Request access to Robot-EQ-Video

Access is granted only after manual review. Incomplete applications will be rejected. You MUST provide (1) an official school/university email with an institutional domain suffix, (2) your advisor's full name, and (3) your advisor's Google Scholar page. Personal emails (Gmail, QQ, 163, Outlook, Hotmail, iCloud, etc.) will be rejected. 申请必须填写学校后缀邮箱、导师姓名、导师谷歌学术主页;个人邮箱一律拒绝。

You request access to Robot-EQ-Video (synthetic 1218 + real 928).
Use is for academic research only. Do not redistribute the videos.
Real videos may contain identifiable people / locations; do not attempt to identify individuals.

Required for approval:

  • Institutional email with a school/university domain suffix (e.g. .edu / .edu.cn / .ac.cn / your university domain). Personal emails such as Gmail, QQ, 163, Outlook, Hotmail, and iCloud will be rejected.
  • Advisor / supervisor full name.
  • Advisor's Google Scholar profile URL (scholar.google.com). Requests without a valid Scholar page will be rejected.

审核要求:必须使用学校/研究所域名邮箱;必须填写导师姓名和导师的 Google Scholar 主页。

Log in or Sign Up to review the conditions and access this dataset content.

Robot-EQ-Video

Video companion to Tongji-Emotion/Robot-EQ.

This repo contains raw mp4 videos and action-level labels for Robot-EQ behavior judgment (synthetic + real). Image-side annotations remain in Tongji-Emotion/Robot-EQ.

Each action is a candidate robot behavior (text) with a majority-vote label. Annotations are released in Chinese and English (same items; only the label wording differs). Raw per-annotator files are not included.

该做 = proper, 不该做 = improper. Action text is the original Chinese wording used in evaluation.

Statistics

Videos Actions Proper Improper
Total 2146 17544 10469 7075
Synthetic 1218 11225 6742 4483
Real 928 6319 3727 2592

Every video has at least 4 labeled actions.

Layout

Robot-EQ-Video/
├── README.md
├── metadata.jsonl
├── annotations/
│   ├── Chinese/
│   │   ├── actions.jsonl
│   │   ├── gt_synthetic.json
│   │   └── gt_real.json
│   └── English/
│       ├── actions.jsonl
│       ├── gt_synthetic.json
│       └── gt_real.json
├── synthetic/videos/*.mp4
└── real/videos/*.mp4

Filenames are GT video_ids. Real videos use unique prefixed ids because original basenames collide.

metadata.jsonl

One video per line:

  • video_id
  • domain: synthetic | real
  • file_name: relative path inside this repo
  • n_actions / n_proper / n_improper

annotations

Chinese (annotations/Chinese/) and English (annotations/English/) are the same 17544 actions.

Chinese actions.jsonl

  • action_id: {video_id}/{opt}
  • video_id, opt, file_name, domain
  • text: action description
  • label: 该做 | 不该做
  • agree / valid_total: annotator agreement on this item

English actions.jsonl

  • action_id: {video_id}/{opt}
  • video_id, opt, file_name, domain
  • text: action description
  • label: proper | improper
  • agree / valid_total: annotator agreement on this item

gt_synthetic.json and gt_real.json in each folder are the same labels in dict form (key = video_id/opt).

Load

from datasets import load_dataset

videos = load_dataset("Tongji-Emotion/Robot-EQ-Video", data_files="metadata.jsonl", split="train")
zh = load_dataset("Tongji-Emotion/Robot-EQ-Video", data_files="annotations/Chinese/actions.jsonl", split="train")
en = load_dataset("Tongji-Emotion/Robot-EQ-Video", data_files="annotations/English/actions.jsonl", split="train")

Videos are under synthetic/videos/ and real/videos/ as listed by file_name.

Downloads last month
99