Vision-Language-Action Models (VLA)

The 2023 RT-2 paper used vision-language-action models as a category name and represented robot actions as tokens alongside language. OpenVLA followed in 2024 with an open 7-billion-parameter model trained on a reported 970,000 robot demonstrations, plus checkpoints and adaptation tools. The π0 work, first submitted as a preprint in October 2024 and later published at RSS 2025, instead used a flow-matching action architecture on top of a pretrained vision-language model. These examples establish a family, not a requirement to encode every action as text.

Sources: s1, s2, s3

VLA models seek to reuse broad visual and language representations while learning physical action, reducing the need to build an isolated policy for every instruction and object. RT-2 reported improved generalization to novel objects and commands in its evaluations. OpenVLA made checkpoints and training tools available for adaptation, and π0 addressed continuous, dexterous control with a different action-generation method. This line of work matters for general-purpose robotics, but results from selected tasks and laboratory setups do not establish dependable operation in uncontrolled environments.

Sources: s1, s2, s3, s4

In a Skills Intelligence illustration, a robot receives camera images and the instruction “put the red cup in the sink.” A VLA policy uses the scene and language to produce control outputs, such as end-effector movements. RT-2 represents actions through discrete tokens; π0 uses flow matching for continuous action generation. The implementation difference matters when adapting to a robot's action space. A system that only captions the scene is a vision-language model, while a generalist robot policy is not necessarily language-conditioned and therefore is not automatically a VLA.

Sources: s1, s2, s3, s4

Robot demonstrations are expensive and uneven, action spaces differ across hardware, and small perception errors can become physical failures. Reported success rates depend on task definitions, embodiments, training data, and laboratory conditions, which complicates comparison. Internet-derived semantics can also be poorly grounded in a specific robot's capabilities. VLA is a broad architecture category, not evidence that a system can safely generalize to arbitrary instructions, objects, people, or environments.

Sources: s1, s2, s3, s4