Sterzhang/video2skill-bench

Video2Skill: From Streaming Experience to Reusable Embodied Skills

Jianshu Zhang^1*, Ce Zhang^2*, Xiyuan Yang³, Chenwei Xu¹, Haoran Lu¹, Yijiang Li⁴, Yaqi Xie², Katia P. Sycara², Han Liu^1
1Northwestern University   ²Carnegie Mellon University   ³University of Illinois Urbana-Champaign   ⁴University of California, San Diego
^*Equal contribution

📄 arXiv:2609.36691  |  🌐 Project Page  |  💻 Evaluation code

The Video2Skill benchmark for Streaming Embodied Skill Discovery (SESD): the two verified test splits used in the paper, HD-EPIC and RoboInter, with their video clips. The evaluation code reproduces the paper's main table.

The task

A model watches a stream of video clips in a fixed order. For each manipulation event it must find when the event happens, express it as a skill call such as grasp(object=onion, source=produce rack), and decide whether to reuse a skill already in its library or add a new one. The library starts empty and carries over from clip to clip. The model never sees the reference class list and names its skills freely; scoring does not depend on names.

Data

Split	Domain	Clips	Video	Reference events	Classes	Median event
HD-EPIC	Egocentric kitchen activity, 29 held-out recordings	40	117.6 min	1,137	46	1.9 s
RoboInter	Robot tabletop manipulation (98 DROID + 34 RH20T clips)	132	60.0 min	534	15	4.5 s

video2skill-bench/
├── test/
│   ├── hdepic.jsonl      # 40 clips in the order models must process them
│   └── robointer.jsonl   # 132 clips in the order models must process them
└── videos/
    ├── hdepic/           # 40 clips (5.6 GB)
    └── robointer/        # 132 clips (48 MB)

Each line of a test file is one clip. The order of lines is the processing order used in the paper (the balanced order). Times are seconds from the start of the clip.

{"clip_id": "ri__robointer__17991_exterior_image_1_left",
 "video": "videos/robointer/ri__robointer__17991_exterior_image_1_left.mp4",
 "duration": 17.5, "fps": 10.0,
 "events": [{"start": 0.0, "end": 11.2, "skill": "grasp", "call": "grasp(object=lid, source=pot)"},
            {"start": 11.3, "end": 13.2, "skill": "transfer", "call": "transfer(object=lid, source=pot, destination=table)"}]}

• skill is the reference class. Scoring uses it to decide which events are the same skill.
• call is the verified skill call, with arguments from seven roles: object, source, destination, instrument, direction, substance, and state.

Evaluation

Predictions are one line per clip: {"clip_id": ..., "events": [{"start": ..., "end": ..., "skill": ...}]}. Score them with score.py:

python score.py --test test/hdepic.jsonl --pred my_model_hdepic.jsonl

It reports the six metrics of the paper: Cov., Pair P, Pair R, ARI, Create R, and Reuse R. The same folder has the outputs of the 19 models in the paper's main table and a script that reproduces the table exactly.

from huggingface_hub import snapshot_download
root = snapshot_download("Sterzhang/video2skill-bench", repo_type="dataset",
                         allow_patterns=["test/*"])   # drop allow_patterns to also get the videos

Dataset construction

The clips come from HD-EPIC and from RoboInter, which draws robot videos from DROID and RH20T. Every candidate event was checked against the video: errors were corrected, and labels whose manipulation is not visible were removed. The test clips come from source videos that are disjoint from the training data used in the paper.

License and attribution

• The annotations in this repository are released under CC BY-NC 4.0 for non-commercial research use.
• The video clips are trimmed from third-party datasets and remain under their original terms. Please cite and follow the upstream licenses:
  • HD-EPIC (40 clips): CC BY-NC 4.0. The clips are trimmed (and re-encoded) segments of the original recordings; no other modification was made.
  • RH20T (34 clips): scenes 0001–0005 (16 clips) are CC BY-SA 4.0; scenes 0006–0010 (18 clips) are CC BY-NC 4.0.
  • DROID (98 clips): see the DROID project page for its release terms.
• If you are a rights holder and have concerns about any content, please contact the authors.

Citation

@misc{zhang2026video2skillstreamingexperiencereusable,
      title={Video2Skill: From Streaming Experience to Reusable Embodied Skills},
      author={Jianshu Zhang and Ce Zhang and Xiyuan Yang and Chenwei Xu and Haoran Lu and Yijiang Li and Yaqi Xie and Katia P. Sycara and Han Liu},
      year={2026},
      eprint={2609.36691},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.36691},
}

Please also cite HD-EPIC, DROID, and RH20T when using the video clips.

Contact

Jianshu Zhang (andyzhang [at] u.northwestern.edu). Questions and issues are welcome in the Community tab.