PHYSICAL AI

Intelligence becomes physical when actions have consequences.

AI systems that perceive, reason and act in the physical world through autonomous machines and robots.

CITATION RECORDCanonical source ↗
MATURITYAUTHORITY EXPLAINER
LAST VERIFIED22 Sep 2026
SOURCE VERSIONCurrent NVIDIA Physical AI materials + OpenRGD architecture
EVIDENCE BOUNDARYVendor-neutral concept explainer; OpenRGD does not claim ownership of the Physical AI category or replace models, simulators or controllers.

FROM MODEL TO MACHINE

The loop is no longer purely informational.

In a digital-only system, a wrong prediction may produce a bad answer. In a physical system, a wrong prediction can produce a wrong motion. Physical AI therefore joins perception, spatial understanding, reasoning, planning, control and safety into one embodied loop.

NVIDIA describes Physical AI as the ability of autonomous systems to perceive, understand, reason and perform actions in the physical world. Simulation and synthetic data become important because a robot cannot safely learn every edge case by trial and error on real hardware.

SYSTEM STACK

Physical AI is not one model.

01

Perception

Cameras, depth, LiDAR, force, proprioception and other sensor streams.

02

World representation

Spatial state, objects, topology, uncertainty and modelled dynamics.

03

Reasoning and policy

Task decomposition, VLA policies, planners and learned action strategies.

04

Simulation

Physics-based environments for training, validation and synthetic evidence.

05

Control

Trajectories, force, locomotion and low-level execution interfaces.

06

Safety

Operational envelopes, deterministic limits, fallback and execution gating.

WHY OPENRGD

A policy needs to know what body it is acting through.

A general AI model can reason about a task while still lacking a trustworthy, machine-readable representation of the specific robot body in front of it: joints, limits, sensors, capabilities, operational constraints, provenance and lifecycle state.

OpenRGD targets this grounding problem. It does not replace the model, simulator, controller or runtime. It provides a semantic description layer that can be validated before information flows toward physical execution.

01Perception / model
02ActionIntent
03OpenRGD grounding
04Safety boundary
05Body Adapter
06Hardware

BOUNDARY

Knowing the body is not permission to move it.

OpenRGD deliberately separates semantic context from actuation authority. The canonical toolchain is non-actuating, and a valid profile does not prove physical compatibility or operational safety. Physical execution remains an explicit downstream responsibility.

PRIMARY SOURCES