Physical Intelligence releases π0 and π0-FAST: Vision-Language-Action Models for LeRobot

AI Impact Summary

Physical Intelligence has released π0 and π0-FAST, Vision-Language-Action (VLA) models designed for general robot control, leveraging pre-training and flow matching for dexterous manipulation. These models, available in the LeRobot repository, represent a significant step towards versatile robot intelligence by integrating vision, language, and action, addressing the challenge of bridging the gap between AI and the physical world. The models utilize a state token to represent the robot's environment, action tokens to define motor commands, and a flow matching approach for generating smooth, real-time action trajectories.