Physical Intelligence Publishes Demonstration Results for Robot Foundation Model π0.6

In a blog post, the company announced demonstration results with its partners Weave and Ultra. Adding Weave data to pretraining reduced grasping errors by 42% and interventions by 50%, while the model achieved a 96.4% autonomy rate at Ultra’s site.

The significance is that the company showed the same foundation model applied to two tasks with substantial variation in shapes and procedures: folding laundry at home and packing customer orders in a warehouse. According to Physical Intelligence, adding Weave task data to pretraining reduced grasping errors and remote interventions, while at Ultra the model achieved 96.4% autonomous operation over one shift using existing workstations. Users can compare concretely which tasks benefit from additional data or operating conditions.