Loading dataset details…
SimRepairBench
CASTAIGNE25 · OpenUSD object repair
Chris von Csefalvay · Pranav Doma · Achyuthan Unni Krishnan · Tamas Foldi
Dataset and objects · Whole objects · Broken objects
Task definitions and verifiers
CASTAIGNE25 contains 25 repair tasks with 128 interacting faults in articulated tools, mechanisms, containers and assemblies. Each task supplies a broken OpenUSD object, its resources, a repair brief and evidence of the intended object.
Run
harbor run --dataset hcltech-robotics/simrepairbench@0.2.0 \
--agent <agent> --model <model> --n-concurrent 1Use an agent that supports inference outside the task container. The task has no
network access; Python 3.12, OpenUSD 26.8 and candidate diagnostics are installed.
Submit the repaired object and all required resources in /workspace/submission.
Harbor transfers that directory to a separate verifier with the source reference.
Scoring
The numeric authored_usd reward combines strict target recovery and continuous
semantic measurements, with a healthy-content preservation gate. Checks cover
visible geometry, joint connectivity and motion coordinates, collision authoring,
mass properties and material networks. Equivalent valid USD authoring is allowed.
Deleting healthy components or disabling their physics does not earn repair credit.
Full credit requires all measured objectives and preservation checks to pass. Dynamic contact, physical stability and rendered or LLM judgments are separate grades; they are not included in the authored-USD reward.
Attribution
Source objects are from NVIDIA SimReady Assets, licensed under CC BY 4.0, with source notices retained in each package. The MIT-licensed SimReadyBench evaluator provides reused USD observation and validation utilities. SimRepairBench supplies the broken objects, repair tasks and task-specific objective verifiers.
SimRepairBench software is MIT-licensed; its task documentation is CC BY 4.0. The complete licence text accompanies this dataset and every task.
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}
}