Source: TechCrunch
Physical Intelligence's π0.7 model transfers knowledge across tasks without explicit training data for each one—a genuine but limited achievement. Robot companies have spent years trapped in task-specific systems requiring constant retraining, so any escape from that cycle matters. The gap between "early sign of generalization" (the company's framing) and production deployment is substantial. Generalization in controlled labs doesn't guarantee performance in messy real-world environments where robots encounter friction, material variation, and edge cases training data never captured. The competition isn't about one model's architecture. It hinges on whether Physical Intelligence can scale training data faster than competitors iterate on their own approaches, and whether any system can justify its deployment costs outside high-volume, standardized warehouses.