---
license: cc-by-nc-sa-4.0
language:
- en
task_categories:
- robotics
- video-classification
tags:
- robotics
- robot-manipulation
- atomic-skills
- skill-segmentation
- video-understanding
- robointer
- automark
pretty_name: RoboInter Atomic-Skill Segment Annotations (AutoMark)
size_categories:
- 100K<n<1M
---

# RoboInter Atomic-Skill Segment Annotations (AutoMark)

Atomic-skill segment **annotations only** for robot-manipulation videos, curated by the
[**AutoMark**](https://github.com/Chenwei-1999/AutoMark) project. Each robot episode is segmented
into atomic primitive-skill intervals (e.g. `pick`, `place`, `pour`) with second-aligned timestamps
and `verb(object)`-style call targets, over the 15-label RoboInter primitive-skill taxonomy.

> ⚠️ **This dataset contains annotations only — NO raw videos.** The labels are *derived from*
> [`InternRobotics/RoboInter-Data`](https://huggingface.co/datasets/InternRobotics/RoboInter-Data),
> which is **access-gated** (you must accept its Community License + Privacy Policy) and licensed
> **CC BY-NC-SA 4.0**. To obtain the actual `.mp4` videos referenced here, request access to
> RoboInter-Data directly and rebuild with the AutoMark export scripts. The `file_name` /
> `video_path` fields are RoboInter video identifiers, not redistributed media.

## Scale

| | |
|---|---|
| Episodes (primary/exterior camera) | **235,880** |
| Atomic-skill segment annotations | **758,397** |
| Skill taxonomy | 15 labels (RoboInter primitive skills) |
| Source FPS | 10.0 (dataset); timestamps in seconds |

**Skill taxonomy:** `transfer, pick, place, press, push, pull, twist, pour, fold, slide, insert,
shake, strike, throw, manipulate`

## Files

- **`annotation.json`** — the full folder-dataset annotation. Top-level metadata
  (`skill_taxonomy`, `dataset_fps`, counts) plus `videos[]`; each video has `file_name`, `fps`,
  `source` (`droid`/`rh20t`), `source_member_path` (path *within* the RoboInter release), and an
  `annotations[]` list of segments: `{start, end, frame_start, frame_end, skill, call, text}`.
  `call` is a `verb(args)`-style function string (args may be empty when the rule-based parser
  could not bind an object).
- **`qwen/text.jsonl`** — 758,397 text-only evaluation rows for skill/call prediction (one segment
  per row): `text` (segment hint), `gt_skill`, `target_call`, `allowed_skills`, and a ready `prompt`.
- **`qwen/visual_smoke_1024.jsonl`** — a 1,024-row visual smoke subset; same schema plus a relative
  `video_path` (`video/<video_id>.mp4`) for grounding against the (separately obtained) videos.
- **`qwen/report.json`** — build report (row counts, taxonomy).

Filesystem paths have been relativized; no absolute paths are included.

## Provenance & licensing

- Curated by **AutoMark** (online atomic-skill induction): <https://github.com/Chenwei-1999/AutoMark>
- Source: **RoboInter-Data** (`InternRobotics/RoboInter-Data`), itself built on
  [DROID](https://droid-dataset.github.io/) and [RH20T](https://rh20t.github.io/) — see each for
  their original licenses.
- **License: CC BY-NC-SA 4.0** (inherited from RoboInter). Non-commercial use only; derivatives
  must carry the same license; attribute RoboInter, DROID, and RH20T.

If you use these annotations, please cite RoboInter-Data and the upstream DROID / RH20T datasets.