---

license: mit
pretty_name: RoboAtom
language:

- en
task_categories:
- robotics
size_categories:
- 100K<n<1M
tags:
- robotics
- atomic-actions
- temporal-action-segmentation
- video
configs:
  - config_name: RoboAtom-Data
    data_files:
      - split: train
        path: RoboAtom-Data/roboatom_data.jsonl
  - config_name: RoboAtom-Bench
    default: true
    data_files:
      - split: test
        path: RoboAtom-Bench/benchmark.jsonl

---

# RoboAtom: Large-scale Atomic-Action Supervision for Robot Learning

![RoboAtom overview](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/figures/roboatom-overview.png)

<p align="center">
  <a href="https://internrobotics.github.io/RoboAtom/" style="display: inline-block; vertical-align: middle; line-height: 0;"><img src="https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/badges/website.svg" alt="Project page" width="146" height="36" style="display: inline-block; margin: 0; vertical-align: middle;"></a>
  <a href="https://github.com/InternRobotics/RoboAtom" style="display: inline-block; vertical-align: middle; line-height: 0;"><img src="https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/badges/code.svg" alt="GitHub" width="112" height="36" style="display: inline-block; margin: 0; vertical-align: middle;"></a>
  <a href="https://arxiv.org/abs/2610.40207" style="display: inline-block; vertical-align: middle; line-height: 0;"><img src="https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/badges/paper.svg" alt="Paper" width="102" height="36" style="display: inline-block; margin: 0; vertical-align: middle;"></a>
  <a href="https://huggingface.co/datasets/InternRobotics/RoboAtom" style="display: inline-block; vertical-align: middle; line-height: 0;"><img src="https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/badges/data.svg" alt="Dataset" width="122" height="36" style="display: inline-block; margin: 0; vertical-align: middle;"></a>
  <a href="https://huggingface.co/InternRobotics/RoboAtom" style="display: inline-block; vertical-align: middle; line-height: 0;"><img src="https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/project_media/badges/models.svg" alt="Models" width="116" height="36" style="display: inline-block; margin: 0; vertical-align: middle;"></a>
</p>

## RoboAtom-Data

Time-aligned atomic-action annotations: each segment has a skill label, an action description, and frame boundaries. Each dataset has one JSONL file; `roboatom_data.jsonl` is their concatenation.


| Dataset            | Annotation file                                                                                                                                    |
| ------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------- |
| EBench             | [EBench.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/EBench.jsonl)                         |
| Galaxea Open-World | [Galaxea_Open-World.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/Galaxea_Open-World.jsonl) |
| InternData-A1      | [InternData-A1.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/InternData-A1.jsonl)           |
| LIBERO             | [LIBERO.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/LIBERO.jsonl)                         |
| RoboCasa           | [RoboCasa.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/RoboCasa.jsonl)                     |
| RoboCerebra        | [RoboCerebra.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/RoboCerebra.jsonl)               |
| RoboDojo           | [RoboDojo.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/RoboDojo.jsonl)                     |
| RoboTwin 2.0       | [RoboTwin_2.0.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/RoboTwin_2.0.jsonl)             |
| RH20T              | [RH20T.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/RH20T.jsonl)                           |
| AgiBot World Beta  | [AgiBot_World_Beta.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/AgiBot_World_Beta.jsonl)   |
| AgiBot World 2026  | [AgiBot_World_2026.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/by_dataset/AgiBot_World_2026.jsonl)   |
| **All datasets**   | [roboatom_data.jsonl](https://huggingface.co/datasets/InternRobotics/RoboAtom/resolve/main/RoboAtom-Data/roboatom_data.jsonl)                      |


```bash
hf download InternRobotics/RoboAtom --repo-type dataset \
  --include 'RoboAtom-Data/roboatom_data.jsonl' --local-dir ./data
```

```python
from datasets import load_dataset

annotations = load_dataset(
    "InternRobotics/RoboAtom", "RoboAtom-Data", split="train", streaming=True
)
```


| Fields                                      | Meaning                                                |
| ------------------------------------------- | ------------------------------------------------------ |
| `id`, `dataset`, `task_id`, `episode_index` | Trajectory and task identities                         |
| `task`, `task_flags`                        | Task text and task-text flags                          |
| `num_frames`, `fps`, `views`, `domain`      | Frame count, FPS, camera names, real/sim               |
| `segments`                                  | Ordered `{start_frame, end_frame, skill, action_text}` |
| `is_human_annotated`                        | Human-QC status: `true`, `false`, or `null`            |
| `public_source_url`, `public_location`      | Source repository and source-relative path, or `null`  |


Frame indices are zero-based and both endpoints are included. The files contain annotations only; get videos and sensor data from the source datasets.

Skills: `Pick`, `Place`, `Handover`, `Move`, `Open`, `Close`, `Push`, `Pull`, `Press`, `Rotate`, `Screw`, `Plug`, `Pour`, `Fill`, `Scoop`, `Stir`, `Dip`, `Shake`, `Fold`, `Tie`, `Wipe`, `Sweep`, `Wash`, `Brush`, `Cut`, `Peel`, `Strike`, `Scan`, `Throw`. An empty `skill` marks an unlabeled span.

## RoboAtom-Bench

A model predicts ordered skill labels, action descriptions, and temporal boundaries for complete robot trajectories.


| Subset               | Trajectories | GT segments | Videos  |
| -------------------- | ------------ | ----------- | ------- |
| Short (1–4 segments) | 200          | 411         | 200     |
| Medium (5–8)         | 100          | 634         | 100     |
| Long (9–62)          | 100          | 2,302       | 100     |
| Multiview            | 100          | 671         | 321     |
| **Total**            | **500**      | **4,018**   | **721** |


```bash
hf download InternRobotics/RoboAtom --repo-type dataset \
  --include 'RoboAtom-Bench/**' --local-dir ./data
```

`benchmark.jsonl` holds model inputs and relative video paths; `gt.jsonl` holds `action_config` segments with the same `benchmark_id`. An input record:

```json
{
  "benchmark_id": "benchmark:004957f135e186fb5052",
  "category": "short",
  "dataset": "RoboMIND 2.0",
  "task": "Transfers the tape from the board to the tray using both arms.",
  "fps": 30.0,
  "length": 242,
  "frame_interval": "closed_inclusive",
  "benchmark_view_mode": "singleview",
  "evaluation_views": ["front"],
  "model_view_key": "front",
  "views": {"front": "videos/004957f135e186fb5052/00_front.mp4"}
}
```

Validate, run inference, and score with the [RoboAtom code](https://github.com/InternRobotics/RoboAtom#quick-start-benchmark):

```bash
roboatom evaluate \
  --gt ./data/RoboAtom-Bench/gt.jsonl \
  --benchmark-input ./data/RoboAtom-Bench/benchmark.jsonl \
  --prediction outputs/roboatom_predictions.jsonl \
  --protocol configs/roboatom_bench.yaml \
  --output-dir outputs/roboatom_evaluation
```

Metrics: Micro F1 at 0.3/0.5/0.7, Skill F1@0.7, mIoU, and Edit.

## 🏆 Leaderboard

<details>
<summary>📦 RoboAtom-Bench</summary>

| Model | mIoU ↑ | F1@0.7 ↑ | Skill F1@0.7 ↑ | Edit ↑ | Type |
| --- | ---: | ---: | ---: | ---: | --- |
| 🥇 <nobr>RoboAtom-Anno-27B</nobr> | 56.4 | 43.2 | 33.8 | 76.8 | Open-Source |
| 🥈 <nobr>RoboAtom-Anno-9B</nobr> | 55.1 | 41.0 | 30.1 | 76.7 | Open-Source |
| 🥉 GPT-6-astra | 53.6 | 34.5 | 29.2 | 74.0 | Proprietary |
| GPT-6.1-sol | 51.2 | 29.8 | 22.8 | 64.6 | Proprietary |
| GPT-6-sol | 47.7 | 30.4 | 23.8 | 67.0 | Proprietary |
| GPT-5.6-sol | 44.0 | 28.0 | 21.4 | 63.4 | Proprietary |
| Gemini 3.1 Pro | 42.5 | 24.4 | 16.1 | 61.8 | Proprietary |
| Qwen3.8-27B | 38.4 | 20.1 | 10.9 | 56.6 | Open-Source |
| GPT-4o | 31.5 | 13.9 | 7.2 | 47.7 | Proprietary |
| Qwen3.5-9B | 9.6 | 6.2 | 3.5 | 28.7 | Open-Source |

Results on all 500 RoboAtom-Bench trajectories, ranked by mIoU in descending order. All metrics are reported on a 0–100 scale; higher is better.

</details>

Subset results use the same predictions as the overall leaderboard. Short, Medium, Long, and Multiview contain 200, 100, 100, and 100 trajectories, respectively. Each table is ranked by mIoU. mIoU pools GT segments; both F1 metrics pool TP/FP/FN; Edit is averaged over trajectories.

<details>
<summary>📦 Short</summary>

| Model | mIoU ↑ | F1@0.7 ↑ | Skill F1@0.7 ↑ | Edit ↑ | Type |
| --- | ---: | ---: | ---: | ---: | --- |
| 🥇 <nobr>RoboAtom-Anno-9B</nobr> | 81.9 | 77.4 | 71.1 | 83.9 | Open-Source |
| 🥈 <nobr>RoboAtom-Anno-27B</nobr> | 77.3 | 77.3 | 71.7 | 81.9 | Open-Source |
| 🥉 GPT-6-astra | 75.8 | 53.6 | 48.8 | 74.8 | Proprietary |
| GPT-6-sol | 72.1 | 47.1 | 43.2 | 71.6 | Proprietary |
| GPT-6.1-sol | 69.3 | 38.1 | 29.6 | 60.8 | Proprietary |
| Gemini 3.1 Pro | 68.7 | 39.6 | 31.9 | 61.8 | Proprietary |
| GPT-5.6-sol | 68.3 | 40.2 | 31.9 | 64.9 | Proprietary |
| Qwen3.8-27B | 65.5 | 37.0 | 23.2 | 62.3 | Open-Source |
| GPT-4o | 57.7 | 20.7 | 15.3 | 56.9 | Proprietary |
| Qwen3.5-9B | 35.5 | 29.2 | 18.0 | 47.5 | Open-Source |

</details>

<details>
<summary>📦 Medium</summary>

| Model | mIoU ↑ | F1@0.7 ↑ | Skill F1@0.7 ↑ | Edit ↑ | Type |
| --- | ---: | ---: | ---: | ---: | --- |
| 🥇 <nobr>RoboAtom-Anno-27B</nobr> | 63.6 | 53.3 | 49.9 | 76.9 | Open-Source |
| 🥈 GPT-6-astra | 62.9 | 39.4 | 35.1 | 79.3 | Proprietary |
| 🥉 <nobr>RoboAtom-Anno-9B</nobr> | 61.9 | 48.8 | 43.7 | 78.2 | Open-Source |
| GPT-6-sol | 61.9 | 37.5 | 32.6 | 74.6 | Proprietary |
| GPT-6.1-sol | 59.1 | 34.0 | 27.8 | 67.3 | Proprietary |
| GPT-5.6-sol | 56.7 | 35.8 | 29.4 | 67.8 | Proprietary |
| Gemini 3.1 Pro | 54.8 | 37.6 | 28.5 | 70.0 | Proprietary |
| Qwen3.8-27B | 48.2 | 24.9 | 16.8 | 57.9 | Open-Source |
| GPT-4o | 47.4 | 22.1 | 11.4 | 48.4 | Proprietary |
| Qwen3.5-9B | 10.5 | 4.5 | 2.6 | 15.2 | Open-Source |

</details>

<details>
<summary>📦 Long</summary>

| Model | mIoU ↑ | F1@0.7 ↑ | Skill F1@0.7 ↑ | Edit ↑ | Type |
| --- | ---: | ---: | ---: | ---: | --- |
| 🥇 <nobr>RoboAtom-Anno-27B</nobr> | 50.8 | 33.8 | 21.4 | 63.5 | Open-Source |
| 🥈 <nobr>RoboAtom-Anno-9B</nobr> | 47.9 | 30.9 | 17.3 | 62.4 | Open-Source |
| 🥉 GPT-6-astra | 46.7 | 28.4 | 22.7 | 67.4 | Proprietary |
| GPT-6.1-sol | 45.8 | 26.2 | 18.9 | 64.5 | Proprietary |
| GPT-6-sol | 39.3 | 24.4 | 17.4 | 56.3 | Proprietary |
| GPT-5.6-sol | 35.5 | 22.7 | 16.6 | 55.8 | Proprietary |
| Gemini 3.1 Pro | 32.3 | 14.8 | 7.2 | 55.3 | Proprietary |
| Qwen3.8-27B | 29.6 | 14.6 | 6.3 | 45.7 | Open-Source |
| GPT-4o | 22.5 | 9.8 | 3.6 | 33.7 | Proprietary |
| Qwen3.5-9B | 3.4 | 1.4 | 0.4 | 5.4 | Open-Source |

</details>

<details>
<summary>📦 Multiview</summary>

| Model | mIoU ↑ | F1@0.7 ↑ | Skill F1@0.7 ↑ | Edit ↑ | Type |
| --- | ---: | ---: | ---: | ---: | --- |
| 🥇 <nobr>RoboAtom-Anno-9B</nobr> | 56.8 | 44.1 | 33.9 | 75.2 | Open-Source |
| 🥈 <nobr>RoboAtom-Anno-27B</nobr> | 56.1 | 45.3 | 37.8 | 80.0 | Open-Source |
| 🥉 GPT-6-astra | 54.8 | 34.4 | 29.3 | 73.8 | Proprietary |
| GPT-6.1-sol | 51.1 | 30.1 | 24.5 | 69.7 | Proprietary |
| Gemini 3.1 Pro | 49.7 | 27.2 | 17.0 | 60.3 | Proprietary |
| GPT-6-sol | 48.1 | 28.1 | 18.2 | 60.6 | Proprietary |
| GPT-5.6-sol | 46.6 | 25.6 | 18.5 | 63.6 | Proprietary |
| Qwen3.8-27B | 42.9 | 17.6 | 8.4 | 55.0 | Open-Source |
| GPT-4o | 31.7 | 10.7 | 5.7 | 42.7 | Proprietary |
| Qwen3.5-9B | 14.1 | 5.9 | 3.7 | 27.8 | Open-Source |

</details>

## License and citation

Annotations, benchmark labels, and project media are released under the MIT license. Source videos and datasets keep their own terms. Citation: [CITATION.cff](https://github.com/InternRobotics/RoboAtom/blob/main/CITATION.cff).

If you use RoboAtom in your research, please cite:

```bibtex
@misc{wei2026roboatomlargescaleatomicactionsupervision,
      title={RoboAtom: Large-scale Atomic-Action Supervision for Robot Learning},
      author={Xueyuan Wei and Xuanxu Wang and Xudong Li and Yuping He and Wanshui Gan and Jiaqi Peng and Wenqi Guo and Jianjun Zhou and Junting Dong and Chunhua Shen and Haoyu Guo and Mulin Yu and Tao Lu},
      year={2026},
      eprint={2610.40207},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.40207},
}
```



## To-do

- [ ] ABC-130k
- [ ] RoboCOIN
- [ ] RoboMIND
- [ ] RoboMIND 2.0