What is a robot foundation model?
The term generally refers to a broadly trained model intended to support many robotic tasks, environments or embodiments rather than one narrowly defined behavior. The exact architecture varies, but the common idea is reusable physical intelligence.
From perception to action
Robot foundation models may combine visual observations, language instructions, proprioception, demonstrations and other sensor information. Their outputs can range from high-level plans to action representations or control commands.
The data problem
Physical data is expensive because collection requires hardware, environments, supervision and safe operation. Training corpora can combine robot trajectories, demonstrations, simulation and video, but each source has different biases and levels of action grounding.
Embodiment is the missing variable
A general model still needs to interact with a specific machine. A target robot defines sensors, action dimensions, reachability, contact mechanics and safety constraints. Foundation-model research therefore intersects with embodiment-aware representations, action abstractions and adaptation.
What to evaluate
- Task diversity and generalization
- Performance across environments
- Transfer across robot embodiments
- Data efficiency during adaptation
- Robustness and recovery
- Latency and control-loop compatibility
Why this matters
The central promise is reuse: retain useful knowledge while changing the body that executes it, moving robotics toward shared intelligence with embodiment-specific interfaces.
Explore our foundation model directory and cross-embodiment learning guide.