Dataset Viewer
codebase_version
string | robot_type
string | total_episodes
int64 | total_frames
int64 | total_tasks
int64 | fps
int64 | splits
dict | data_path
string | features
dict |
|---|---|---|---|---|---|---|---|---|
v2.0
|
egocentric_human
| 50
| 1,800
| 1
| 6
|
{
"train": "0:50"
}
|
data/episode_{episode_index:06d}.npz
|
{
"observation.images.top": {
"dtype": "image",
"shape": [
360,
640,
3
],
"names": [
"height",
"width",
"channel"
]
},
"observation.state": {
"dtype": "float32",
"shape": [
4
],
"names": [
"hand_bbox_x_min",
"hand_bbox_y_min",
"hand_bbox_x_max",
"hand_bbox_y_max"
]
},
"action": {
"dtype": "float32",
"shape": [
2
],
"names": [
"hand_delta_x",
"hand_delta_y"
]
},
"episode_index": {
"dtype": "int64",
"shape": [
1
]
},
"frame_index": {
"dtype": "int64",
"shape": [
1
]
},
"timestamp": {
"dtype": "float32",
"shape": [
1
]
},
"next.done": {
"dtype": "bool",
"shape": [
1
]
},
"index": {
"dtype": "int64",
"shape": [
1
]
}
}
|
Ego2Robot: Factory Manipulation Episodes
Dataset Description
50 curated episodes of factory worker manipulation tasks, converted from egocentric video into LeRobot-compatible format for robot learning research.
Key Features
- 50 episodes (~1,800 frames total)
- Real factory work from 85 manufacturing facilities
- 10 skill clusters discovered via unsupervised learning
- LeRobot v3.0 format with observations + pseudo-actions
- Rich annotations: VideoMAE embeddings, CLIP labels, quality scores
Data Structure
Each episode contains:
- Observations:
observation.images.top: RGB frames (360x640, 6fps)observation.state: Hand bounding box [x_min, y_min, x_max, y_max]
- Actions: 2D hand motion vectors
[delta_x, delta_y](pseudo-actions for representation learning) - Metadata: Skill cluster ID, zero-shot action label, quality scores
Skill Distribution
- Quality Inspection: 50% (25 episodes)
- Assembly: 17% (9 episodes)
- Fastening: 17% (8 episodes)
- Machine Operation: 8% (4 episodes)
- Mixed: 8% (4 episodes)
Intended Use
Primary: Representation learning and pretraining for vision-language-action (VLA) models
- Pretrain visual encoders on diverse manipulation tasks
- Learn spatial reasoning from egocentric perspective
- Discover manipulation primitives via clustering
- Domain adaptation for manufacturing robotics
NOT intended for: Direct robot policy learning (actions are pseudo-actions from human hand motion, not robot joint commands)
Data Collection
- Source: BuildAI Egocentric-10K dataset
- Processing pipeline:
- Quality filtering (motion + hand visibility)
- VideoMAE embeddings (768-dim)
- CLIP zero-shot labeling
- K-means clustering (10 skills)
- Hand tracking (MediaPipe)
- LeRobot format conversion
Citation
@dataset{ego2robot2025,
title={Ego2Robot: Factory Manipulation Episodes for Robot Learning},
author={Michelle Sun},
year={2025},
url={https://huggingface.co/datasets/msunbot1/ego2robot-factory-episodes}
}
License
Apache 2.0 (inherits from Egocentric-10K source dataset)
Contact
For questions or collaboration: x.com/michellelsun
- Downloads last month
- 40
Size of downloaded dataset files:
1.28 kB
Size of the auto-converted Parquet files:
15.4 kB
Number of rows:
1