the company showed Atlas rotating its torso 180 degrees, squatting to pick up a mini-fridge, and carrying it across a lab floor while adjusting to shifting weight inside the object. The behavior was developed within weeks of Atlas’ public debut earlier this year.

The latest demo marks a shift from choreographed robot movements toward adaptable industrial behaviors designed for factories, warehouses, and construction sites. Boston Dynamics said Atlas is being developed as a “general purpose tool for physical work.” Rather than relying mainly on cameras, the robot uses proprioception, internal body awareness, to sense weight, balance, grip, and resistance while moving objects, allowing Atlas to adapt in real time to unstable loads and changing conditions.

Boston Dynamics trained Atlas through reinforcement learning, where the robot repeatedly practiced the same lifting task in simulation under different conditions. The company varied factors such as object weight, floor friction, grip strength, and fridge positioning to force the robot to adapt. “Atlas practiced the moves for millions of hours in simulations in parallel on GPUs,” the company said. The process starts with a reference trajectory, then Atlas is rewarded for completing tasks correctly while external disturbances are introduced.

One major advantage of the new Atlas platform is its reduced “sim-to-real gap”—a longstanding robotics challenge where behaviors trained in simulation fail in the physical world. Boston Dynamics said Atlas’ simplified hardware architecture makes accurate simulation easier, using only two actuator types across its body with symmetrical arms and legs.

Read more here!

Locus Robotics acquires Nexera Robotics! 🪬

Locus Robotics, the leader in Flexibility-First Warehouse Automation, has announced the acquisition of Nexera Robotics, a Vancouver-based robotics company specializing in advanced robotic grasping. The integration of Nexera’s proprietary NeuraGrasp end-effector technology into the Locus Robotics physical AI platform significantly expands the company’s autonomous mobile manipulation capabilities and broadens what Locus Array can handle across end-to-end fulfillment workflows.

NeuraGrasp combines AI-driven grasping intelligence, onboard sensory inputs, computer vision, and a patented soft membrane structure to adapt dynamically to the physical characteristics of each item. This enables a single gripper to conform to variations in shape, surface texture, material, porosity, and weight, creating reliable grasps across high-variability inventory found in real warehouse operations.

Locus Robotics CEO Rick Faulk stated that the frontier of warehouse robotics today is AI-driven mobile manipulation at enterprise scale, and being able to efficiently grasp millions of SKU types with both speed and precision is where the next decade of value gets created. Nexera CEO Roy Belak said they built NeuraGrasp to solve the manipulation challenges that have held robotic picking back for years, and joining Locus gives them the platform, scale, and customer base to bring this breakthrough technology into high-velocity fulfillment environments.

Read more here about it!

Robotics maps - Europe 🇪🇺

Alpine Tech Corridor is the undisputed deeptech capital of continental Europe. The data from the European Deeptech Report 2026:

Access the report here:

Rivian spinoff Mind Robotics raises another $400M!

Rivian’s spinoff company Mind Robotics has raised another $400 million, just two months after raising $500 million, as it works to develop industrial robotics that can further automate factory operations. The funding round, first reported by the Wall Street Journal, was led by Kleiner Perkins, with participation from the venture arms of Volkswagen (partnered with Rivian on a software joint venture) and Salesforce. The new funding brings total raised to more than $1 billion and values Mind Robotics at more than $3 billion.

Rivian CEO RJ Scaringe, who is chairman of Mind Robotics, told TechCrunch in March that he created the company because he felt other startups were not fully equipped to automate industrial work. He started the project—initially known as “Project Synapse”—as an effort to build “robotics with human-like skills.” Mind Robotics had previously raised $115 million from Eclipse after it was created in 2025.

Scaringe also helped create and spin out a micromobility company called Also, which has raised more than $300 million to date. The rapid fundraising pace, $900 million in just over two months, demonstrates strong investor confidence in Mind Robotics’ vision to bring advanced industrial robotics to factory operations and manufacturing environments.

Read more here about it!