π0 and π0-FAST: Vision-Language-Action Models for General Robot Control

Hugging Face Blog · 1 year ago

Physical Intelligence's π0 and π0-FAST Vision-Language-Action models have been integrated into Hugging Face's LeRobot repository, enabling robots to perform complex manipulation tasks across different embodiments. π0 was trained on data from seven robotic platforms and 68 unique tasks, generating smooth action trajectories at 50Hz using flow matching. π0-FAST, an autoregressive variant using Frequency-space Action Sequence Tokenization, achieves 5x faster training than diffusion-based models while improving generalization across robot morphologies.