Three inputs, one output: motor commands
A vision-language-action model, usually called a VLA, takes in camera images of the current scene, a natural language instruction describing the task like 'put the apple in the bowl,' and often the robot's own proprioceptive state such as current joint angles. It outputs a sequence of low-level actions, typically target joint positions, velocities, or end-effector poses, that the robot executes over the next fraction of a second to few seconds before the loop repeats with a fresh observation. This perceive, decide, act cycle runs continuously, closing the loop between what the robot sees and what it does.
The 'language' part is what makes VLAs distinct from earlier learned robot policies that only mapped vision to action. By training on paired image-instruction-action data, the model learns to condition its behavior on the specific instruction, so the same visual scene produces different actions depending on whether you asked it to 'pick up the cup' or 'push the cup away.' This is directly analogous to how a language model conditions its next-word prediction on the prompt it was given, except here the 'next word' is a motor command.