?0.5 keeps the ?0 recipe — a vision-language-model backbone with a flow-matching action expert — but is built around generalisation to unseen environments. Where ?0 and most contemporary VLAs are evaluated in settings that closely match training, ?0.5 is trained on a deliberately heterogeneous mixture of knowledge sources: multi-robot demonstrations, web-scale vision-language data, verbal-feedback data and high-level semantic subtask labels. The model can break a long-horizon instruction into intermediate subtasks and act on them, and can improve from verbal corrections given by a human. Physical Intelligence reports that ?0.5 can control a mobile manipulator to tidy kitchens and bedrooms it has never seen before, which the company frames as a meaningful step toward broadly generalisable physical intelligence. Limitations: it remains a research model; the company notes substantial work remains on knowledge transfer, autonomous self-improvement and reliability; and unlike ?0, full open weights for ?0.5 were not released on the same terms at announcement, so teams should verify current availability. A later ?*0.6 model — a VLA that learns from its own experience — has since been described by the company.
?0.5 extends Physical Intelligence's ?0 model with a focus on open-world generalisation. It is designed to control a mobile manipulator in entirely new homes -- cleaning an unseen kitchen or bedroom -- rather than only environments close to its training data.
?0.5 keeps the ?0 recipe — a vision-language-model backbone with a flow-matching action expert — but is built around generalisation to unseen environments. Where ?0 and most contemporary VLAs are evaluated in settings that closely match training, ?0.5 is trained on a deliberately heterogeneous mixture of knowledge sources: multi-robot demonstrations, web-scale vision-language data, verbal-feedback data and high-level semantic subtask labels. The model can break a long-horizon instruction into intermediate subtasks and act on them, and can improve from verbal corrections given by a human. Physical Intelligence reports that ?0.5 can control a mobile manipulator to tidy kitchens and bedrooms it has never seen before, which the company frames as a meaningful step toward broadly generalisable physical intelligence. Limitations: it remains a research model; the company notes substantial work remains on knowledge transfer, autonomous self-improvement and reliability; and unlike ?0, full open weights for ?0.5 were not released on the same terms at announcement, so teams should verify current availability. A later ?*0.6 model — a VLA that learns from its own experience — has since been described by the company.