| --- |
| license: cc-by-sa-4.0 |
| task_categories: |
| - robotics |
| language: |
| - en |
| tags: |
| - embodied-ai |
| - robotics |
| - egocentric |
| - hand-tracking |
| - trajectory |
| - slam |
| - vla |
| - manipulation |
| size_categories: |
| - n>1T |
|
|
| gated: true |
| extra_gated_heading: "Access Gen-HumanEgo" |
| extra_gated_description: "Please provide a few details about yourself and your intended use of Gen-HumanEgo. Access is automatically granted after you submit the form." |
| extra_gated_button_content: "Agree and access dataset" |
|
|
| extra_gated_fields: |
| Institution / Organization: |
| type: text |
| Intended use: |
| type: select |
| options: |
| - Research |
| - Education |
| - Industry / Product Development |
| - Other |
| Brief description of intended use: |
| type: text |
| --- |
| |
| # Gen-HumanEgo |
|
|
| <p align="center"> |
| <img src="assets/banner.jpg" width="100%"> |
| </p> |
|
|
| **1,800+ hours of egocentric human demonstrations with synchronized, structured supervision for embodied AI and robot learning.** |
|
|
| Gen-HumanEgo contains real-world first-person demonstrations collected across diverse tasks, people, environments, and ways of performing activities using a unified six-camera [DAS-Ego](https://www.genrobot.ai/products/ego) setup. GenRobot's **Data Foundation Model (DFM)** processes the recordings to provide complementary training signals for human motion, spatial understanding, and task understanding. |
|
|
| The dataset includes: |
|
|
| - **Egocentric RGB:** synchronized multi-view video recordings. |
| - **3D hand reconstruction:** 3D hand keypoints, MANO parameters, and full hand geometry. |
| - **Spatial understanding:** large-field-of-view Ego-Depth. |
| - **Semantic understanding:** hierarchical video-, task-, and subtask-level annotations with temporal boundaries. |
|
|
| For a detailed introduction to the dataset, see our [Gen-HumanEgo Technical Blog](https://www.genrobot.ai/blog/gen-human-ego) |
|
|
|
|
| ## Dataset at a glance |
|
|
| | Attribute | Value | |
| |---|---:| |
| | Total duration | **1,847.7 hours** | |
| | Episodes | **44,632** | |
| | Unique tasks | **10,257** | |
| | Domains | Home, business, industry, agriculture | |
| | Native RGB views | **6 synchronized cameras** | |
| | Resolution | **1600 × 1300** | |
| | Frame rate | **30 FPS** | |
| | File foramt| MCAP | |
| | Modalities | **21 3D keypoints per hand, full handmesh, depth maps, task+subtask annotation** | |
|
|
|
|
| ## Repository structure |
|
|
| The task-organized directory tree follows the dataset's scenario/skill hierarchy. The final two levels are a two-character file-prefix directory and the episode's MCAP filename. |
|
|
| ```text |
| Gen-HumanEgo/ |
| ├── README.md |
| ├── egov4_urdf.zip |
| ├── assets/ |
| │ └── ... |
| └── data/ |
| ├── industry/ |
| │ └── logistics/ |
| │ └── sorting_and_packing/ |
| │ └── daily_work/ |
| │ └── 00/ |
| │ └── 005c469cf7424c4c87054569b2234ebc.mcap |
| ├── domestic_services/ |
| │ └── living_room/ |
| │ └── clothing_organization/ |
| │ └── iron_clothes/ |
| │ └── c0/ |
| │ └── c096609eca54434e9fa93b0fff55ce3d.mcap |
| └── business/ |
| └── restaurant/ |
| └── bakery/ |
| └── daily_work/ |
| ├── a9/ |
| │ └── a91e57586c33437292d763fb8fd3777c.mcap |
| └── 10/ |
| └── 101207375b8e41fea31bacd94db480e6.mcap |
| ``` |
|
|
|
|
| ## What's inside an episode? |
|
|
| Each episode is packaged as an MCAP recording. The release includes synchronized video and structured outputs, summarized below. |
|
|
| | Data | What it provides | Exact topic/schema | |
| | --- | --- | --- | |
| | Multi-view RGB | First-person visual observations | /robot0/sensor/camera[0-6]/compressed | |
| | Hand reconstruction | 3D keypoints, MANO parameters, hand geometry and associated validity/quality information where provided | /robot0/handtracking/left, /robot0/handtracking/right | |
| | Ego-Depth | Large-FOV depth for the surrounding workspace | /robot0/sensor/camera2/depth | |
| | Hierarchical annotations | Episode description, task segments, and fine-grained subtasks | /robot0/annotation_v2/ | |
| |
| The hierarchical annotation design describes activity at three levels: |
| |
| 1. **Video:** an overall description of the episode. |
| 2. **Task:** the objective of a continuous segment, with time/frame boundaries and scene information. |
| 3. **Subtask:** a more detailed action caption, time/frame boundaries, and a success field (`is_success`), with objects, observable attributes, and spatial relationships expressed in the caption when applicable. |
|
|
| For example, an episode about tidying a sofa can include the task *“Tidy up the sofa”* and subtasks such as *“Place the red and black throw pillow on the sofa back”* and *“Flatten the sofa cover on the right seat of the sofa.”* |
|
|
|
|
|
|
| ## Documentation and tools |
|
|
| - **GenRobot DAS-Ego Data Introduction** — topic definitions, message schemas, coordinate systems, timestamp conventions, and transformation definitions. [GenRobot DAS-Ego Data Introduction](https://docs.genrobot.ai/guides/das-ego-data-introduction) |
| - **GenRobot MCAP Visualization Tool** — browser-based visualization of MCAP episodes. [GenRobot MCAP Visualization Tool](https://monitor.genrobot.click/#/index) |
| - **GenRobot DAS DataKit** — tools for reading, validating, visualizing, and converting DAS-Ego data. [GenRobot DAS DataKit](https://github.com/genrobot-ai/das-datakit) |
| - **Ego URDF** — [`egov4_urdf.zip`](egov4_urdf.zip), containing the DAS-Ego device link structure and sensor-frame definitions. |
|
|
|
|
| ## Contact |
|
|
| Questions, suggestions, and requests for additional scenarios or skills are welcome. |
|
|
| - X: [@GenrobotAI](https://x.com/GenrobotAI) |
| - LinkedIn: [GenRobot](https://www.linkedin.com/company/108767412/) |
| - Email: opendata@genrobot.ai |
| - Discord: https://discord.gg/rSSb5thgu |
|
|
