physical-intelligence/openpi is an open-source repository of models and packages for robotics released by the Physical Intelligence team. It holds training and inference implementations of VLA models that connect visual and language inputs to robot actions, along with checkpoints.
The base checkpoints provided by the repository cover three models: π0, π0-FAST and π0.5. The README states that these checkpoints were pre-trained on 10k+ hours of robot data, and it organizes examples for using the models as they are as well as for fine-tuning them with your own data. In other words, both how to download and use the models and how to collect data and train again can be found in one place. The implementation language is Python and the license is Apache-2.0. The repository was created on October 21, 2024 and has 14318 GitHub stars.
That said, what can be confirmed and what cannot differ. Release information is not verified, so from the published material alone there is no way to know what exactly changed this time. Neither comparative performance figures for the three models nor actual robot operation results are confirmed. Out of 10 points, only the momentum score of 4.6 is filled in, while novelty, technology and quality are left blank. What is clearly established is limited to the fact that the checkpoints and example code are published under Apache-2.0, and the scale of the data used for pre-training. Right now, rather than ranking the models against one another, it is more realistic to read the code and documentation first and check whether the fine-tuning examples are in a form that is easy to follow.