What Physical Intelligence does
Physical Intelligence (π) is a robotics AI company developing general-purpose vision-language-action (VLA) models intended to control robots across a wide range of tasks and embodiments. The company’s stated goal is to build learning algorithms and model architectures that can control “any robot to do any task,” using a foundation-model approach adapted to the physical world.
More
Rather than treating robot capabilities as a per-task software engineering problem, Physical Intelligence focuses on training “robot brains” that take visual observations plus language instructions and produce actionable robot behaviors. Its research and published model lineup (e.g., π0, π0.5, π*0.6, π0.7) emphasizes generalization, compositional instruction following, and improvements to autonomy and throughput when models are deployed on real robotic systems. Go-to-market positioning: Physical Intelligence describes a “Physical Intelligence Layer” approach—partners integrate their robot hardware with PI models so the models can be used as general-purpose control components, with partners’ systems handling deployment-specific interfacing. In a partner write-up (Ultra), PI describes measurable improvements in real warehouse performance as models progress across generations, including higher autonomy and reduced human interventions.