Physical Intelligence π0.7

2026-04-16 · Publisher : Physical Intelligence · Domain : Robotics

Physical Intelligence publishes its work on π0.7, a model that, per the company, gets robots to perform tasks they were never taught, coached step by step in plain language, with "early signs" of generalization.

### Background On April 16, 2026, Physical Intelligence published new research on π0.7. According to the company, as reported by TechCrunch, the model can direct robots through tasks they were never explicitly trained on, a capability that surprised its own researchers. ### Specifications - **Compositional generalization**: instead of memorizing each task, the model recombines learned skills to solve new problems. - **Plain-language coaching**: an operator can walk the robot step by step through an unfamiliar task; co-founder Sergey Levine says it then tends to work pretty well. - **Comparison**: lacking standardized robotics benchmarks, the team compared π0.7 with its own specialist models; the generalist matched them on making coffee, folding laundry and assembling boxes (results published by the company). - **Caution**: the paper speaks of "early signs" of generalization and "initial demonstrations"; these are research results, not a deployed product. ### Why it matters Learned robotics has so far worked by rote: collect data for a task, train a specialist, repeat. π0.7 hints at an inflection point similar to large language models, where capabilities compound beyond what the data would predict. ### Key takeaway After π0, π0.5 and π*0.6, Physical Intelligence targets generalization to unseen tasks; external validation is still pending.

Tags : VLA, Généralisation, Physical Intelligence, Guidage par le langage