What Physical Intelligence does
Physical Intelligence (π) is a San Francisco–based robotics AI company building general-purpose vision-language-action (VLA) foundation models intended to control a wide range of robots across many physical tasks. The company’s core thesis is that the path to broadly useful robot intelligence is a reusable “robot brain” (a generalist policy) rather than bespoke models per robot and per task.
On its website, Physical Intelligence describes itself as developing learning algorithms and foundation models to create a model that will “control any robot to do any task,” and it frames its research as a sequence of generalist policy improvements (e.g., π0, π0.5, π0.6, π0.7) plus supporting methods for faster training, real-time action execution, memory/learning over longer horizons, and steerability.
Business model and go-to-market signals are limited in the public materials available from the company’s site. The company does, however, position its work as enabling “a Cambrian explosion of robotics applications,” and its research posts emphasize capabilities that would be valuable to robot manufacturers, robotics integrators, and developers who need general-purpose robot control.
A notable differentiation is the company’s emphasis on *generalist* policies (cross-embodiment, instruction-following, and compositional generalization) and on practical execution constraints for real robots (e.g., efficient online RL and efficient action tokenization / chunking).