manus-egocentric-sample

Egocentric video dataset with Manus glove hand tracking data, converted to LeRobot v3.0 format.

Dataset Description

This dataset contains egocentric (first-person view) recordings of human hands performing various manipulation tasks, captured with:

• Manus Metagloves: High-precision finger tracking (~70Hz)
• OAK-D Camera: RGB video (1920x1080, 30fps) + Depth (640x400, 30fps)
• IMU: Accelerometer and gyroscope data

Dataset Statistics

Property	Value
Total Episodes	20
Total Frames	96,544
FPS	30
Robot Type	manus_gloves
LeRobot Version	v2.1
Dataset Size	~15.3 GB

Tasks

Pick_and_pack_task, fold_laundry

Features

Feature	Type	Shape
episode_index	int64	[1]
frame_index	int64	[1]
timestamp	float32	[1]
task_index	int64	[1]
observation.images.egocentric	video	[1080, 1920, 3]
observation.state.hand_joints	float32	[150]
observation.state.finger_angles	float32	[40]
observation.state.gestures	float32	[26]
observation.state.imu	float32	[6]

Depth Data

This dataset includes raw depth data as a custom extension (LeRobot v3.0 doesn't officially support depth yet).

Property	Value
Format	Raw uint16 binary
Resolution	[400, 640] (H, W)
Unit	millimeters
Episodes with depth	17
Storage	Episode-based (depth/episode_XXXXXX.bin)

To load depth data:

import numpy as np

def load_depth(dataset_path, episode_index):
    depth_path = dataset_path / f"depth/episode_{episode_index:06d}.bin"
    if depth_path.exists():
        data = np.fromfile(depth_path, dtype=np.uint16)
        # Reshape based on frame count (need timestamps)
        return data
    return None

Usage

Load with LeRobot

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("opengraph-labs/manus-egocentric-sample")

# Access a sample
sample = dataset[0]
print(sample.keys())
# dict_keys(['observation.images.egocentric', 'observation.state.hand_joints', ...])

# Get hand joint positions (25 joints x 3 coords x 2 hands = 150 dims)
hand_joints = sample["observation.state.hand_joints"]
print(hand_joints.shape)  # torch.Size([150])

Data Structure

manus-egocentric-sample/
├── meta/
│   ├── info.json           # Dataset schema and features
│   ├── stats.json          # Feature statistics (mean/std/min/max)
│   ├── tasks.jsonl         # Task definitions
│   ├── episodes.jsonl      # Episode metadata
│   └── episodes/           # Episode info (parquet)
├── data/chunk-XXX/         # Time-series data (parquet)
├── videos/                 # RGB video (mp4)
└── depth/                  # Raw depth data (uint16 binary)

Hand Skeleton Structure

Each hand has 25 tracked joints:

Hand (Root)
├── Thumb: MCP → PIP → DIP → TIP (4 joints)
├── Index: MCP → IP → PIP → DIP → TIP (5 joints)
├── Middle: MCP → IP → PIP → DIP → TIP (5 joints)
├── Ring: MCP → IP → PIP → DIP → TIP (5 joints)
└── Pinky: MCP → IP → PIP → DIP → TIP (5 joints)

Citation

If you use this dataset, please cite:

@dataset{manus_egocentric_2025,
  title = {Manus Egocentric Hand Tracking Dataset},
  author = {OpenGraph Labs},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/opengraph-labs/manus-egocentric-sample}
}

License

This dataset is released under the Apache 2.0 License.