Physical Intelligence works on foundation models for robots. Most robots are programmed for one job, but the company aims to train a general model that can control many kinds of robots to do many tasks, such as folding laundry, clearing tables, and packing items, by learning from large amounts of robot data.
It has released research and open models, including the pi-zero family of vision-language-action models, along with code that lets other researchers try them. These models take camera images and instructions in natural language and output actions for the robot. The company collaborates with robot makers and collects data across different hardware.
Physical Intelligence is a research company, not a product for consumers, so its work reaches users through research releases and partnerships. Robot foundation models are early, and reliability in unstructured settings is still a challenge. Peers include Google DeepMind’s robotics work, Skild AI, Figure, Tesla Optimus, and NVIDIA’s GR00T. Physical Intelligence suits robotics researchers and developers following progress in general-purpose robots.