Physical Intelligence published π0.7, describing it as "a step-change in generalization" for robotic foundation models. Unlike prior vision-language-action models, π0.7 demonstrates compositional generalization: it can combine skills learned across different tasks and robot platforms to complete jobs it was never explicitly trained on, such as using an air fryer or folding laundry on an unseen robot arm. The model accepts language instructions, execution-quality metadata, and visual subgoal images as prompts, and its laundry-folding performance on a UR5e robot it had never seen matched zero-shot expert human teleoperators (TechCrunch). The result is significant because it suggests robots could be deployed in new environments and improved through plain-language coaching without retraining.