VLA Models

VLA (Vision-Language-Action) – “VLM + action output”: see and reason plus act with hands.

Two-level architecture:

slow long-term planning – what to do: decompose instruction into subgoals;
fast VLA control – how to do: continuous control signals.

Two directions of action representation:

discrete tokens (RT-2, OpenVLA) – autoregression as text generation;
continuous trajectory (π₀) – flow matching, smoother for precise manipulations.

Foundation – imitation learning on cross-platform demonstrations; RL – an additional layer. Action decomposition into blocks: one model output – a sequence of future actions, execution at high frequency.