π0, and open weights for a robot foundation
Physical Intelligence released π0, a generalist policy trained on over 10,000 hours of data from 7 different robot configurations and 68 tasks — and three months later published its weights openly.
Why it matters
This is the point at which a private laboratory’s generalist policy covered long, dexterous jobs on arms of different shapes from one set of weights. Publishing those weights in February put a frontier robot foundation model in other people’s hands for the first time.
The architecture is a flow-matching action output on top of a pre-trained vision-language model. Pre-training ran on over 10,000 hours of dexterous manipulation data collected on 7 different robot configurations and 68 tasks, in addition to the open Open X-Embodiment dataset; single-arm, dual-arm and mobile manipulators are covered. Specific scores were published on the evaluation tasks: bussing easy 0.971, bussing hard 0.875, shirt folding 1.0, grocery bagging 0.786, toast out of toaster 0.750. On 4 February 2025 the company opened the openpi repository with code and weights for the base pre-trained model and several fine-tuned checkpoints; the same page states that 1 to 20 hours of data sufficed for fine-tuning in their experiments. Licence terms are not stated on that page.