High-precision multimodal motion data for humanoid robots, embodied AI, and virtual production.
Motion capture studio panorama → Actor capture → Full-body motion capture → Rigid body tracking of objects → Real-time display of bones → Robot redirection → Multi-scene montage
ChingMu 1000-Hour Embodied Motion Dataset It is a high-precision, multi-modal, and reproducible action data infrastructure specifically designed for humanoid robots, dexterous hands, embodied AI, and virtual production.
The data is collected through an optical motion capture system, covering full-body human movements, 6D poses of objects, multi-angle videos, task labels and quality inspection reports; the scenarios cover typical applications such as industrial manufacturing, household services, supermarket retail, medical care and rehabilitation, logistics and warehousing, ball interaction and entertainment performances, etc.
By implementing standardized data collection, automated cleaning of annotations, robot redirection, and quality assessment processes, a high-quality data foundation is provided for imitation learning, action generation, simulation training, and real robot deployment.
1000 hours of optical motion capture data, with sub-millimeter accuracy, capturing full body movements + fine finger actions + 6DoF synchronization of objects.
During the 1000h → 2000h expansion, it covers over 15 real scenarios, aligns multi-angle videos with mocap frame clocks, and comes with structured task labels.
3000h+ target direction, collecting natural human behaviors in uncontrolled environments, enhancing the model's generalization ability in the real physical world.
Walking, running, jumping, bending over, turning around, carrying objects, etc. - these are all full-body movements suitable for the training of the overall control of humanoid robots.
Data size: 200+ hours · 100+ types of actions
The human hands, props and goods are simultaneously collected in the same coordinate system, and the 6DoF pose of each frame is output.
Data size: 150+ hours · 100+ rigid bodies
Human→Robot skeleton retargeting + Simulation verification + Quality report
Data volume: Over 300 hours · Various robot models
Multi-camera images aligned with the mocap frame clock, used for input and verification of the visual model.
Data size: 400+ hours · 8 camera positions synchronized
Multiple data redirection effects, visually presenting the diversity of data and its migration capability.
Dance
Upstairs
Punch
Parkour
Lifting and moving
Basketball
Table tennis
Two-person confrontation
Walking backward
The dataset is hosted on
Hugging Face 🤗
，
It can be loaded directly from datasets the library.
View the complete README, file structure, License, sample loading script and release notes on Hugging Face.
🤗 Open in Hugging FaceBased on the 1000h optical motion capture technology, we are currently expanding the data dimension further.
Based on the Layer 1 full-body + finger model, further expand the manipulation, twisting, button-pressing, tool operation and other millimeter-level finger joint trajectories, and output them in the form suitable for dexterous hand IL/RL training.
2000h Under expansion · Q3 2026Multilingual voice commands, human-computer interaction dialogues, and synchronous acquisition of environmental sound fields. Aligns with mocap / video time sequence and supports joint modeling of speech and actions.
Exploring · Q4 2026From the scene design to the release of the dataset, every step is auditable and reproducible.
Customized collection · Data collection center solution · Long-term data subscription · Robot format adaptation（MuJoCo / Isaac / URDF）
📧 MotionDecode@chingmu.com
📱 Contact via WeChat and include in the remarks the name of your organization, your name, and the main purpose.
-->