SimRepairBench

Repairing simulation-ready OpenUSD objects

CASTAIGNE25 · 25 modelling-error repair tasks

Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan and Tamas Foldi

HCLTech

SimRepairBench is a benchmark for repairing OpenUSD models used in robotic simulation. Its CASTAIGNE25 sample contains 25 tasks drawn from workshop equipment, tools, medical furniture and household objects. Each task starts with a complete object whose representation has been deliberately corrupted. The model inspects the supplied USD package and observation, then repairs the error while keeping the healthy parts and their behaviour.

A joint may turn about the wrong axis, mesh faces may connect to the wrong vertices, or a surface may have lost its material assignment. The correctly modelled object can still be dented, worn or partly open. Those features belong to the object and should survive the repair.

Browse the cases · Harbor tasks · Scoring · Dataset format · Model and attempt records · SimReadyBench

Objects and repair tasks

The cases use the identifiers R01–R25. Release 0.8.0 contains all 25 tasks and 28 modelling errors. The vise, rack and drum each combine two errors in the same object. The case definitions describe what needs repairing, and the release index links each case to its Harbor task. A task revision fixes the input package, repair requirements and verifier so that models can be compared on the same problem.

Objects	Modelling errors
Articulated equipment and guided mechanisms	Joint ownership, local frames, motion axes and contact geometry
Conveyor, ladder and folding knife	Incorrect face-to-vertex connectivity
Pallet jack and patient chair	Incorrect effective surface normals
Dishwasher, drill, car door and drum	Shader resolution and surface assignment

The viewer shows the correctly modelled reference beside the broken input. You can browse both as complete USD packages or download their archives. The reference images show the correctly modelled object, while each broken image comes from that task's supplied observation. Some show the result of a recorded physical test, such as a drop or a pull at a handle. The image_state field identifies these observations, while the USD file retains the object's starting pose. The native 3D gallery opens the broken inputs and uses square thumbnails from their observations.

During a repair, the agent receives the broken package, its observation and the repair brief. The reference and grading code belong in the separate verifier environment. Keep the public reference out of the agent's environment when running a model comparison.

Use the data

from datasets import load_dataset

cases = load_dataset(
    "hcltech-robotics/simrepairbench", "cases", split="castaigne25"
)
case = cases[0]
print(case["case_id"], case["title"], case["faults"])

Each row includes the reference and broken images, a list of faults and the repair brief. It also links to both USD entrypoints, the complete archives and the versioned Harbor task. The verification_scope list describes the checks used for that case.

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    "hcltech-robotics/simrepairbench",
    case["package_path"],
    repo_type="dataset",
)

Extract the whole archive so that textures and other resources keep their relative paths. Open asset.usdc as the broken entrypoint and use visual-observation.png as the initial observation. Read the file named by repair_brief for what the submitted package must restore and preserve. To run the task through Harbor, use the published collection, which includes the agent environment and a separate verifier.

Evaluation

The repair reward measures the submitted asset on a scale from 0 to 1. The verifier checks that the package is valid and that healthy content survives, then awards partial credit for the corrections it can measure. Surface tasks measure recovered area. Alignment tasks measure how far the error has been reduced. Some requirements, such as resolving a shader program, have a categorical result.

The vise, rack and drum combine two equally weighted requirements. Within each requirement, half the credit comes from meeting its tolerance and half from the measured partial correction. A bounded material-redundancy penalty also applies to the drum. Repair scoring explains these rules and the preservation gates.

The repair harness records the score after every model turn alongside the response, tool outputs and resulting package. After edits, it returns qualitative check progress so the model can decide what to fix next. Read-only turns reuse a measurement only when the complete package is unchanged. Missing measurements remain unmeasured. The final reward always describes the final submitted state.

Per-turn scoring and feedback require the SimRepairBench repair-progress integration. A standard or hosted Harbor agent can use the same dataset with terminal verification. Publishing the tasks does not enable that integration automatically.

Keep task_revision and model settings fixed when comparing attempts. Each attempt links its broken input to the trajectory, submitted USD and repair report. Episode records describe these connections. The current evaluation uses the objective repair reward and requires neither a separate diagnosis submission nor a model judge.

CASTAIGNE25 is a public sample, so comparisons should report prior exposure to its objects or repair traces. Evaluation on unseen objects requires separate asset lineages, with near-duplicate geometry excluded from training. The per-turn record makes it possible to follow each correction from the supplied broken object to the final repair.

Attribution and licence

The objects come from NVIDIA's SimReady asset collection, pinned to revision 4763fbfaa4f8e97658e68b57d01b4c86659e08b5. They retain their Creative Commons Attribution 4.0 licence and package notices. When redistributing them, retain NVIDIA's attribution and notices, and mark packaging changes, injected modelling errors and repairs as modifications. Inclusion in SimRepairBench does not imply NVIDIA endorsement.

Documentation and dataset records use CC BY 4.0, while SimRepairBench software is licensed under MIT. Reused SimReadyBench evaluator code retains its original MIT notice.

Citation

@misc{simrepairbench2026,
  title = {SimRepairBench: CASTAIGNE25 OpenUSD Asset Repair Tasks},
  author = {von Csefalvay, Chris and Doma, Pranav and
            Unni Krishnan, Achyuthan and Foldi, Tamas},
  organization = {HCLTech},
  year = {2026},
  url = {https://huggingface.co/datasets/hcltech-robotics/simrepairbench}
}