A WORKING TAXONOMY
The stack is best understood as explicit responsibility layers.
There is no single universal Physical AI stack. NVIDIA organizes current robotics development around training, simulation and deployment infrastructure, while modern robotics models add world understanding, embodied reasoning and VLA policies. The taxonomy below is a vendor-neutral architectural view intended to clarify boundaries rather than define a new industry standard.
EIGHT LAYERS
From evidence to motion, every layer should know what it owns.
Data + evidence
Robot datasets, demonstrations, sensor logs, synthetic data and source-backed body descriptions.
Simulation + digital world
Physics simulation, scene composition, synthetic environments and sim-to-real workflows.
World models + reasoning
Predictive state, spatial reasoning, task decomposition and possible-future evaluation.
Policies + VLA
Models that map perception, language and context toward robot action.
Embodiment context
The actual robot body, capabilities, constraints, provenance and current configuration.
Runtime + safety
Intent handling, capability grounding, policy gates, deterministic limits and fallback.
Control + Body Adapter
Trajectories, middleware, buses, device units, actuator interfaces and body-specific execution.
Hardware + observation
Physical sensors, actuators and measured outcomes that close the loop.
END-TO-END FLOW
Intelligence becomes useful only when assumptions survive the trip to the body.
THE EMBODIMENT GAP
A capable model can still be underspecified for a specific robot.
Multi-embodiment models increase pressure on the interface between model output and a concrete machine. Different robots expose different sensors, ranges, payloads, control surfaces, calibration states and operational envelopes. A model that understands the task still needs trustworthy information about the body through which the task will be executed.
OPENRGD POSITION
OpenRGD targets the semantic layer between robot assets and physical authority.
OpenRGD does not replace datasets, simulators, world models, VLA policies, controllers or certified safety systems. It focuses on machine-readable embodiment context and explicit contracts around how semantic intent moves toward physical execution. Its canonical toolchain remains non-actuating; several downstream agent contracts remain candidate rather than normative.