Skip to main content
← Back to glossary

Vision-Language-Action (VLA)

Architectures

A model architecture that takes visual and language input and produces actions directly — keyboard and mouse movements, robot commands — in real time. Trained end-to-end on trajectories (miHoYo's Gamer uses 10K+ hours of gameplay), it is the class behind GUI agents and robotics policies that act by watching the screen, with no APIs or privileged state.

This term in other languages