What is a vision-language-action (VLA) model?

OpenVLA is a concrete VLA example. Its project describes a model for robot manipulation and publishes evaluation and adaptation information. Use the project and paper to inspect the supported robot setup, observation inputs and action representation. A VLA label alone does not specify control frequency, safety mechanisms or deployment readiness.

Record the model version, robot platform, camera configuration and action space. Ask whether a result comes from the released checkpoint or from a checkpoint fine-tuned on the target task. Note how success was measured, the number of attempts and the role of human resets. Keep simulation results separate from physical trials. A task score on one setup does not answer whether a different robot can use the same policy.

These terms describe different properties. VLA describes the connection between modalities and actions. Foundation model describes a model intended for adaptation across downstream uses. A model may fit both descriptions, but neither term establishes general physical competence. Compare a release through its documented tasks and transfer tests, rather than through the breadth of its name. See the robot foundation model guide for the broader context.