ncsu-digital-manufacturing/teeth_retainer_v2

Teeth Retainer v2

A robotics simulation scene and retainer scan mesh for exploring dental-retainer handling and manufacturing workflows in NVIDIA Isaac Sim.

This repository contains a 3D simulation asset. No trained policy, neural-network weights, training code, or evaluation results are included. The retainer geometry is bundled as a local USD asset generated from the included STL scan; the scene references it with a relative path.

Asset details

Field	Details
Repository	ncsu-digital-manufacturing/teeth_retainer_v2
Asset type	OpenUSD scene with an accompanying STL scan
Main scene	ur5_teeth_v2.usd (binary USDC format); equivalent text version: ur5_teeth_v2.usda
Scan	meshes/retainer_scan/retainer_scan.stl
Text retainer asset	meshes/retainer_scan/retainer_scan.usda (equivalent geometry and material in readable USDA format)
Local retainer asset	meshes/retainer_scan/retainer_scan.usdc (1,385,355 triangles; 702,614 unique vertices; scale matches supplied URDF)
Environment	Conveyor belt, Thorlabs table, stand, and Stereolabs ZED X sensor references
Coordinate system	Z-up; one scene unit equals one meter
Timeline	Start: 0; end: 1,000,000; 60 time codes per second
Physics settings	PhysX scene; 240 physics steps per second; continuous collision detection enabled; GPU dynamics disabled
Camera integration	Isaac Sim OmniGraph nodes for render-product creation and ROS 2 camera publishing
Software evidence	Included screenshot shows Isaac Sim Full 6.0.1; external asset URLs point to the Isaac 5.0 asset collection. Compatibility has not been tested.
License	Not specified in the inspected repository

The filename includes ur5, and some saved render products target /World/ur5e/.... However, the inspected root layer does not define or reference a UR5/UR5e robot. Do not assume that a robot or functioning manipulation controller is included.

Repository contents

.
├── .gitattributes
├── README.md
├── ur5_teeth_v2.png
├── ur5_teeth_v2.usd
├── ur5_teeth_v2.usda
└── meshes/
    └── retainer_scan/
        ├── retainer_scan.stl
        ├── retainer_scan.usdc
        └── retainer_scan.usda

Choosing a file

File	Use	Approximate size
ur5_teeth_v2.usd	Open the complete workcell scene in compact binary format.	55 KB
ur5_teeth_v2.usda	Inspect or edit the same workcell scene as readable text.	55 KB
meshes/retainer_scan/retainer_scan.usdc	Load the standalone retainer with its yellow material; used by both scene files.	17.9 MB
meshes/retainer_scan/retainer_scan.usda	Inspect or edit the standalone retainer geometry and material as readable text.	67.4 MB
meshes/retainer_scan/retainer_scan.stl	Original scan geometry used to generate the retainer USD assets.	69.3 MB

The USDA exports were checked for complete layer equivalence with their binary counterparts when generated. Both main scene files retain the relative reference ./meshes/retainer_scan/retainer_scan.usdc; opening the USDA scene still requires that USDC file. To use the text retainer asset in a scene, change the reference on /teeth_retainer to ./meshes/retainer_scan/retainer_scan.usda.

The binary and text files are separate copies. After editing one, re-export its counterpart to keep both versions consistent. The STL contains geometry only; the yellow material is stored in the retainer USD assets.

Getting started

1. Download the complete repository, including the USD scene, preview, and mesh directory. Preserve the directory structure. If cloning with Git, use Git LFS to retrieve any LFS-managed assets.

2. Open ur5_teeth_v2.usd or ur5_teeth_v2.usda through File → Open in Isaac Sim. For the retainer alone, open meshes/retainer_scan/retainer_scan.usdc or its USDA counterpart.

3. Allow the referenced NVIDIA assets to resolve, or replace their references with accessible local copies. The scene directly references these assets under https://omniverse-content-staging.s3-us-west-2.amazonaws.com/Assets/Isaac/5.0/Isaac/:

   Component	Asset path beneath the prefix
   Conveyor	Props/Conveyors/ConveyorBelt_A06.usd
   Table	Props/Mounts/ThorlabsTable/table_instanceable.usd
   Stand	Props/Mounts/Stand/stand.usd
   Stereo sensor	Sensors/Stereolabs/ZED_X/ZED_X.usdc

4. Keep meshes/retainer_scan/retainer_scan.usdc with the scene so the relative reference resolves. The retainer asset has default prim /teeth_retainer and mesh /teeth_retainer/base_link/retainer_scan. Its existing scene position, orientation, and scale are preserved. Scan coordinates were multiplied by 0.001, matching the supplied URDF's mesh scale, producing local dimensions of approximately 64.1 × 16.2 × 53.0 mm. Physical dimensions have not been independently measured.

5. Inspect the camera and render-product targets. Several saved render products target an absent /World/ur5e/ZED_X/... hierarchy; other products target the sensor under /World/stand/ZED_X_01/.... Select valid camera prims for the workflow you intend to run.

6. For ROS 2 camera output, enable the Isaac Sim ROS 2 bridge and inspect the graph at /World/stand/ZED_X_01/ActionGraph. Confirm the camera target, image dimensions, topic, ROS domain, and frame ID. The authored frame ID is turtle; several other inputs have no explicit values in the root layer.

7. Review asset-resolution and graph errors before starting simulation. Configure collision geometry, physical properties, and controls as needed for your experiment.

Retainer material and URDF source

Both retainer USD formats include a bound LightYellow material using UsdPreviewSurface. The source color from the supplied teeth_retainer.urdf is RGB (228, 209, 119); dividing by 255 gives the authored diffuse color (0.894118, 0.819608, 0.466667). The material is located at /teeth_retainer/Looks/LightYellow and bound to /teeth_retainer/base_link/retainer_scan.

The supplied URDF also specifies 0.001 0.001 0.001 scaling for its visual and collision meshes. The generated USD geometry uses that same scale. The URDF was supplied separately and is not currently included in this repository; neither scene requires it at runtime. The generated retainer assets contain scan geometry, the color material, and the physics setup described below. The unavailable source USD's robot metadata has not been reconstructed.

Retainer physics

Both retainer_scan.usdc and retainer_scan.usda include the same physics configuration:

• /teeth_retainer/base_link: enabled dynamic UsdPhysics.RigidBodyAPI, with kinematic mode disabled and zero initial linear and angular velocity.
• /teeth_retainer/base_link/collision_proxy: simplified mesh with 19,999 triangles, enabled UsdPhysics.CollisionAPI, and UsdPhysics.MeshCollisionAPI set to convexDecomposition. Isaac Sim cooks the convex collision shapes when loading the asset; cooked hulls are not precomputed here.
• /teeth_retainer/base_link/retainer_scan: the original 1,385,355-triangle visual mesh and LightYellow material are preserved. Its collision API is explicitly disabled so the annotation script's add-missing-colliders helper does not enable collisions on the detailed scan.

The collision proxy is hidden and marked as guide geometry. This affects rendering, not its enabled physics collision setting. The proxy was generated using quadric mesh simplification; its local bounds differ from the visual bounds by at most 0.057 mm. Bounds agreement does not establish surface or contact accuracy.

No measured mass, density, friction, restitution, center of mass, or inertia is assigned. The simulator derives unspecified mass properties from collider geometry and its defaults; configure suitable values for the experiment. Replicator_autoannotation_2.py currently overrides mass to 0.05 kg. That value is an experiment setting, not a measured retainer mass.

Physics schema placement, collider enablement, material binding, visual geometry preservation, and USDC/USDA layer equivalence were checked. Convex cooking, contact behavior, settling, and rendered output still require validation in Isaac Sim.

Intended uses

• Research and teaching on digital manufacturing scenes and retainer-handling workcells.
• Scene layout, camera placement, and visualization experiments.
• A starting environment for perception, synthetic-data, or manipulation experiments after dependencies, sensors, annotations, and controls are configured and validated.

These are potential uses of the assets, rather than demonstrated task capabilities. No autonomous pick-and-place behavior or validated retainer-processing pipeline is provided.

Limitations

• Reconstructed retainer: the unavailable retainer USD was replaced by a local asset generated from the bundled STL. The URDF's LightYellow material is applied, and a new dynamic rigid body with a simplified collision proxy is supplied. These physics settings are newly authored; the missing USD's original physics and robot metadata could not be recovered.
• External dependencies: conveyor, table, stand, and sensor assets are referenced remotely rather than bundled. Their availability and transitive dependencies have not been verified.
• Saved scene state: the layer contains multiple render products, including targets for an absent robot hierarchy. Their presence does not establish working stereo capture or ROS 2 publishing.
• Geometry and physics: scan units, acquisition method, dimensional accuracy, physical material properties, and collision suitability are not documented. The conversion uses the supplied URDF's 0.001 scale, interpreted as millimeters to meters; the STL itself does not declare units and physical dimensions remain unverified.
• Generalization: a single scan and scene layout do not establish coverage of different retainers, materials, workcells, lighting conditions, or sensor noise.
• Validation: inspection covered the root USD layer, repository inventory, and preview image. The repaired retainer reference was verified by composing its USD subtree and checking triangle topology and coordinates against the STL. The LightYellow material binding and diffuse color were also verified through the main scene. USDA exports were checked for complete layer equivalence with their binary source layers, and the USDA scene's retainer reference was verified. The scene was not executed in Isaac Sim, and physics, rendering, ROS 2 output, and real-world transfer were not tested.

Training and evaluation

Training data, architecture, hyperparameters, compute, and emissions are not applicable to this asset release because no trained machine-learning model is included. The scan's provenance is not documented. No task success rates, perception metrics, throughput measurements, or other benchmarks are supplied.

Responsible use

The assets do not establish clinical suitability, patient fit, manufacturing accuracy, or safe physical robot operation. Validate geometry and simulation behavior independently for the intended application. If the scan represents an individual's dental geometry, establish its provenance and permission for use and redistribution.

License and attribution

No license file was found in the inspected repository, so this card does not assign a license. Confirm the applicable permissions before reusing or redistributing the scene and scan. Externally referenced NVIDIA and other third-party assets may have separate terms; inclusion of a reference does not grant redistribution rights.

Asset authors, scan provenance, and a preferred citation have not been specified. For questions or corrections, use the repository's Hugging Face community discussions if enabled.