Physical Intelligence pi2 Foundation Model Powers 12 Commercial Robot SKUs Across Logistics and Retail

## The Foundation Model Approach to Robotics Physical Intelligence's core thesis is that robotics should adopt the same paradigm shift that transformed language AI: instead of training a bespoke model for each robot hardware variant and task, train a single generalist foundation model with broad physical understanding, then fine-tune it for specific hardware and tasks with a small amount of demonstration data. The pi2 model was trained on over 10,000 hours of robot demonstration data spanning manipulation, navigation, and object recognition across more than 50 hardware platforms. A hardware partner can access the model via API, provide 200-400 demonstrations of their target task, and receive a fine-tuned checkpoint ready for deployment in approximately 72 hours. ## Deployed Commercial SKUs The 12 commercially active robot product lines using pi2 include picking arms from three major warehouse robotics vendors, shelf-scanning units from two retail automation companies, a parcel-sorting system deployed across DHL facilities in Europe, and a last-metre indoor delivery robot used in hospital logistics. Physical Intelligence reports that pi2-powered robots achieve task success rates 18-27% higher than the same hardware running vendor-specific models, with the gains most pronounced on tasks involving irregular or loosely packed items. ## Next Steps Physical Intelligence is developing a pi2-Mobile variant optimised for legged and wheeled mobile platforms. A commercial preview is expected at the International Conference on Robotics and Automation in November 2026.