facebook/wearable-ai

Wearable AI Dataset (ECCV 2026)

Part of the Wearable AI Workshop at ECCV 2026.

A benchmark of egocentric (first-person, head-mounted wearable camera) videos paired with three complementary video question-answering tasks for evaluating wearable-AI assistants on real-world everyday activity videos.

▶ Baseline code & evaluation scripts: see starter_kit/README.md. The starter kit ships inside this repo, so git clone gives you the code and the data together.

Tasks

EgoLongQA — Long-form video question answering with multiple-choice answers. Given a video and a question with four options (A/B/C/D), predict the correct answer.

EgoConv — Conversational video question answering. Given a video and a multi-turn conversation, generate free-form answers for each turn. The model sees only the video up to the current turn (no future leaking) and uses its own previous answers as context.

EgoProactive — Proactive assistant over streaming egocentric video. Given a video stream, the model must decide when to speak and what to say at each candidate moment, simulating a wearable AI that volunteers helpful information without being prompted.

Configurations

Config	Split	# samples	Description
egoconv	val	700	Multi-turn conversational QA grounded in egocentric video
egolongqa	val	700	Long-video MCQ over single, longer egocentric clips
egoproactive	val	700	Proactive-assistant moments over streaming egocentric video

Repo Layout

facebook/wearable-ai/
├── README.md                   # this file
├── LICENSE                     # CC-BY-NC-4.0 (dataset)
├── egoconv/
│   ├── wearable_ai_2026_egoconv_val_700.jsonl
│   └── val/<id>.mp4            # 700 videos, ~91 GB
├── egolongqa/
│   ├── wearable_ai_2026_egolongqa_val_700.jsonl
│   └── val/<id>.mp4            # 700 videos, ~203 GB
├── egoproactive/
│   ├── wearable_ai_2026_egoproactive_val_700.jsonl
│   └── val/<id>.mp4            # 700 videos, ~23 GB
└── starter_kit/                # baseline code + evaluation scripts
    ├── README.md
    ├── LICENSE                 # CC-BY-NC-4.0 (code)
    ├── run_evaluation.py
    └── ...

Cloning & Disk Usage

The full dataset is ~317 GB across 2,100 videos. Pick the recipe that matches what you need:

Full clone (code + annotations + all videos, ~317 GB):

git clone https://huggingface.co/datasets/facebook/wearable-ai

Code + annotations only (~25 MB; videos left as LFS pointers, downloadable on demand):

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/facebook/wearable-ai

One task only (annotations + the task's videos):

huggingface-cli download facebook/wearable-ai \
    --repo-type dataset \
    --include "egolongqa/**" "starter_kit/**" "*.md" "LICENSE" \
    --local-dir wearable-ai

Per-task video sizes: egoconv ≈ 91 GB · egolongqa ≈ 203 GB · egoproactive ≈ 23 GB.

Loading

from datasets import load_dataset

egoconv_val = load_dataset("facebook/wearable-ai", "egoconv", split="val")
egolongqa_val = load_dataset("facebook/wearable-ai", "egolongqa", split="val")
egoproactive_val = load_dataset("facebook/wearable-ai", "egoproactive", split="val")

Schemas

egoconv — Conversational QA over egocentric video

Field	Type	Description
video_path	str	Basename of the video file (<id>.mp4)
duration_in_sec	float	Video duration in seconds
video_intervals	list[[float, float]]	Time intervals (start, end) covered by the conversation
questions	list[str]	Sequence of user questions over the video
answers	list[str]	Reference answers, aligned to questions
task	str	Coarse activity tag (e.g. Tourism, Cooking)
dialog	list[dict]	Raw turn-by-turn dialog. Each turn: `{text: str, role: "P0"

egolongqa — Long-form MCQ over egocentric video

Field	Type	Description
video_path	str	Basename of the video file (<id>.mp4)
question	str	Long-form question over the full video
answer	str	Open-ended reference answer
mcq_options	str	Formatted MCQ choices: A. ... B. ... C. ... D. ...
mcq_answer	str	Correct option letter (A/B/C/D)
category	str	Activity category (e.g. Travel-Sightseeing (Indoors))

egoproactive — Proactive assistant over streaming egocentric video

Field	Type	Description
video_path	str	Basename of the video file (<id>.mp4)
duration_in_sec	float	Video duration in seconds
video_intervals	list[[float, float]]	Candidate decision intervals (start, end) over the streaming video
query	str	The user's initial high-level query (e.g. "How do I decorate a notebook cover with stickers?")
domain	str	Coarse domain tag (e.g. Arts and Crafts, Cooking)
task	str	Specific task description (e.g. "Decorating a notebook cover with stickers")
answers	list[str]	Reference assistant decisions, one per interval. Each entry is either "$silent$" (stay silent) or "$interrupt$<utterance>" (speak with the given utterance)
dialog	list[list[dict]]	Cumulative dialog state before each interval. Each turn: `{role: "user"

Videos

Video files are bundled in this repository under each config's val/ folder. The video_path field in each row is the basename of a .mp4 file located at:

<repo_root>/<config>/val/<video_path>

where <config> is the config name (egoconv, egolongqa, or egoproactive). The repo holds 2,100 videos (700 per task, ~317 GB total). Videos are H.265 / 1080p / 15 fps, audio-stripped, and face-blurred. See Cloning & Disk Usage above for download recipes.

Example: looking up a video locally after downloading the repo snapshot:

import os
from huggingface_hub import snapshot_download

repo_root = snapshot_download("facebook/wearable-ai", repo_type="dataset")
sample = egoconv_val[0]
video_file = os.path.join(repo_root, "egoconv", "val", sample["video_path"])

License

The dataset in this repository (videos + JSONL annotations) is released under CC-BY-NC-4.0 — see LICENSE.

The starter-kit code under starter_kit/ is separately licensed under CC-BY-NC-4.0 — see starter_kit/LICENSE.

Both licenses permit academic / research use only; commercial use is not granted. Pre-trained model weights downloaded from HuggingFace (e.g., Llama 4 Scout, Llama 4 Maverick, Qwen2.5-VL) are governed by their respective licenses on the model pages.

Citation

For datasets in general:

@misc{wearableaiworkshop2026,
  title = {Wearable AI Workshop at ECCV 2026},
  author = {Tuyen (Harry) Tran and Maxim Arap and Seungwhan Moon and Raffay Hamid and Alessandro Suglia and Zsolt Kira and Pascale Fung and Mubarak Shah},
  year = {2026},
  howpublished = {\url{https://wearable-ai-workshop.github.io/}},
  note = {Workshop at the European Conference on Computer Vision (ECCV) 2026}
}

For EgoProactive dataset specifically:

@misc{kundu2026planwatchrecoverbenchmark,
      title={Plan, Watch, Recover: A Benchmark and Architectures for Proactive Procedural Assistance}, 
      author={Kaustav Kundu and Ritvik Shrivastava and Maxim Arap and Nanshu Wang and Xianhui Zhu and Quintin Fettes and Gautam Tiwari and Parth Suresh and Théo Moutakanni and Alejandro Castillejo Munoz and Allen Bolourchi and Pascale Fung and Pinar Donmez and Babak Damavandi and Anuj Kumar and Seungwhan Moon},
      year={2026},
      eprint={2606.04970},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.04970}, 
}