SIMULATION FIRST
SDF can describe both a robot and the world around it.
SDFormat defines models, links, joints, sensors, collision, visual geometry, lights, physics and world-level configuration. A model can represent a complete robot or another physical object, while a world can contain models, scene and physics configuration.
RICH MODEL
The format reaches beyond a minimal kinematic tree.
Models
Complete robots or physical objects can be represented as named models.
Links and joints
Articulated multibody structure is part of the core model.
Sensors
A broad range of simulated sensor types can be declared.
Physics
Simulation physics and surface behavior are first-class concerns.
Worlds
Environments can contain multiple models, scene settings and physics.
Plugins
Simulation-specific extensions can be attached to world, model and sensor contexts.
BOUNDARY
Simulation completeness and semantic embodiment are not the same thing.
SDF is intentionally strong at describing what a simulator needs about models and environments. OpenRGD focuses on a different set of questions: canonical identity, provenance, operational semantics, declared capability, lifecycle and machine-readable governance context.
COMPLEMENTARITY
A simulation world can inform an embodied agent without becoming its semantic identity.
In a layered system, SDF may describe the simulated robot and environment while RGD represents the semantic embodiment profile consumed by AI or runtime components. The two can overlap on physical facts without sharing the same responsibility.
COMPARE
SDF and RGD meet around the robot, but from different directions.
Links, joints, geometry, inertials, collision and joint constraints in a ROS-centric XML model.
Not a general scene-composition system or semantic embodiment graph.
USDComposable scene descriptionLarge hierarchical scenegraphs, references, variants, time-sampled data and physics schemas.
Not a robot-specific semantic capability or governance model by default.
SDFSimulation-oriented model and world descriptionRobots, environments, models, sensors, physics, lights, plugins and simulation context.
Not designed as a general AI embodiment-governance or provenance graph.
MJCFMuJoCo-native dynamic-system modelNested bodies, joints, actuators, tendons, sensors, defaults and simulator-native dynamic modeling.
Optimized for MuJoCo rather than cross-stack semantic embodiment.
RGDSemantic embodiment graphIdentity, physical evidence, operational constraints, capability semantics, provenance and lifecycle context.
Does not replace scene composition, a physics engine, middleware or a physical runtime.