Physical Intelligence

Skild AI's $1.4 billion Series C, led by SoftBank, and Figure AI's $1 billion Series C, led by Parkway Venture Capital, are the two largest humanoid and embodied-AI rounds in MEKAWIRE's verified records since January 2025, and together make up 41% of the $5.86 billion disclosed across 22 rounds. Humanoid makers took 70% of disclosed money in 2025, but developers of robot foundation models have taken 71% in 2026 so far, and NVIDIA joined seven rounds without leading one.

A vision-language-action (VLA) model is one neural network that reads camera images and a plain-language instruction and outputs a robot's next motor commands, built by training a vision-language model on robot demonstrations. Google DeepMind, Physical Intelligence, NVIDIA, Figure AI, Skild AI and AgiBot lead, and the commercial race now turns on the cost of action data rather than model design.

Runway announced Praxis-1 on 30 September 2026, an open-weight world action model that converts its large-scale video pretraining into control for real robots. Early tests run with Noble Machines, Standard Bots and Ultra on bimanual, six-axis arm and mobile-base hardware, and public weights are promised in the coming months.

General Intuition, which trains AI models on action-labelled gameplay clips from the Medal platform, said on 29 September 2026 that it had raised $220 million at a $6.2 billion valuation from investors including Khosla Ventures and General Catalyst. The figure is 2.7 times the $2.3 billion it disclosed with its Series A 96 days earlier, and the company has opened a waitlist for partners in robotics, simulation and games.

Skild AI said on September 10 that it crossed $100 million in annual recurring revenue ten months after its first commercial deployment, with more than 60 paying customers. The figure follows the August 18 launch of S1, a robot foundation model that the company says can pick up a new task from a single video demonstration without retraining.

Generalist AI said on August 19, 2026 that GEN-1.5 can learn a new manipulation task from one demonstration of 3 to 12 seconds placed in its context, averaging 59% success over 10 tasks with no fine-tuning and 83% after 10 gradient steps on about 5 minutes of data. The company says the ability emerged from more than eight months of pretraining, but calls the tasks simple and the skills brittle, and has named no customers.

Google DeepMind on July 30 released Gemini Robotics 2, a family of three models covering whole-body humanoid control, high-level embodied reasoning and on-device inference. The reasoning model ER 2 is available to developers through the Gemini API and Google AI Studio, while the action models remain limited to early-access partners and trusted testers.

Walden Robotics, spun out of Toyota Research Institute in January 2026, emerged from stealth on July 15, 2026 with a $300 million seed round at a $1.1 billion valuation co-led by Toyota and Deviation Capital. The Cambridge, Massachusetts company says its robots, driven by large behavior models, have been doing production work at a Toyota plant in North America since February, and on July 28 it added Samsung SDS as a manufacturing partner.

Hugging Face released version 0.6.0 of its open-source LeRobot robotics library on July 7, adding three world-model policies, five vision-language-action models including NVIDIA's GR00T N1.7, two reward models and six simulation benchmarks. NVIDIA said a day earlier that it is bringing GR00T, Isaac Teleop, Isaac Lab-Arena, Jetson Thor support and, soon, Cosmos 3 into LeRobot.

Korea's trade ministry said on May 29 that the AI robot division of its M.AX manufacturing alliance now has about 280 members and is running industrial humanoid pilots with HD Hyundai Heavy Industries, SK Energy and BGF. It also highlighted Korean robot foundation models, a 13 km/h humanoid lower body from KAIST and plans for a 500 billion won fund with a mandatory humanoid share.

Sereact, the Stuttgart developer of the Cortex robot model, raised US$110 million in a Series B led by Headline on 27 April 2026, citing more than 1 billion production picks and one remote intervention in about 53,000. The money funds Cortex 2.0, which adds a world model to a vision-language-action model, and a first US office in Boston; Zalando later joined, lifting the round to US$116 million.

Sony AI's autonomous table tennis system Ace won three of five matches against elite players and lost both matches against two Japanese T.League professionals in April 2025 trials, according to a paper published in Nature on April 22, 2026. The robot combines event-based vision sensors from Sony Semiconductor Solutions, an eight-axis arm and reinforcement learning policies trained in simulation, and reaches a perception latency of 10.2 milliseconds.

An 82-author team from Toyota Research Institute, Cornell University and MIT reported in Science Robotics on April 15, 2026 that multitask large behavior models, built on diffusion policy, outperform single-task baselines in blind, randomized trials and need far less task-specific data. The preprint version describes about 1,700 hours of robot demonstrations across more than 500 tasks and 1,800 controlled real-world trials.

Skild AI said on April 15, 2026 that it had acquired the robotics division of Zebra Technologies, the business Zebra bought as Fetch Robotics in 2021. Zebra's quarterly SEC filing shows the sale closed on March 27 for $20 million in cash, escrow and a minority stake in Skild, against the $301 million Zebra recorded when it bought Fetch.

Google DeepMind released Gemini Robotics-ER 1.6 on April 14, 2026, an embodied reasoning model built on Gemini 3.0 Flash and offered as a preview through the Gemini API and Google AI Studio. DeepMind said the model reads industrial instruments correctly 86% of the time, or 93% with an agentic zoom and code execution step, against 23% for ER 1.5. Boston Dynamics is using the capability for Spot facility inspections.

Generalist AI released GEN-1 on April 2, 2026, an embodied foundation model pretrained on more than 500,000 hours of human physical interaction data captured with wearable devices and no robot data. The company reports 99% average success on tasks where its previous model reached 64%, about one hour of robot data per task and task times roughly three times faster than the prior state of the art; the results are the company's own and have not been independently replicated.

Munich-based Agile Robots said on 24 March 2026 it had formed a strategic research partnership with Google DeepMind to integrate Gemini Robotics foundation models with its hardware, starting with manufacturing use cases. The German company, which says it has installed more than 20,000 robotic solutions, becomes the latest hardware maker in Google's orbit after Boston Dynamics and Apptronik. No financial terms were disclosed.

NVIDIA used its GTC keynote on March 16, 2026 to preview Isaac GR00T N2, a robot foundation model built on its DreamZero world action model research that it says succeeds at new tasks in new environments more than twice as often as leading VLA models. It also put the 3-billion-parameter GR00T N1.7 into commercial early access and announced Cosmos 3 and Isaac Lab 3.0, with N2 due by the end of 2026.

Mind Robotics, the industrial AI robotics company founded and led by Rivian chief executive RJ Scaringe, said on March 11, 2026 that Accel and Andreessen Horowitz co-led a $500 million Series A, about four months after a $115 million seed. Rivian's filings show it deconsolidated the start-up in March, valued its remaining 37.6% at about $569 million using the Series A share price and recorded a $506 million gain, which by MEKAWIRE arithmetic implies an equity value near $1.51 billion.

Palo Alto start-up Rhoda AI emerged on 10 March 2026 with a $450 million Series A from Khosla Ventures, Mayfield, Temasek and six other institutions, plus John Doerr. Its Direct Video Action model predicts video of a task and converts it into robot actions, and the company says new tasks can need as little as ten hours of teleoperation data. No lead investor, valuation or customer was disclosed.

Physical Intelligence on February 24, 2026 laid out its plan to sell general-purpose robot models as a ready-made ‘physical intelligence layer’ for other robot makers, the way developers call language models through an API. It named home robot maker Weave and warehouse robot maker Ultra as partners, reporting a full packing shift at 96.4% autonomy and a 50% cut in laundry-folding interventions, but gave no pricing or access terms.

Figure AI on Jan. 27, 2026 released Helix 02, a robot foundation model that runs a humanoid from camera pixels and fingertip touch down to joint commands at 1 kHz, replacing 109,504 lines of hand-written C++ with a 10-million-parameter learned controller. The company showed a Figure 03 unloading and reloading a dishwasher across a full kitchen in a four-minute, 61-action run it says had no resets and no teleoperation.

Microsoft Research on January 21, 2026 introduced Rho-alpha, its first robotics model, derived from the Phi vision-language family and trained partly on synthetic data from NVIDIA Isaac Sim on Azure. The model adds tactile sensing and learning from operator corrections to the usual vision-language-action recipe, and is offered through a research early access program ahead of a later Microsoft Foundry release.

Skild AI said on 14 January 2026 it had closed a $1.4 billion Series C led by SoftBank Group, valuing the company at more than $14 billion. NVIDIA's venture arm, Macquarie Capital and Bezos Expeditions joined, alongside new strategic backers LG, Schneider Electric, CommonSpirit and Salesforce Ventures. The start-up, founded in 2023, said its Skild Brain software went from no revenue to about $30 million within a few months of 2025.

1X Technologies said on January 12, 2026 that its NEO humanoid now runs on the 1X World Model, a 14-billion-parameter video generator that imagines a task before an inverse dynamics model turns the clip into joint commands. The system was mid-trained on 900 hours of first-person human video and fine-tuned on only 70 hours of NEO data. 1X's own trials show it working on simple household tasks and failing on fine ones such as pouring cereal, months before the first $20,000 NEOs were due in US homes.

Robot foundation model developer Physical Intelligence has raised $600 million at a $5.6 billion valuation in a Series B led by Alphabet's growth fund CapitalG, according to its investors. The round followed the release of π*0.6, a model the company says more than doubles throughput on hard tasks after reinforcement learning on real robots.

Generalist AI introduced GEN-0 on November 4, 2025, a family of embodied foundation models that the company says improves predictably with more physical interaction data. It reported a capacity threshold at 7 billion parameters, models scaled past 10 billion, and a pretraining corpus of more than 270,000 hours growing by over 10,000 hours a week. The findings were published as a company blog post, not a peer-reviewed paper.

Google DeepMind on 25 September 2025 introduced Gemini Robotics 1.5, a vision-language-action model that reasons in natural language before moving and transfers skills across robot bodies, alongside Gemini Robotics-ER 1.5, an embodied reasoning model opened to developers as a preview in the Gemini API. In its technical report, the pairing of the two cut total subtask failures on long-horizon tasks to 22% from 44.5% when a general Gemini 2.5 Flash model did the planning.

Dyna Robotics closed a US$120 million Series A on 15 September 2025, led by Robostrategy, CRV and First Round Capital with NVIDIA, Amazon, Salesforce, Samsung and LG venture arms joining, six months after a US$23.5 million seed. The Redwood City company did not disclose a valuation; it says its DYNA-1 model already runs paying deployments in hotels, restaurants, laundromats and gyms for up to 16 hours a day.

NVIDIA made its Jetson AGX Thor developer kit and Jetson T5000 production modules generally available on 25 August 2025, pricing the kit from $3,499 and claiming 7.5 times the AI compute of Jetson Orin with 128 GB of memory. Figure, Agility Robotics, Boston Dynamics and Amazon Robotics are among eight named early adopters, while 1X, OpenAI and Physical Intelligence are evaluating the computer.

Boston Dynamics and TRI showed Atlas completing a long workshop task under a single Large Behavior Model that commands walking and manipulation together. The 450M-parameter Diffusion Transformer runs at 30 Hz on camera, body-state and language inputs and learns new behaviours from VR teleoperation data without new code. No success rates or data volumes were disclosed.

FieldAI, an Irvine, California developer of robot autonomy software, said on August 20, 2025 that it had raised $405 million across two consecutive rounds from investors including Bezos Expeditions, BHP Ventures, Intel Capital, Khosla Ventures, NVIDIA's NVentures and Temasek. The company did not disclose a valuation and says its Field Foundation Models already run on quadrupeds, humanoids, wheeled robots and passenger-scale vehicles at industrial sites in Japan, Europe and the US.

Skild AI on July 29, 2025 named and described its robotics foundation model, the Skild Brain, a two-layer system pre-trained on simulation and internet video that the company says runs across quadrupeds, humanoids, table-top arms and mobile manipulators. The launch carried no customers or prices, but set out the body-agnostic strategy that later underpinned a $14 billion-plus valuation and more than $100 million in annual recurring revenue.

Genesis AI, a physical AI lab based in Paris and Palo Alto, came out of stealth on July 1, 2025 with $105 million co-led by Eclipse and Khosla Ventures, with Bpifrance, HSG, Eric Schmidt and Xavier Niel also investing. The company, which later called the round seed funding, plans a universal robotics foundation model trained on a data engine that mixes in-house physics simulation, generative modelling and real robot data, and did not disclose a valuation.

Google DeepMind on June 24, 2025 released Gemini Robotics On-Device, a vision-language-action model built to run locally on a robot with no network link. It is the first VLA the lab is letting outside developers fine-tune, through a new Gemini Robotics SDK offered to trusted testers, and DeepMind says it adapted the model from its ALOHA training rig to a bi-arm Franka FR3 and Apptronik's Apollo humanoid.

Physical Intelligence on 22 April 2025 published π0.5, a vision-language-action model that drove mobile manipulators through 10- to 15-minute cleaning jobs in three homes absent from its training set. The San Francisco start-up says co-training on data from other robots, web tasks and spoken coaching, rather than more home-robot data alone, is what lets the model generalise, and its own ablations show which ingredients matter most.

NVIDIA on 18 March 2025 released Isaac GR00T N1, a dual-system vision-language-action model for humanoid robots, as downloadable weights on Hugging Face and GitHub, with 1X, Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics named as early-access partners. It also announced Newton, an open-source physics engine under development with Google DeepMind and Disney Research, a synthetic manipulation-data blueprint and a 15-terabyte open physical AI dataset.

Google DeepMind on 12 March 2025 introduced Gemini Robotics, a vision-language-action model built on Gemini 2.0 that issues motor commands, and Gemini Robotics-ER, a spatial reasoning model for robot makers. The lab said the action model more than doubles performance on a generalization benchmark against other leading models, named Apptronik as its humanoid partner and opened Gemini Robotics-ER to four trusted testers including Boston Dynamics and Agility Robotics.

Figure AI on 20 February 2025 introduced Helix, a vision-language-action model that pairs a 7-billion-parameter vision-language model running at 7 to 9 Hz with an 80-million-parameter control policy running at 200 Hz across 35 degrees of freedom. Figure says the model was trained on about 500 hours of teleoperation data and runs entirely on two low-power embedded GPUs inside the robot.