EMBODIED REASONING
Perception and action form a closed loop.
An embodied agent does not observe the world from a fixed distance. Its actions change what it can perceive next. Moving a camera changes the visual evidence. Reaching changes contact. Walking changes the map. Manipulation changes the scene.
DeepMind uses the term embodied reasoning for this bridge between multimodal intelligence and physical action. Recent robotics models combine vision, language and action while separate reasoning models operate over spatial and physical context.
EMBODIMENT MATTERS
The same task is different in different bodies.
Kinematics
Reachability and action spaces depend on links, joints and physical topology.
Sensing
What can be inferred depends on available sensors, placement and uncertainty.
Actuation
Torque, velocity, force and control modes shape what actions are possible.
Morphology
Hands, wheels, legs and end effectors change strategies and affordances.
History
Wear, calibration and adaptation can change the effective body over time.
Constraints
Safety and operational envelopes bound what intelligence may physically attempt.
GENERALIZATION
A policy should not silently assume every robot is the same robot.
One of the central challenges in embodied intelligence is transferring skills across different embodiments. A bimanual manipulator, a mobile base and a humanoid may share task semantics while exposing very different sensors, kinematics, control surfaces and safety constraints.
This is where a machine-readable embodiment graph becomes useful: the policy or planner can consume explicit body context instead of relying on hidden assumptions embedded in training data or middleware configuration.
OPENRGD POSITION
Embodiment should be inspectable, versioned and provenance-aware.
OpenRGD treats the robot body as structured context with explicit maturity and provenance. That is different from claiming that the graph itself is intelligent. OpenRGD describes the embodiment; models, planners and runtimes reason and act through it.