Physical Intelligence Announces RL Token Method to Speed Up Precise Manipulation

On March 19, 2026, Physical Intelligence announced an online reinforcement learning method using RL tokens. According to the company, RLT adds special output tokens to a VLA model and provides an interface to a lightweight RL policy, allowing the model to adapt its behavior in a few hours without fine-tuning the entire model. Across four precision manipulation tasks, it accelerated the most precise stages by up to 3 times and was able to exceed the speed of human teleoperation.

For tasks such as tightening screws and inserting cables, where fine adjustments to position and orientation determine success, the key point is that existing vision-language-action models can be tuned quickly without readjusting the entire model. The company said the method made the most difficult stages of four precision manipulation tasks up to three times faster and surpassed human teleoperation, addressing the final refinements needed for delicate work.