eRLT: Efficient VLA Reinforcement Learning via Action-Relevant Token Routing
eRLT improves sample efficiency in online reinforcement learning for robotic manipulation by efficiently adapting frozen Vision-Language-Action (VLA) models to specific tasks. It constructs a task-specific state representation by routing action-relevant information across tokens and layers of the VLA, allowing the model to learn how to best aggregate features for better action refinement and value estimation from limited online interactions.