Physical Intelligence

Physical Intelligence is a company founded in 2024 that develops foundation models and learning algorithms to power robots and physically actuated devices.

The company's flagship product, π0 (pi-zero), is a vision-language model trained on 3.3 billion parameters dedicated to robot control. The model enables robots to perform complex tasks through natural language commands, similar to how users interact with chatbots. The system allows robots to perform household tasks like folding laundry, clearing tables, assembling boxes, and loading coffee into grinders, aiming to revolutionize how robots are utilized across multiple industries.

To achieve dexterous robot manipulation, which involves highly precise, coordinated movements required for grasping, balancing, or adjusting objects, π0 must generate motor commands with fine control. This requires the model to output continuous, real-time actions at high frequencies, up to 50x per second, ensuring smooth, responsive interactions with objects. To train π0 for this level of control, the model adapts a vision-language model to produce continuous outputs instead of language, using a flow-matching technique that enables the generation of fine-tuned motor commands.