deeproboticslab/simready-assets

DEEP Robotics SimReady assets

robots at DEEP Robotics.

Organisation: DEEP Robotics
Contact: Percy Peng pengbo@deeprobotics.cn

Prepared USD robot assets for NVIDIA SimReady Central, version 2026.10.22_00. Each package includes its bill of materials, required profile-validation results, runtime benchmark results, and original source license notices.

Asset	Robot type	Source
M20S	Quadruped with wheels	Existing OEM USD
M20	Quadruped with wheels	High-resolution OEM URDF, updated mounting frames
X30	Quadruped	High-resolution OEM URDF, updated mounting frames
S10	Wheeled robot	Existing OEM USD
M20_Piper	Mobile manipulator	Existing OEM USD
DR02-std	Humanoid	High-resolution OEM URDF
DR02-pro	Humanoid with wrist joints and rubber hands	High-resolution OEM URDF
Lite3	Quadruped	New 20260821 OEM USD mechanics, local high-resolution visuals, three mounting frames

Lite3 is included following the newly supplied 20260821 source. Its USD mechanics match published Isaac Sim 6.2, while its three URDF mounting sites are new. The local high-resolution visuals are retained. The supplied USD/MJCF thigh centers of mass are preserved; the new URDF still records older values. Lite3, M20 and X30 mounting frames preserve source positions and rotations and do not implement camera or lidar simulation.

Lite3 corrects four invalid source knee position-drive targets from 0 degrees to their unchanged OEM lower limit of 30.022985 degrees. This avoids Newton's midpoint fallback (approximately 94.996402 degrees). Drive gains, hardware torque caps, limits, masses, collisions and source geometry are retained. Controlled ground-drop checks pass on both engines; the final packaged runtime results remain authoritative, including unresolved slope findings.

Simulation preparation

The staged models use relative package dependencies and UsdPreviewSurface materials. Repairs preserve robot geometry and mechanical structure except for removing exactly zero-area triangles and duplicate rendered surfaces. Rigid-body hierarchies are flattened where needed while retaining world poses and remapping joint targets. DR02 Pro retains its extra wrist DOFs, visible rubber hands, and mesh-derived convex-hull hand collisions.

X30's mirrored render transforms are baked into its visual geometry while preserving all world-space vertices and normals. Its visual instances are expanded for portable material overrides. Its undersized torso collision box is replaced with a convex envelope derived from existing CAD vertices; the original box is retained disabled. OEM torso mass, inertia, and zero center of mass are preserved, and the active collider count remains unchanged.

DR02 uses acceleration-drive stiffness 625 and damping 50. Its native torque variant includes a passive-viscosity simulation preset resisting 2% of rated torque at 1 rad/s. This is a simulation preset rather than measured hardware friction. Native torque limits, joint limits, masses, and inertias are retained.

S10 and M20S include selectable Source/Locomotion control and Source/Ground placement variants. Defaults use a simulation locomotion preset (leg position control and continuous-wheel velocity control) and place the lowest visual/collision bound 5 mm above ground. Source choices restore original control and frame placement. PhysX leg gains are 1000/100 for S10 and 2000/100 for M20S; the Newton/MuJoCo variant uses runtime-specific angular gain serialization. Geometry, relative mounting poses, mass, inertia, and hardware torque/joint limits are retained. Newton slope rebound remains a known runtime finding; read each package's results for exact verdicts.

Newton mesh collision representations and articulation self-collision settings are explicit. X30's CAD hull and Piper's existing mesh colliders declare the Newton mesh collision schema. S10/M20S retain their existing self-collision setting; Piper retains the prior Newton importer default. These declarations restore backend validation and runtime test coverage without changing geometry. Lite3 explicitly preserves its source-disabled self-collision setting.

The X30 state-accuracy benchmark commands three 0-degree waypoints outside its OEM knee minimum of about 20 degrees; the observed errors match that target mismatch. Original joint limits and official failed verdicts are retained.

Newton runtime tests use the intended 1000-contact MuJoCo allocation. Dense robot contacts can exceed the SDK default of 200; allocate enough contact capacity when configuring your simulation.

Required Robot-Body profile validation and runtime physics benchmarks are separate checks. Read the JSON results inside each package for exact findings; package availability does not imply every runtime test passed. NVIDIA review is pending. Source BSD notices remain in the packages alongside the CC-BY-4.0 submission license.

NVIDIA submission

Open pull request — awaiting review; not approved or merged.

Current package results

Required authoring validation and runtime verdicts are separate. Skipped and unrecorded checks are not passes.

Asset	Required Robot-Body 3.0.0	Runtime verdicts	Immutable package
M20S	Pass	1 error, 4 fail, 6 missing, 22 pass, 3 skipped	Package
M20	Pass	1 fail, 6 missing, 17 pass, 12 skipped	Package
X30	Pass	3 fail, 20 pass, 13 skipped	Package
S10	Pass	1 error, 4 fail, 8 missing, 21 pass, 2 skipped	Package
M20_Piper	Pass	2 fail, 15 missing, 19 pass, 2 skipped	Package
DR02-std	Pass	1 fail, 6 missing, 17 pass, 12 skipped	Package
DR02-pro	Pass	1 fail, 8 missing, 16 pass, 11 skipped	Package
Lite3	Pass	5 fail, 4 missing, 21 pass, 6 skipped	Package

Exact failure messages, engine names and warnings remain in each immutable package's .metadata/<asset>_benchmark.json. The linked PR preserves official runtime findings; package availability does not imply every benchmark passed.