What is a VLA model? Robot foundation models explained, from RT-2 to pi0.5, and why their makers raised $2.94bn in 2026
A vision-language-action (VLA) model takes camera images and a written instruction and outputs robot motor commands directly; Google DeepMind coined the term with RT-2 in July 2023. In 2026 the companies building these robot foundation models disclosed USD 2.94bn in six rounds in the MEKAWIRE database, about a quarter of all disclosed robotics funding and twice what humanoid hardware makers raised.
A vision-language-action model, or VLA, is a single neural network that looks at camera images, reads an instruction such as "put the cup in the sink" and outputs the robot's next motor commands. It is built by taking a vision-language model trained on internet images and text and teaching it to emit actions as well as words. In 2026 the companies that build these robot foundation models have disclosed USD 2.94 billion in six rounds in the MEKAWIRE database, a quarter of all disclosed robotics funding this year.
That money is a bet that the robot's brain, not its body, is where the value will sit. This explainer sets out how the models work, who makes them, what the money buys and what recent research says about how far they still have to go, using the developers' papers, company releases and our verified funding and paper rows only.
Google wrote the definition in 2023, and everyone kept it
Google DeepMind introduced the term in its RT-2 paper of July 28, 2023. Its recipe was simple: write robot actions as text tokens and train a vision-language model on robot trajectories and web question-answering at the same time, so that one model answers questions and drives an arm. The team ran 6,000 evaluation trials and reported that the robot could follow commands absent from its robot data, such as choosing a rock as an improvised hammer.
The open answer came in June 2024. OpenVLA, a 7-billion-parameter model built on Meta's Llama 2 with DINOv2 and SigLIP vision encoders, was trained on 970,000 real robot demonstrations. Its authors reported 16.5 points higher task success than the 55-billion-parameter RT-2-X across 29 tasks, and showed it could be fine-tuned on consumer GPUs. That combination made the VLA a research platform rather than a lab exhibit.
How the second generation changed the action half
Text tokens are a clumsy way to move a seven-joint arm smoothly. Physical Intelligence's pi0, published Oct 31, 2024, kept a pretrained vision-language model for understanding but generated continuous action chunks with flow matching, and showed laundry folding and box assembly across single-arm, dual-arm and mobile robots. Its successor pi0.5, published Apr 22, 2025, co-trained on robot data, web data and semantic subtask labels and, the company reported, cleaned kitchens and bedrooms in homes it had never seen.
NVIDIA's GR00T N1 (Mar 18, 2025) split the job explicitly into a vision-language System 2 that interprets the scene and a diffusion-transformer System 1 that produces motion in real time, trained on a mixture of real-robot data, human videos and synthetic data, and ran it on Fourier's GR-1 humanoid. Google DeepMind's Gemini Robotics (Mar 25, 2025) sits on Gemini 2.0 and an embodied-reasoning model, and learned new short tasks from as few as 100 demonstrations.
- RT-2
- Google DeepMind
- July 2023
- No
- Actions written as text tokens; emergent reasoning such as picking an improvised hammer; 6,000 evaluation trials
- OpenVLA
- University-led team
- June 2024
- Yes, 7B
- 970,000 demonstrations; +16.5 points over RT-2-X (55B) on 29 tasks
- pi0
- Physical Intelligence
- Oct 2024
- Not stated in paper
- Flow-matching actions on a VLM; laundry folding, table cleaning, box assembly
- GR00T N1
- NVIDIA
- Mar 2025
- Yes
- Dual system (VLM plus diffusion transformer); real-robot, human-video and synthetic data; Fourier GR-1
- Gemini Robotics
- Google DeepMind
- Mar 2025
- No
- Built on Gemini 2.0; new short tasks from as few as 100 demonstrations
- pi0.5
- Physical Intelligence
- Apr 2025
- Not stated in paper
- Co-training on heterogeneous data; long-horizon cleaning in unseen homes
Sources: the arXiv papers listed under sources. Capabilities are the developers' own reported results, not independent tests.
As of Oct 17, 2026
The models in that table share one architecture idea and differ in what they will give away. That difference shapes the funding.
Where the USD 2.94 billion went
Six robot-foundation-model and VLA companies disclosed rounds in 2026. Skild AI's USD 1.4 billion Series C at a USD 14 billion post-money valuation, announced Jan 14 and led by SoftBank Group, is 47.6% of the total on its own and a 9.3-fold valuation step-up from its USD 1.5 billion Series A of July 2024, 18 months earlier. Rhoda AI (USD 450 million), Generalist AI (USD 400 million), Walden Robotics (a USD 300 million seed at USD 1.1 billion) and Spirit AI (USD 280 million) follow, with Germany's Sereact at USD 110 million.
- Skild AI
- US
- Jan 14
- Series C
- 1,400.0
- 14.0
- SoftBank Group
- Spirit AI
- CN
- Feb 25
- Other
- 280.0
- n/d
- n/d
- Rhoda AI
- US
- Mar 10
- Series A
- 450.0
- n/d
- n/d
- Sereact
- DE
- Apr 27
- Series B
- 110.0
- n/d
- Headline
- Generalist AI
- US
- June 4
- Growth
- 400.0
- n/d
- Radical Ventures
- Walden Robotics
- US
- July 15
- Seed
- 300.0
- 1.1
- Deviation Capital
- Total
- No data
- No data
- 6 rounds
- 2,940.0
- No data
- No data
As of Oct 17, 2026. Verified rows with a disclosed amount from company or investor releases; companies whose primary category is robot foundation model, VLA or embodied AI. X Square Robot (two 2026 rounds) and ROBOTERA (two) are excluded because the amounts are not stored as verified figures. Physical Intelligence's rounds are known only from press reports and are excluded. Method: share = 2,940.0 / 11,468.8 (all 57 disclosed 2026 robotics rounds); average round = total / number of rounds.
As of Oct 17, 2026
The average disclosed round in this group is USD 490.0 million, against USD 201.2 million across all 57 disclosed robotics rounds this year and USD 181.6 million for humanoid-primary companies. Four of the six are American and they account for 86.7% of the money. The largest name in the field, Physical Intelligence, is missing from the table because it has not published its round sizes; MEKAWIRE carries them only as press reports.
MEKAWIRE analysis
The VLA is turning into the robot industry's shared operating layer, and investors are paying for control of the pretrained checkpoint and the data that improves it, not for arms or legs.
Our paper database shows the layer forming. Of 585 robotics papers ingested since Sept 1, 101 (17.3%) are VLA papers, and the three we briefed on Oct 17 all start from Physical Intelligence's pi0.5 and add something around it: MBZUAI's ForkCast forecasts whether to move the camera or replan the arm, Beihang and BAAI's LTM gives the policy a memory, and a Seoul National University, KAIST and Dnotitia team's verifier decides how many planned actions to trust. Research is moving from building VLAs to wrapping them, which is what happens when a base model becomes a standard.
The same papers carry the strongest counter-argument. A fine-tuned pi0.5 succeeded in 13.3% of disturbed RoboTwin episodes and 11.25% of real counting trials before the add-ons, and the gains came from task-specific data and engineering. If deployment success depends on per-site data and integration, value flows to whoever owns robot fleets and customer sites, which favours vertically integrated humanoid makers and incumbents over model start-ups.
The bull case is that one checkpoint fine-tuned with about a hundred demonstrations per task becomes cheaper than programming robots, and the model owners collect a toll on every deployment. The bear case is that open models such as OpenVLA and GR00T N1 keep closing the gap, VLA capability commoditises within two years, and USD 14 billion valuations meet software-like margins without software-like lock-in.
What would change our view is evidence from customer sites rather than benchmarks. Watch for any VLA developer publishing per-task success rates at a paying customer by Dec 31, 2026; for Physical Intelligence disclosing a round through its own channel; and for whether VLA papers rise above October's share in MEKAWIRE's November paper count, published in our December data story.
- Google DeepMind introduced the term vision-language-action model with RT-2 (arXiv, July 28, 2023), which writes robot actions as text tokens inside a vision-language model and was tested in 6,000 evaluation trials.
- OpenVLA (arXiv, June 13, 2024) is a 7-billion-parameter open model trained on 970,000 real robot demonstrations; its authors report it beat the 55-billion-parameter RT-2-X by 16.5 points of task success across 29 tasks.
- Physical Intelligence's pi0 (Oct 31, 2024) added a flow-matching action generator on top of a vision-language model; pi0.5 (Apr 22, 2025) co-trained on robot, web and semantic data and cleaned kitchens and bedrooms in homes it had never seen.
- NVIDIA's GR00T N1 (Mar 18, 2025) is an open VLA with a vision-language System 2 and a diffusion-transformer System 1, deployed on Fourier's GR-1 humanoid; Gemini Robotics (Mar 25, 2025) learned new short tasks from as few as 100 demonstrations.
- MEKAWIRE database, as of Oct 17, 2026: six robot-foundation-model and VLA companies disclosed USD 2,940.0m in 2026, 25.6% of USD 11,468.8m disclosed across 57 robotics rounds, and 2.0 times the USD 1,452.4m raised by humanoid-primary companies.
- Skild AI's USD 1.4bn Series C at a USD 14bn post-money valuation (Jan 14, 2026) is 47.6% of the 2026 foundation-model total and a 9.3-fold valuation step-up from its USD 1.5bn Series A of July 2024.
- 101 of the 585 robotics papers MEKAWIRE ingested since Sept 1, 2026 (17.3%) are VLA papers; all three briefed on Oct 17 use a frozen or fine-tuned pi0.5 as the base policy.
Why it matters
Every robot company now has to choose whether to build, license or fine-tune its manipulation brain. The six disclosed 2026 rounds put USD 2.94bn behind the idea that this layer can be separated from hardware and sold across many robot bodies, the way Android was sold across phones.
For buyers the practical question is switching cost. A plant that fine-tunes a licensed model on its own demonstrations owns the data but not the weights; an open checkpoint inverts that. Procurement teams should treat VLA licensing terms as a strategic decision now, before the first production deployments lock them in.
Rival analysis
Three camps compete. Platform owners with their own chips, clouds or consumer models (Google DeepMind with Gemini Robotics, NVIDIA with GR00T) can give models away to sell compute or keep them closed to sell services. Independent model labs (Physical Intelligence, Skild AI, Generalist AI, Rhoda AI) must charge for the model itself and have raised the largest private rounds. Vertically integrated robot makers build in-house models for their own hardware.
In China, Spirit AI, X Square Robot, ROBOTERA and Galaxea combine models with their own humanoids or arms and are financed partly by industrial strategics. Europe's disclosed entrant is Sereact, which sells VLA-driven picking into logistics. Korea has research depth (seven of 41 VLA papers with a recorded institution since Sept 1 come from Korean institutions) but no disclosed VLA start-up round in our database this year.
Valuation context
Skild AI's USD 14bn post-money in January is the only disclosed valuation among this year's foundation-model rounds besides Walden Robotics' USD 1.1bn seed. Skild's step-up from USD 1.5bn in July 2024 is 9.3 times in 18 months, faster than any humanoid hardware company in our database over a comparable period.
The average round of USD 490.0m is 2.4 times the robotics average of USD 201.2m and 2.7 times the humanoid average of USD 181.6m. These are disclosed figures only; undisclosed rounds of X Square Robot and ROBOTERA, and press-reported Physical Intelligence rounds, would raise the group's totals further.
Supply-chain implications
A VLA business consumes GPUs for training and inference, robot time for data collection and, increasingly, simulation. The papers we briefed show the inference side: pi0.5 policy inference at 66.71 ms on an NVIDIA H100 with memory added, and a verifier costing 0.45 to 1.24% of generation time. Training clusters are NVIDIA-dominated, which is why NVIDIA can afford to release GR00T openly.
On the robot side, the research base runs on a few platforms: Franka Panda arms, AgileX PiPER dual-arm kits, Unitree and Galbot humanoids. Data collection is labour-intensive teleoperation, which favours countries with low-cost operators and state-funded data centres, a point China's humanoid data-collection build-out exploits.
Signals to watch
First, customer-site evidence: any VLA developer publishing per-task success and intervention rates from a paying deployment by Dec 31, 2026. Second, Physical Intelligence disclosing a round or licensing terms through its own channel. Third, MEKAWIRE's November paper count: if VLA papers stay near or above October's share of about one in six, the research community is still converging on the architecture; a drop would suggest a shift to world models or other designs.
Analyst view
Thesis: the VLA is becoming a shared operating layer and investors are paying for control of the checkpoint and its data flywheel. Confidence: medium. The architecture has converged and the research ecosystem is standardising on a few base models, but no developer has yet shown audited production economics, and open models narrow the moat. We would lower confidence if open checkpoints match closed ones on customer tasks within 12 months, and raise it if a model developer signs multi-site licensing deals with disclosed terms.
Questions you should be asking
For model developers: what share of revenue comes from licences versus engineering services, and what success rate do customers see without per-site fine-tuning? How many demonstrations does a new task need in production, not in papers?
For investors in Skild AI and its peers: what does the USD 14bn valuation assume about the price per robot per year? For robot makers: at what fleet size does building an in-house model beat licensing one?