What Is a VLA Model?

A VLA model (Vision-Language-Action model) is an end-to-end robot foundation model: it takes camera images and a natural-language instruction (e.g. "put the cup in the drawer") as input and directly outputs robot actions — joint angles, end-effector poses, or gripper commands. By compressing perception, language understanding, and motion generation into one network, VLA has become one of the most watched policy architectures in embodied AI.

Relationship to LLMs

VLA models are typically built on top of vision-language models (VLMs): they inherit the semantic and commonsense knowledge of large language models, then replace or extend the output head with action tokens or continuous actions. A useful mental model: an LLM predicts the next word; a VLA predicts the next action — it is the language-model family extended into the physical world.

Why Real-Robot Data Is Essential

The internet holds vast text and images, but almost no paired "image + instruction → joint action" data. VLA training therefore depends heavily on real-robot demonstrations, mostly collected via teleoperation. Simulation data can add scale, but must cross the Sim-to-Real gap. Data diversity — across tasks, scenes, and embodiments — often matters more for generalization than parameter count.

Representative Work

• RT-2 (Google DeepMind): co-trains a VLM with robot action tokens, demonstrating that web-scale knowledge transfers to manipulation;
• OpenVLA: an open-source VLA trained on large open robot datasets, designed for community fine-tuning;
• π0 (Physical Intelligence): uses flow matching to generate continuous actions, targeting a general cross-embodiment manipulation policy.

Deploying and fine-tuning VLA on real hardware requires a platform that executes high-rate action commands reliably — BXI's humanoid robot and dual-arm platform both expose ROS2 interfaces suited to VLA research.