Today we are excited to announce the release of our most ambitious hand motion dataset to date: Dexterous Hand Movements Pro.
With 139 carefully selected videos and 123,732 high-quality exported frames, this dataset represents a significant step forward in providing clean, rich, and ready-to-use data for dexterous manipulation research and humanoid robotics development.
Why This Dataset Matters
Modern humanoid robots need more than basic pose estimation. They need precise understanding of finger movements, grip transitions, wrist dynamics, and temporal consistency. This dataset delivers exactly that.
Key Highlights:
Mean Quality Score: 0.9747
Mean Landmark Visibility: 0.9495
Acceptance Rate: 57.13% (strict HQ filtering)
Full 21-point MediaPipe Hands landmarks
Wrist-relative coordinates for easier robot retargeting
Rich dexterous_hand features: finger flexion angles, grip type proxies (Pinch, Precision, Power, Open, Neutral), tip velocities, and wrist speed
Motion Intelligence v3 (hand domain) with activity classification and grip change detection
Gaussian temporal smoothing + wrist-pivot augmentations
Who Is This Dataset For?
This pack is specifically designed for teams working on:
Dexterous in-hand manipulation
Humanoid robot hand control (G1, GR-1, Walker, AgiBot, etc.)
Imitation learning and teleoperation
Fine-grained gesture recognition
AR/VR hand tracking systems
What’s Inside
Clean JSONL format with per-frame records
Per-video splits for easy experimentation
Full metadata, quality reports, and schema documentation
Interactive viewer to inspect keypoints before training
Transparent rejection logs and global statistics
Whether you are training a new policy from scratch or fine-tuning an existing model, this dataset provides the high-fidelity temporal data needed to push manipulation performance further.
Ready to explore?
Check the full dataset on Gumroad or contact us for enterprise licensing and custom weekly supplies (up to 140k+ frames per month).
We are also open to extracting custom hand movements from your own videos within 24 hours.