π0.5 (pi-0.5)

π0.5, published by Physical Intelligence on 22 April 2025, was the company's answer to the hardest question in robot learning: what happens when the robot is dropped into a house nobody trained it in. It builds on the π0 vision-language-action architecture but is co-trained on a deliberately heterogeneous mixture — around 400 hours of mobile-manipulation demonstrations, data from non-mobile robots working in many different homes, the cross-embodiment corpus from π0, multimodal web data (question answering, captioning, object detection) and 'verbal instruction' episodes in which a human coaches the robot through a task step by step. The model reasons in two stages. First it predicts a high-level semantic step in discrete tokens — effectively telling itself what to do next — then a 300-million-parameter flow-matching action expert turns that intent into continuous motor commands. Physical Intelligence's central finding was about breadth rather than depth: once training covered roughly a hundred distinct environments, performance in a completely new home approached that of a policy trained in that home. Demonstrations showed multi-minute tasks such as clearing dishes into a sink and making a bed in unseen kitchens and bedrooms. Unlike the closed π*0.6 and π0.7 that followed, π0.5 did reach the public: Physical Intelligence added a pi05_base checkpoint to the openpi repository in September 2025, under Apache 2.0 and alongside PyTorch support, so the model can be run and fine-tuned outside the company.

▸Videos

No videos yet.

▸Reviews