---
pretty_name: RobotAgentBench Training Data
language:
- en
task_categories:
- robotics
size_categories:
- 1M<n<10M
tags:
- robotics
- lerobot
- robotagentbench
- arxiv:2610.34649
configs:
- config_name: default
  data_files:
  - split: train
    path: data/chunk-*/*.parquet
---

<div align="center">

![RobotAgentBench](assets/robotagentbench.png)

</div>

# RobotAgentBench

[Paper](https://arxiv.org/abs/2610.34649) · [Project](https://aetherlabsai.github.io/RobotAgentBench/) · [Code](https://github.com/AetherLabsAI/RobotAgentBench) · [Model](https://huggingface.co/AetherLabs-AI/InternVLA-A1.5-RAB)

**RobotAgentBench: Towards Open-world Robot Autonomy** introduces a benchmark for evaluating robot agents in realistic, open-world household environments. An agent-assisted, human-in-the-loop real-to-sim pipeline constructs household environments with interactive objects and physically grounded tasks.

The task hierarchy comprises atomic tasks, short-horizon tasks, long-horizon tasks, and open-ended scenarios. Evaluation spans task discovery, context reasoning, manipulation, failure recovery, robustness, and autonomous fulfillment of high-level user needs.

## Dataset summary

| Category | Episodes | Frames | Hours |
| --- | ---: | ---: | ---: |
| Atomic | 4,195 | 3,151,810 | 43.775139 |
| Short-horizon | 750 | 2,475,637 | 34.383847 |
| Long-horizon | 70 | 971,943 | 13.499208 |
| Open-ended | 40 | 1,520,713 | 21.121014 |
| **Total** | **5,055** | **8,120,103** | **112.779208** |

The release contains 15,165 camera videos and 203 language instructions. The `train` split contains all 5,055 episodes. Tablecloth-manipulation tasks are excluded.

[Per-source task counts](meta/task_counts.csv) · [Episode duration statistics](meta/duration_summary.csv)

## Data format

Demonstrations use **LeRobot v2.1** at **20 Hz**.

| Feature | Type / shape | Content |
| --- | --- | --- |
| `observation.state` | `float32[23]` | Twenty joint channels, world-frame base `(x, y)`, and unwrapped yaw |
| `action` | `float32[22]` | Twenty joint-position targets, body-forward velocity, and yaw rate |
| `observation.images.top_head` | RGB video, native resolution | Head camera |
| `observation.images.hand_left` | RGB video, native resolution | Left wrist camera |
| `observation.images.hand_right` | RGB video, native resolution | Right wrist camera |
| `task_index` | `int64` | Language instruction in `meta/tasks.jsonl` |
| `timestamp` | `float64` | Episode time in seconds |
| `episode_index`, `frame_index`, `index` | `int64` | Episode, frame, and dataset indices |

Torso lift and base position use metres; angular joint channels and yaw use radians. Base commands use metres per second and radians per second. Channel order and units are recorded in [meta/info.json](meta/info.json).

```text
data/        Episode data in Parquet format
videos/      Head and wrist camera videos
meta/        Tasks, episodes, features, statistics, and provenance
auxiliary/   Original source fields and supporting data
```

Episode provenance and original diagnostic labels are retained in [meta/episode_provenance.jsonl](meta/episode_provenance.jsonl) and [meta/qualification.jsonl](meta/qualification.jsonl). Signal origins and alignment are recorded in [meta/component_origin_index.json](meta/component_origin_index.json) and [meta/alignment_conventions.json](meta/alignment_conventions.json).

## Video format and resolution

In progress: 1,402/1,539 videos in this encoding update. See [encoding manifest](meta/video_encoding.json).

Videos use native camera rectangles, including 640 × 480, 1280 × 800 and 848 × 480. Actual per-episode dimensions are listed in [video dimensions](meta/video_dimensions.jsonl). Spatial dimensions are `null` in global feature metadata when episodes have different sizes. The portable loader resolves a selected episode's concrete shapes in memory. For mixed-resolution training batches, configure your model's image preprocessing or group episodes by resolution.

Updated videos use MP4 / H.264 Main level4.0, 8-bit YUV420, limited-range BT.709 SDR with web-optimized metadata placement. Encoding includes normal chroma subsampling and quantization. Frame counts, timestamps and native image dimensions are retained. Numeric action/state data, task identities and original qualification annotations are unchanged.

The files under `videos/` are individual MP4s.

## Download

```bash
hf download AetherLabs-AI/RobotAgentBench-TrainingData --repo-type dataset --local-dir RobotAgentBench
```

## Load

```bash
python -m pip install "lerobot==0.3.3" av
cd RobotAgentBench
python portable_loader.py --root . --episode 0 --frame 0 --cache-dir ../rab-cache
```

## Citation

```bibtex
@misc{mao2026robotagentbenchopenworldrobotautonomy,
      title={RobotAgentBench: Towards Open-world Robot Autonomy}, 
      author={Lingjun Mao and Jinglin Cao and Lukun He and Wenpeng Xu and Yuchen Yan and Hengyu Jin and Zijun Zhang and Jinzhou Tang and Songyao Jin and Ziche Liu and Yifei Shao and Junbo Huang and Fang Nan and Ruobing Han and Ziqiao Xi and Ziming Xu and Shuang Liang and Sibo Zhu and Wenyi Wu and Xinyue Wang and Yiming Huang and Xinkai Zou and Gao Zhu and Kun Zhou and Biwei Huang},
      year={2026},
      eprint={2610.34649},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.34649}, 
}
```