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

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
Private	Private.jsonl
EBench	EBench.jsonl
Galaxea Open-World	Galaxea_Open-World.jsonl
InternData-A1	InternData-A1.jsonl
LIBERO	LIBERO.jsonl
RoboCasa	RoboCasa.jsonl
RoboCerebra	RoboCerebra.jsonl
RoboDojo	RoboDojo.jsonl
RoboTwin 2.0	RoboTwin_2.0.jsonl
RH20T	RH20T.jsonl
AgiBot World Beta	AgiBot_World_Beta.jsonl
AgiBot World 2026	AgiBot_World_2026.jsonl
All datasets	roboatom_data.jsonl

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

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 (InternData-A1: 30 mapped to 15), 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

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:

{
  "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:

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.

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.

If you use RoboAtom in your research, please cite:

@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