Precise Manipulation with Efficient Online RL

Extract an RL Token from VLA models to enable fast online RL and improve throughput on precise tasks with few hours of data.