| --- |
| license: apache-2.0 |
| task_categories: |
| - robotics |
| tags: |
| - lerobot |
| - ego-vision |
| - embodied-ai |
| - teleoperation |
| - robot-learning |
| - factory-automation |
| - 3c-manufacturing |
| - manipulation |
| pretty_name: RealFactory Ego Data |
| --- |
| |
| # RealFactory-Ego-Data |
|
|
| Real-world factory production line ego-perspective manipulation dataset collected by **FlyBirdData (飞鸟数据)**, a data service brand of **Shenzhen Mingzhi DataTech Co., Ltd. (深圳市明志数科信息有限公司)**. |
|
|
| ## About |
|
|
| This dataset contains **5 high-quality episodes** of real 3C electronics manufacturing tasks recorded from an ego-centric (first-person) perspective on an actual touchscreen production line. Each episode includes synchronized RGB video, hand pose landmarks, and holistic action annotations in LeRobot-compatible format. |
|
|
| ## Tasks |
|
|
| | Episode | Task | Duration | Size | |
| |---------|------|----------|------| |
| | 触摸屏贴边2026081304 | Edge alignment & bonding | ~3 min | 203 MB | |
| | 触摸屏贴边20260818012 | Edge alignment & bonding | ~3 min | 207 MB | |
| | 触摸屏割膜20260819007 | Protective film cutting | ~3 min | 271 MB | |
| | 触摸屏抓取放置20260819001 | Pick & place on conveyor | ~2 min | 102 MB | |
| | 触摸屏撕膜20260818018 | Protective film peeling | ~3 min | 223 MB | |
|
|
| ## Data Format |
|
|
| Each `.zip` file contains a **LeRobot-compatible holistic output** with: |
|
|
| - **RGB video** — Ego-perspective recording of the assembly task |
| - **Hand pose** — 21-landmark hand keypoint tracking per frame |
| - **Action annotations** — Frame-level action boundary labels (holistic_detect output) |
| - **Metadata** — Task type, timestamps, and quality metrics |
| |
| ## Collection Environment |
| |
| - **Scenario**: Real 3C touchscreen manufacturing production line |
| - **Viewpoint**: Ego / first-person (head-mounted camera) |
| - **Tasks**: Edge bonding, film cutting, pick-and-place, film peeling |
| - **Worker**: Experienced production line operator |
| - **Annotation pipeline**: FlyBirdData automated holistic detection + manual QC |
| |
| ## Why Ego Data Matters |
| |
| Ego-perspective data captures the exact visual input a human worker sees during assembly, including occlusions, depth cues, and fine motor control. This is critical for training imitation learning and video-based manipulation policies that must generalize to real factory conditions — not just lab environments. |
| |
| ## About FlyBirdData |
| |
| **FlyBirdData (飞鸟数据)** is a specialized robot training data provider under Shenzhen Mingzhi DataTech Co., Ltd. We are one of the few providers in China with real 3C manufacturing production line access for ego-perspective data collection. Our services include: |
| |
| - Ego + UMI (Universal Manipulation Interface) fused data collection |
| - Force/torque, joint torque, and 3D pose annotation |
| - LeRobot standard format conversion |
| - Full pipeline: collection → annotation → QC → delivery |
| |
| **Website**: https://mz-datatech.com/ |
| **Contact**: contact@mz-datatech.com |
| **Code**: https://gitee.com/ok_vincey/flybird-data |
|
|
| ## License |
|
|
| Apache-2.0 |
