Over the past eight months, a general-purpose robot foundation model named π0 has been developed to advance artificial physical intelligence. This model is unique in combining vision, language, and action, directly outputting motor commands for diverse tasks such as folding laundry and bussing tables. Unlike narrowly specialized industrial robots, π0 can be prompted or fine-tuned for various complex tasks, representing an essential step towards versatile robots capable of learning from experience. The model leverages Internet-scale pretraining and datasets from distinct robots to achieve unprecedented dexterity and physical capability.
Table of contents
The promise of generalist robot policiesA cross-embodiment training mixtureInheriting Internet-scale semantic understandingPost-training for dexterous manipulationEvaluating and comparing π 0 to prior modelsWhere do we go from here?Sort: