Simulation Assets

The robot, object, and scene models used in simulation, which need to look right and also carry physical properties.

Simulation assets are the raw material for building a simulated environment: robot models (usually described with URDF or MJCF for links and joints), manipulable objects, furniture, and entire room scenes. Unlike a game model, a simulation asset needs more than an appearance mesh and material — it also needs collision geometry (complex meshes are usually broken into convex pieces via convex decomposition), mass, friction coefficients, articulated joints (such as a drawer slide or a door hinge), and semantic labels. NVIDIA's OpenUSD-based SimReady assets specifically emphasize carrying physical properties, semantic annotations, and sensor properties. The quantity and diversity of assets determines how many kinds of training scenes can be produced, which directly affects a policy's generalization; common sources include Objaverse, PartNet-Mobility, and generative 3D models.

ExampleRoboTwin 2.0's object library, RoboTwin-OD, contains 731 objects: 534 custom-built meshes, 153 taken from Objaverse, and 44 articulated objects from PartNet-Mobility.

Also called

3D Assets, Scene Assets, Sim Assets

Related

SimReady Assets · Convex Decomposition · Articulated Object · Objaverse · Universal Scene Description (OpenUSD) · Unified Robot Description Format

Sources

NVIDIA SimReady 文档 (Chinese)
RoboTwin 官网（RoboTwin-OD） (Chinese)