Physical Intelligence open-sources π0

Physical Intelligence, a robotics startup building general-purpose control models for robots, released the code and weights behind π0 (pi-zero), its vision-language-action model, as an open-source project called openpi. The release included the base pre-trained model along with checkpoints fine-tuned for two existing hardware platforms, ALOHA and DROID, plus inference and fine-tuning code so other teams could adapt it to their own robots.

π0 takes camera images and language instructions as input and outputs low-level robot commands, intended to generalise across different robot bodies rather than being trained separately for each one. Physical Intelligence said its own experiments found that between one and twenty hours of task-specific data was enough to fine-tune the base model for a new task, and the company positioned the release as experimental, cautioning that results would vary across different robots and setups. Hugging Face published a parallel PyTorch port of the model around the same time, widening the pool of developers who could run it without Physical Intelligence’s own training stack.

The release put a general-purpose robot foundation model into the open at a point when most comparable work — at Figure AI, at Google DeepMind’s robotics group, and inside Physical Intelligence’s own commercial roadmap — remained closed. It did not include the data or training pipeline behind the model, so outside groups could run and adapt π0 but not reproduce it from scratch. Physical Intelligence went on to raise substantially larger funding rounds later in 2025, including a $600 million Series B that valued the company at roughly $5.6 billion, on the strength of the same underlying model family.