Physical Intelligence Announces New RL Method “RLT” to Speed Up Precision Manipulation

Physical Intelligence announced a new method, “RLT,” that enables robots to learn precise manipulation from minutes to hours of real-world data. The company said RLT adds special output tokens to a VLA model and adapts actions through a small RL policy. Across four tasks, it said the method improved speed by up to 3x in the most precise stage.

Physical Intelligence said RLT uses the added output tokens as cues for a small reinforcement learning policy to correct actions, without fine-tuning the entire large VLA model. The method is intended to teach the final stages of tasks—such as aligning a screw with a small screw, tightening it with an electric screwdriver, bundling items or inserting a cable—where small errors in position, orientation or timing can lead to failure, using a short period of real-world data.