π0.7 released, showing early compositional generalisation across tasks and robots

Physical Intelligence/π model family

Physical Intelligence released π0.7, trained with rich contextual prompts that include task instructions and demonstration metadata. The model transfers skills to unseen robots and tasks without retraining—for example, performing laundry folding on new robot platforms with no folding training data—and matches specialized models on tasks like espresso machine operation.

• Compositional generalization: π0.7 recombines learned skills to solve new problems and control unseen robot types.
• Rich prompting approach: includes task instructions, episode metadata on quality/speed, control mode labels, subgoal images.
• Transfer learning: model performs laundry folding on robot platforms with no folding training data provided.
• Performance: matches specialized reinforcement-learning-tuned models on complex tasks like espresso machine operation.