Physical Intelligence π0.5

2025-04-22 · Publisher : Physical Intelligence · Domain : Robotics

Physical Intelligence unveils π0.5, a VLA derived from π0 and co-trained on heterogeneous data, that cleans a kitchen or bedroom in entirely new homes.

### What changes On April 22, 2025, Physical Intelligence released π0.5, a new vision-language-action model built on π0. The goal is no longer only lab dexterity but "in the wild" generalization: a robot dropped into a home it has never seen should be able to work there. ### Specifications - **Co-training**: data from multiple robots, high-level semantic prediction, web data and other sources mixed in a single training run. - **Hybrid examples**: each example combines images, language commands, object detections, next-subtask prediction and low-level actions. - **Result**: according to the authors, an end-to-end learned system performs long-horizon, dexterous tasks, such as cleaning a kitchen or bedroom, in entirely new homes for the first time. - **Openness**: π0.5 weights were added to the openpi repository in September 2025. ### Why it matters π0.5 shows that data diversity matters as much as volume. Knowledge transfer across heterogeneous sources becomes the key to generalization, an argument later echoed by Figure, Generalist AI and Dyna Robotics. ### Key takeaway With π0.5, VLAs leave the lab: the target becomes the unknown home, not the demo table.

Tags : VLA, Généralisation, Robotique domestique, Physical Intelligence