General-purpose robot foundation model (Pi0)

Pi0 created by Physical Intelligence introduces a groundbreaking robotic model designed to integrate vision, language, and action (VLA) for general-purpose robotic tasks.

Model Architecture:

• A larger VLM backbone with weights initialized from PaliGemma that is pre-trained from large-scale Internet pre-training.

• A smaller action expert (Diffusion Transformer) addresses robot states and generates actions.

More Information:

• Full paper: https://www.pi.website/download/pi0.pdf

• Homepage: https://www.pi.website/blog/pi0

• Github link: Physical-Intelligence/openpi

Model Conversion