Vision-language-action models

What Vision-language-action models is

VLA models aim to give robots the generality of foundation models, so a single policy can follow instructions across tasks and objects it was not explicitly trained on.

How it works

A multimodal backbone is trained on robot demonstration data paired with instructions, outputting action tokens that a controller executes, with safety limits layered on top.

Why it matters

It is the most active route towards general-purpose robots, and progress here is closely watched as a physical-world capability signal.

Common uses

• →Manipulation research
• →Warehouse and lab automation prototypes