Representative End User Clients
Physical Intelligence refers to the capability of an artificial system (robot, machine, or autonomous vehicle) to perceive, reason, and act effectively within the complex, unstructured, and physics-governed reality of the physical world. Unlike "disembodied" AI that processes text or images in the abstract realm of the cloud, Physical Intelligence possesses a deep, intrinsic understanding of physical laws—gravity, friction, inertia, material properties—allowing it to manipulate objects, navigate dynamic environments, and adapt to physical variability without rigid pre-programming.
Strategic Analysis and Context
It is vital to distinguish between Physical Intelligence the Concept and Physical Intelligence the Company. The latter is a high-profile startup (backed by OpenAI and others) aiming to build "foundation models for robots." However, the ARC definition focuses on the broader concept as a new frontier of industrial capability.
The core loop of Physical Intelligence is Perceive -> Reason -> Act.
Perceive: Using multi-modal sensors (vision, lidar, tactile) to build a real-time understanding of the environment.
Reason: Using foundation models to plan a path or a manipulation strategy that accounts for the physical properties of the object (e.g., "this box is cardboard and might crush," "this part is oily and might slip").
Act: Executing the motion with precise force and feedback control.
Digital Twins are the "engine" of Physical Intelligence. Because the real world is slow and dangerous for training (a robot learning to walk might break itself), Physical Intelligence is cultivated in physics-accurate simulations (like NVIDIA Omniverse) where millions of hours of "experience" can be compressed into days. This allows the AI to learn generalizable skills—how to pick any bin, not just this bin.