Dataset Viewer
Auto-converted to Parquet Duplicate
definition
stringclasses
1 value
frames
dict
"source frames 0, round(fps), 2*round(fps), ... < frames; no extra endpoint (rewardbox SamplingSpec((...TRUNCATED)
{"table30_v2_2945_12869c56-5cd0-43ef-a65b-c8df2c341bdc__main__videos_front":[0,30,60,90,120,150,180,(...TRUNCATED)

RRM-Bench-1K

RRM-Bench-1K tests whether a robot reward model follows the task state of a real robot execution, rather than elapsed time or motion. It accompanies the survey Robotic Reward Models in the Foundation-Model Era and is released together with the RewardBox toolbox (RRM-Hub).

  • 1,117 real-robot videos of closed-loop policy executions recorded on RoboChallenge Table30 and Table30 V2.
  • Task-state labels: a progress curve per clip that rises as effects are achieved and falls when progress is lost.
  • Camera-evidence labels: whether each camera view shows enough to judge the task (sufficient, partial, or insufficient evidence).
  • Four contrast suites that separate the credited condition from its usual cues: (A) Action–Effect, (B) Semantic/Ordered Constraint, (C) Persistence/Reversal, and (D) Visibility.

Files

inputs/     model inputs (single-view and three-camera), labels, event marks, evaluation grid
episodes/   the videos of each clip, with frame maps to the original recordings
tools/      the official evaluator and its tests

The file formats, the evaluation protocol, and the metric definitions are in the benchmark README of RewardBox.

Usage

Download the data, then run any RewardBox model and score it with the official evaluator:

huggingface-cli download Ethanliang99/RRM-Bench-1K --repo-type dataset --local-dir RRM-Bench-1K
cd RRM-Bench-1K
rewardbox bench --manifest inputs/single_view.jsonl --model MODEL \
    --strategy fps --fps 1 --no-include-last -o runs/MODEL
python tools/evaluate.py --predictions runs/MODEL/MODEL.json --name MODEL -o runs/MODEL/eval.json

To evaluate your own model, write its predictions in the format described in the benchmark README and run tools/evaluate.py on them.

License

The data follows RoboChallenge's license, CC BY-NC-SA 4.0: non-commercial use, with attribution, and derivatives shared under the same terms. The code in tools/ is released under the Apache License 2.0, like the rest of RewardBox.

Acknowledgement

All videos in RRM-Bench-1K come from RoboChallenge and record policy executions on its Table30 and Table30 V2 real-robot benchmarks. We thank the RoboChallenge team for building the platform and for making its runs public.

Citation

@article{liang2026rrm,
  title   = {Robotic Reward Models in the Foundation-Model Era},
  author  = {Liang, Guoqiang and Lin, Yiqi and Zhou, Shangchen and
             Zeng, Wenzheng and Chen, Yanzhe and Bai, Zechen and
             Shou, Mike Zheng},
  journal = {arXiv preprint arXiv:2611.20741},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.20741}
}

Please also cite RoboChallenge, the source of the videos:

@article{yakefu2025robochallenge,
  title   = {RoboChallenge: Large-scale Real-robot Evaluation of Embodied Policies},
  author  = {Yakefu, Adina and Xie, Bin and Xu, Chongyang and others},
  journal = {arXiv preprint arXiv:2510.17950},
  year    = {2025}
}
Downloads last month
20

Papers for Ethanliang99/RRM-Bench-1K