Video · Multimodal / AI & Machine Learning
Egocentric Hand Manipulation Dataset | Cloth-Folding Episodes with Wrist-Camera Video & IMU | LeRobot v3.0, Global Use
Advertised by providerLicensing rights not independently verifiedPrice on request
Overview
Egocentric Hand Manipulation Dataset | Cloth-Folding Episodes with Wrist-Camera Video & IMU | LeRobot v3.0, Global Use is a multimodal dataset. Typical applications include LLM pre-training, alignment & RLHF and robotics & embodied AI. Relevant sectors: AI & Machine Learning, Manufacturing & Industrial and Technology & IT. Pricing is not published and is provided on request.
Primary applications
This dataset contains ~20-second episodes of a person naturally folding cloth, captured by the provider’s wrist-worn AI smartwatch. No teleoperation rig, no head-mounted camera, no lab setup — the participant simply wears the watch and does the task, so the viewpoint stays where the hands are and grasps, contacts and object states remain in frame throughout. MEASURED EFFECT ON POLICY PERFORMANCE We ran a controlled test on DexGarmentLab (NeurIPS 2025, Isaac Sim). A Diffusion Policy was trained on a small set of robot demonstrations, and the only variable was whether its vision encoder had been pretrained on this data. Task success doubled — 5.8% to 12.5% — reproduced across four pretraining variants over 840 evaluation rollouts, every arm scored on the same unseen garments and placements. The full write-up is available on request. WHAT EACH EPISODE CONTAINS - Wrist-camera video: 640x480, 30 fps, egocentric palm-facing view with hands and workspace in frame - observation.state: frame-aligned 6-axis IMU in SI units (gyroscope rad/s, accelerometer m/s2) - Raw IMU stream: one CSV per episode at 104 Hz, on a shared device uptime clock with per-episode video alignment metadata — the lossless source of truth - Derived motion signals: one JSON per episode with a normalized per-frame motion-intensity curve and high-activity intervals, computed by optical-flow residual analysis. These are computed signals, not labels: the dataset ships without hand-annotated action labels, and we say so
Related datasets
More in Video & Physical AI →Video · MultimodalAdvertised
Custom Manipulation Data Collection | Egocentric Wrist-Camera Video + IMU Recorded to Your Task Spec | LeRobot v3.0
Off-the-shelf manipulation datasets rarely match the exact task, objects, or environment your policy needs.
- Language
- —
- Volume
- —
- Domain
- AI & Machine Learning +3
- Use case
- LLM pre-training
Price on request
More video & physical ai datasets
All 302 video & physical ai datasets →- Egocentric head-mounted video of everyday household tasks
- Egocentric Hotel & Lodging Housekeeping Video Data: Cleaning, Turnover & Laundry | SLAM/VIO Head-Motion Tracking | Embodied AI, Physical AI
- Egocentric Household Activity Demonstrations — LeRobot
- Egocentric human, synthetic, and teleoperated robot video
- Egocentric Multi-View Dataset with 21-points 3D hand Keypoints Annotations
- Egocentric Personal-Care & Salon Video: Haircutting, Manicure & Pedicure | Machine Learning (ML) Data, Synced IMU | Fine-Motor Robotics
- Egocentric POV video of commercial kitchen service workflows
- Egocentric Restaurant & Commercial-Kitchen Video: Cooking, Dishwashing & Prep | Dual Global-Shutter Stereo Camera | Physical AI, Robotics Data