SmartDeer/egocentric-manipulation-sample

SmartDeer — Egocentric Human Demonstration Data (Public Sample)

SmartDeer collects egocentric human demonstration data for embodied AI: real operators, real tasks, real environments. No teleoperation rigs, no lab mock-ups, no actors.

This repository is a small public sample, published so you can inspect our capture quality, annotation schema and file conventions. Production data is delivered under commercial terms.

This sample

Task twist_cap — bimanual bottle-cap opening across three bottle geometries. 3 episodes, one stereo pair each (6 MP4 files), 30 fps, 48 s of task time, audio removed, single operator in a single indoor scene.

Episode	Content	Duration
ep0001	Purple water bottle, large cap	17 s
ep0002	Short bottle, red cap — one hand stabilises, one hand twists	12 s
ep0003	Tall bottle, yellow label — one hand stabilises, one hand twists	19 s

Fields: video, episode_id, camera_id (left / right), task, caption_en, caption_zh, duration_seconds, fps

from datasets import load_dataset

ds = load_dataset("SmartDeer/egocentric-manipulation-sample", split="train")
print(ds[0])

Collection network

• 800+ active collectors
• Coverage: Mainland China, Hong Kong SAR, Cambodia, Vietnam, Indonesia and others
• Scene categories: household, factory and industrial, food service, agriculture and others
• Review before delivery: task representativeness, action boundaries, caption accuracy, and a full-frame privacy check on both views

Because demonstrations come from people performing tasks in the environments where those tasks actually happen, object variety, lighting, clutter and hand strategy carry the natural long tail that scripted lab capture removes.

Privacy and consent

Collectors participate under written agreements covering data use and likeness. Filenames are anonymised, audio is stripped, and every released frame has passed human review for faces, identity documents, on-screen text and reflections.

CC BY 4.0 governs the dataset. It does not by itself grant third-party likeness rights.

Custom collection

We run task-specified collection programmes for foundation-model teams, robotics companies and capture-hardware partners. Send us a task list, a target episode count, sensor requirements and a delivery format, and we will come back with a capture plan.

Contact: paul@ideal-careerbridge.com

---

SmartDeer — 第一人称人类演示数据（公开样例）

SmartDeer 采集面向具身智能的第一人称人类演示数据：真实操作者、真实任务、真实环境。不使用遥操作设备，不搭实验室场景，不用演员。

本仓库是一个小规模公开样例，用于查看我们的采集质量、标注结构和文件规范。生产级数据以商业方式交付。

样例内容

任务 twist_cap —— 双手旋开瓶盖，覆盖三种瓶型。共 3 段，每段一组双目视频（6 个 MP4），30 fps，48 秒任务时间，已去除音频，来自同一操作者、同一室内场景。

片段	内容	时长
ep0001	紫色水瓶，大瓶盖	17 秒
ep0002	红盖短瓶 —— 一手固定瓶身，一手旋盖	12 秒
ep0003	黄标长瓶 —— 一手固定瓶身，一手旋盖	19 秒

字段：video、episode_id、camera_id（left / right）、task、caption_en、caption_zh、duration_seconds、fps

采集网络

• 800+ 名活跃采集员
• 覆盖地区：中国大陆、中国香港、柬埔寨、越南、印尼等
• 场景大类：家居环境、工厂环境、餐饮、农业等
• 交付前复核：任务代表性、动作边界、描述准确性，以及双侧画面的全帧隐私检查

演示数据来自人们在任务真实发生的环境中完成任务，因此物体多样性、光照、杂乱程度和用手策略保留了脚本化实验室采集会抹掉的长尾分布。

隐私与授权

采集员在书面协议下参与，协议覆盖数据使用与肖像授权。文件名已匿名化，音频已去除，公开的每一帧都经过人工复核，检查人脸、身份证件、画面文字与反光。

CC BY 4.0 适用于数据集本身，不构成对第三方肖像权的授权。

定制采集

我们为基座模型团队、机器人公司和采集硬件伙伴提供按任务定制的采集服务。请提供任务清单、目标片段数量、传感器要求和交付格式，我们会给出采集方案。

联系：paul@ideal-careerbridge.com

Citation

@misc{smartdeer_hf_demo,
  title  = {SmartDeer Egocentric Human Demonstration Data (Public Sample)},
  author = {SmartDeer},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SmartDeer/SmartDeer/egocentric-manipulation-sample}
}