SimRepairBench

Repairing simulation-ready OpenUSD objects

CASTAIGNE25 · 25 object repair tasks

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

HCLTech

SimRepairBench provides OpenUSD repair tasks for robotic simulation. Its public sample, CASTAIGNE25, contains 25 deliberately damaged objects, from articulated workshop equipment to everyday tools. An object can look convincing while its joints, collision geometry or dependencies are wrong; the task is to restore the behaviour specified in its brief without losing the parts that still work.

The sample supports Wilde and other repair agents by keeping each problem connected to the work done to solve it. Every task starts from a supplied damaged package and defines what a satisfactory repair must preserve or restore. The case viewer pairs the original object with its broken input, using labelled diagnostic overlays where an internal fault would otherwise be difficult to see.

Browse the cases · Harbor tasks · Case definitions · Dataset format · SimReadyBench

Cases and faults

CASTAIGNE25 contains 128 authored faults across 25 objects. The identifiers R01–R25 connect each object to its fault descriptions and repair requirements, while identifiers such as R01-D01 distinguish the individual faults within a case.

Objects	Repair challenges
Jacks, support arm, conveyor and ladder	Joint graphs, local frames, travel limits and moving clearances
Tool cabinet, medical cart, X-ray table and dishwasher	Component correspondence, telescoping mechanisms and obstructed travel
Fan, vise, drill, pliers, wrench and knife	Axes, pivots, thin geometry and mass properties
Bed, patient chair and locker unit	Repeated parts, hierarchy, missing components and independent motion
Brake assembly, parts washer, car door, rack, drum and engine	Collision cavities, attachments, material bindings and asset dependencies

The faults represent mistakes such as a partial unit conversion, a namespace migration or a transform baked into the wrong part of a hierarchy. Several faults can interact within one object, so correcting an isolated property is not necessarily enough to restore the mechanism.

Use the data

Load the case table with Hugging Face Datasets and select an object:

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"])

The row brings together the previews, fault descriptions, repair requirements and package location. package_path identifies the archive and package_sha256 records its digest. Download the complete archive to retain the relative paths of the object's accompanying resources.

from huggingface_hub import hf_hub_download

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

Each archive contains asset.usdc, its resources, instruction.md, evidence.json, NVIDIA notices and a per-file SHA-256 manifest. Start with the instructions and evidence, then use the case's required_outcome and scope to determine what the repair must achieve. The verification_scope array describes the checks recorded for the input package. Dataset format explains the fields in more detail.

Repair environments and episodes

The Harbor dataset, version 0.2.0, pins all 25 tasks and their separate verifiers. Download the immutable collection with Harbor:

harbor download hcltech-robotics/simrepairbench@sha256:8d7b1e3a7e2c97f8021beeb9f1cf7d401a06629caf1bea7382d7c86118c6ff93 --export --output-dir simrepairbench-0.2.0

Task versions connect each case to its published task and verifier hashes. Each actor receives the damaged object, tools and brief; reference data and scoring code remain in the separate verifier environment. To assemble the same environments from source:

python scripts/export_harbor_tasks.py --materialize --fetch --case all

The objective reward combines per-fault correctness with graded progress toward the required state. It preserves partial credit when other faults remain, while invalid submissions and damage to healthy parts receive zero. Repair scoring describes the measurements, preservation rules and separate physical and visual grades.

A repair episode retains the connection between this starting package and the final candidate. Stable case and episode identifiers link the trajectory, submitted asset, objective measurements, visual assessment and result analysis. Keep the input unchanged and evaluate the exact final bytes; unsuccessful attempts belong in the record too. The trajectory schema describes the observable episode data, and visual assessment explains the separate judge result.

CASTAIGNE25 is public training and demonstration material for Wilde and other repair agents. For unseen evaluation, use separate asset lineages and exclude near-duplicate geometry from the training material. Keeping the input, actions and independently assessed result together lets a reader follow the same object from its original repair problem to the submitted solution, and see whether the behaviour in its brief has been restored.

Attribution and licence

The objects come from NVIDIA's SimReady asset collection, pinned to revision 4763fbfaa4f8e97658e68b57d01b4c86659e08b5. The source assets retain their Creative Commons Attribution 4.0 licence and package notices. When distributing a SimRepairBench derivative, identify NVIDIA as the source, retain those notices and mark source corrections, injected faults and repairs as modifications. Inclusion in this dataset does not imply NVIDIA endorsement.

Documentation and dataset records use CC BY 4.0. SimRepairBench software uses MIT, and 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}
}