Short answer: a robot foundation model is a large neural network trained on many robots and many tasks so that it can be adapted to new tasks with far less task-specific engineering. The dominant form is the VLA — vision-language-action — model: camera images and a language instruction go in, robot actions come out. Instead of engineers hand-building a pipeline of perception, planning and control for every new job, the model learns the mapping from demonstration data. VLAs are the engine layer of embodied AI, and the most contested layer of the current robotics stack.
The named systems you will hear about
Three reference points anchor most discussions. NVIDIA used its GTC 2025 keynote to launch Isaac GR00T N1, an open foundation model aimed at humanoid robots, alongside its Isaac simulation ecosystem. Google DeepMind announced the Gemini Robotics family in 2025, extending its Gemini multimodal models toward physical action. And Physical Intelligence published its π0 model in 2024 as a general-purpose robot policy, showing one model performing varied manipulation tasks across different robot bodies. Companies such as Skild AI are pursuing the same prize from the startup side — a general "robot brain" licensed across many hardware platforms. These are public product and research announcements; treat each as evidence of direction, not of deployment scale.
How a VLA differs from the classic stack
| Classic engineered stack | VLA / foundation model | |
|---|---|---|
| Structure | Separate modules: detection, pose estimation, planning, control — each engineered and tuned | One learned model maps images + instruction to actions; conventional control still runs underneath for safety and smoothness |
| New task | An engineering project: new code, new tuning | Demonstrations and fine-tuning; sometimes just a new instruction |
| Strength | Predictable, inspectable, precise within its envelope | Handles variation and novel objects within its training distribution |
| Weakness | Brittle outside the situations it was built for | Opaque failure modes; reliability varies by task and object; hard to certify behaviour in advance |
| Scarce resource | Skilled integration engineers | Action-labelled robot data, and compute |
In practice, shipping systems are hybrids: the VLA handles the variable, judgement-heavy part — which object, what grasp, what order — while conventional perception and control handle precision and safety-critical motion. "Replaces programming" is the headline; "changes what gets programmed" is the reality.
The data problem underneath everything
Language models trained on the internet's text. No equivalent corpus of robot action exists, so model companies manufacture their data: teleoperation farms where humans drive robots through thousands of demonstrations, simulation at scale with randomised conditions, and logs harvested from deployed fleets. The economics of that pipeline — cost per useful hour of robot experience — quietly decide which models improve fastest. It is also why model companies partner with hardware makers and logistics operators: access to robots doing real work is access to training data.
The generalisation boundary, stated plainly
Current VLAs generalise impressively within reach of their data: new instances of familiar object categories, modest changes of lighting and layout, instructions phrased in new words. They degrade on objects far outside training, on tasks needing force judgement their sensors cannot feel, and over long task horizons where small errors compound. Cross-embodiment transfer — one model, many different robot bodies — is improving but still requires adaptation per platform. Any claim of "zero-shot, any task, any robot" should be met with a request for unedited, many-trial evidence.
What buyers and partners should ask
- Measured success, not demos: success rate over at least a hundred unedited attempts, on objects like yours.
- Adaptation cost: how many demonstrations and how much engineering to bring the model to your task and your robot?
- Uncertainty behaviour: when confidence is low, does it stop, retry differently, or escalate to a human?
- Update discipline: how are new model versions tested before they reach production machines?
- Data terms: does your operation's data train the vendor's general model, and on what terms?
The companies building these models — and the simulation, data and fleet-software firms around them — are listed under embodied AI and robot software in the directory, which is the fastest-changing of its six categories.
Frequently asked questions
What is a VLA model in robotics?
A vision-language-action model is a neural network trained to take camera images and a natural-language instruction as input and produce robot actions as output — joint movements or end-effector commands — directly, rather than through a chain of separately engineered modules. It is the robotics counterpart of the large language model: one general model, fine-tuned or prompted for many tasks.
What data are robot foundation models trained on?
Mostly robot experience that has to be deliberately created: teleoperated demonstrations (a human driving the robot through tasks), logs from deployed robots, and simulated episodes. Some projects add internet video to teach visual understanding, but action-labelled robot data is the scarce ingredient — which is why data collection operations are a core part of every serious model company.
Will a VLA model work on any robot?
Not automatically. Models are trained on particular robot bodies, camera setups and action spaces; moving to a different arm or gripper requires adaptation data and fine-tuning. Cross-embodiment training — pooling data from many robot types — is an active research direction and is improving transfer, but "one model, any robot, zero setup" is not the current reality.
How should a buyer evaluate a robot that claims a foundation model?
Ask for task success rates over many unedited attempts on objects like yours, the amount of site-specific fine-tuning required, what happens when the model is uncertain, and how updates are validated before deployment. A model's brand name matters less than its measured behaviour in your cell.