FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models
FastVLA introduces a training-free method to compress visual tokens in Vision-Language-Action (VLA) models for better efficiency. It adaptively determines how many visual tokens to keep at every step by measuring task relevance and the concentration of important information, leading to significant speedups on both simulated and real robots while maintaining task success.