What does VLA mean?
Vision-language-action describes a model or system that connects visual observations and language instructions to physical actions. A VLA approach seeks a shared pathway from multimodal context to behavior.
A simple pipeline
A typical system can receive images or video, a natural-language goal and robot state. It then predicts an action representation, trajectory or intermediate command that a lower-level controller executes.
Why language helps
Language provides a compact interface for goals such as locating an object, changing a task or specifying a constraint. But language alone cannot solve embodiment: the model still needs grounding in the target robot's sensors, capabilities and action interface.
The embodiment challenge
A VLA model trained on one action representation may not transfer directly to another. Different robots expose different joints, grippers, cameras and control frequencies. Practical VLA systems therefore need action grounding, adaptation or embodiment-specific interfaces.
Where VLA systems fit
- Instruction following for manipulation
- Long-horizon task execution
- Human-robot interaction
- Generalist robot policies
- Transfer across tasks and potentially embodiments
What to watch
For physical AI, the important question is whether behavior remains reliable when the environment, task, sensor configuration or robot body changes.
Explore the VLA field guide and model directory.