Fast-VLA: You Only Think Once for Robot Manipulation
Fast-VLA introduces Fast-VLA to improve vision-language-action (VLA) policies by separating task understanding from real-time control. It freezes a 4B Vision Language Model (VLM) once per episode to create fixed, multi-layer task memory. A lightweight online actor then uses this fixed memory and current sensory feedback to predict actions without repeatedly running the large VLM, significantly boosting efficiency and performance.