README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - robometer
  - rbm-1m-ood
  - reward-model
  - evaluation

RBM-1M-OOD evaluation dataset used in Robometer. It contains over 1k trajectories used for evaluation of general-purpose reward models.

Dataset Description

Official evaluation in the paper uses only these 6 data sources: usc_trossen, mit_franka, utd_so101, usc_xarm, usc_franka, usc_koch. Reported benchmarks and metrics in the paper are computed on this subset.

The repository may also include trajectories from additional data sources (e.g. utd_so101_wrist, usc_koch_paired, utd_so101_clutter, utd_so101_human, usc_koch_rewind_og) for the full dataset. Filter by data_source to restrict to the 6 eval sources when reproducing paper results.

Each row has a video (MP4) and metadata including task, quality label (success/failure/suboptimal), and data source.

• Paper: https://arxiv.org/abs/2603.02115
• License: Apache 2.0

Dataset Structure

The repo follows the VideoFolder layout with one subset (split) per data source:

• usc_trossen/*.mp4, usc_trossen/metadata.parquet — USC Trossen
• mit_franka/*.mp4, mit_franka/metadata.parquet — MIT Franka
• utd_so101/, usc_xarm/, usc_franka/, usc_koch/ — and other sources
• meta/info.json — dataset schema and split info

Load a single source: load_dataset("videofolder", data_dir="robometer/rbm-1m-ood", split="usc_trossen").

meta/info.json (generated on upload):

{
  "dataset_name": "rbm-1m-ood-eval",
  "total_videos": "<total>",
  "splits": { "usc_trossen": "0:N1", "mit_franka": "0:N2", ... },
  "video_path": "{split}/{file_name}",
  "metadata_path": "{split}/metadata.parquet",
  "eval_data_sources": ["usc_trossen", "mit_franka", "utd_so101", "usc_xarm", "usc_franka", "usc_koch"],
  "data_sources": ["... full list of all sources in repo ..."],
  "features": {
    "id": { "dtype": "string", "description": "Unique identifier (UUID)" },
    "file_name": { "dtype": "string", "description": "Video filename in split folder" },
    "quality_label": { "dtype": "string", "description": "success, failure, or suboptimal" },
    "data_source": { "dtype": "string", "description": "Source dataset" }
  }
}

Key metadata columns:

Column	Description
id	Unique identifier (UUID)
file_name	Video filename (e.g. video_000000.mp4)
quality_label	success, failure, or suboptimal
data_source	Source identifier. Eval (paper): usc_trossen, mit_franka, utd_so101, usc_xarm, usc_franka, usc_koch. Additional sources may appear in the full dataset.

Load with the datasets library (one subset per data source; use split=<data_source>):

from datasets import load_dataset
# Load a single data source (subset)
ds = load_dataset("videofolder", data_dir="robometer/rbm-1m-ood", split="usc_trossen")
# Or load multiple and concatenate
eval_sources = ["usc_trossen", "mit_franka", "utd_so101", "usc_xarm", "usc_franka", "usc_koch"]
ds_list = [load_dataset("videofolder", data_dir="robometer/rbm-1m-ood", split=s) for s in eval_sources]
from datasets import concatenate_datasets
ds_eval = concatenate_datasets(ds_list)

Citation

BibTeX:

@article{liang2026robometer,
  title={Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons},
  author={Liang, Anthony and Korkmaz, Yigit and Zhang, Jiahui and Hwang, Minyoung and Anwar, Abrar and Kaushik, Sidhant and Shah, Aditya and Huang, Alex S. and Zettlemoyer, Luke and Fox, Dieter and Xiang, Yu and Li, Anqi and Bobu, Andreea and Gupta, Abhishek and Tu, Stephen and Biyik, Erdem and Zhang, Jesse},
  journal={arXiv preprint arXiv:2603.02115},
  year={2026}
}