MotionDecode Data Open Program: 1,000 Hours of High-Quality Human Motion Data – Now Free to Apply.

On June 16, CHINGMU and EnDecodeX officially launched the "MotionDecode Data Open Program," offering the first 1,000 hours of high-quality human motion data free to applicants worldwide.

Since its release, MotionDecode has drawn significant attention. Today, we're breaking down the core capabilities of the MotionDecode dataset – and filling you in on the highlights we haven't had the chance to cover yet!

01丨What is MotionDecode?

The development of Embodied AI increasingly relies on real, high-quality data. For robots to truly understand human motion, video alone is far from enough – what's needed is high-precision, multimodal motion data ready for direct training.

MotionDecode is a high-quality motion data system jointly built by CHINGMU and EnDecodeX for Embodied AI training, robot learning, and simulation validation. It covers a full range of data modalities – body, hand, object, video, environment, and task labels – and delivers end-to-end data services spanning real-world motion acquisition, multimodal synchronization, quality screening, robot retargeting, and customized data delivery.

▎Key Features

• Larger Scale

3,000 hours of data built to date, with the first 1,000 hours of high-precision human motion data free to use, and an estimated annual production capacity of 500,000 hours.

• Higher Precision

Captured by CHINGMU's optical motion capture system, delivering sub-mm spatial positioning and joint angle error < 0.5° – ensuring industrial-grade precision that meets the stringent requirements of robot control training.

• More Modalities

Body | Hand (with force feedback) | Ego (first-person view) | Exo (third-person view) | Object 6D | Environmental scene

• Broader Scenarios

Covering diverse application scenarios including industrial manufacturing, home services, retail, medical rehabilitation, logistics & warehousing, ball-based interactions, and digital entertainment. Includes 500+ human-robot tasks and 200+ object types and tools – providing abundant material for algorithm training in vertical scenarios.

• Easy to Use

Compatible with 10+ standard formats including BVH, FBX, and CSV, with support for mainstream robot simulation formats such as MuJoCo, Isaac, and URDF – lowering the barrier to integration so algorithm teams can stay focused on the models themselves.

▎Standardized Data Production Workflow

MotionDecode is not just about capturing a single take. Every piece of data goes through a standardized workflow – from task planning to final asset delivery:

02｜Multi-Scenario Motion Demos

The MotionDecode dataset covers a rich variety of motion types. After robot retargeting, the data can be directly transferred to humanoid robot training, imitation learning, and skill acquisition.

03｜How to Access the Open-Source Data?

Join the MotionDecode Early Access & Feedback Program. The first 1,000 hours of high-quality human motion data are now open for free application via QR code, with continuous updates rolling out. More sample data will be released on GitHub and Hugging Face. Stay tuned for the latest updates on the CHINGMU official website.

A solid data foundation is the key to an intelligent future. CHINGMU looks forward to working with Embodied AI developers, research institutions, and industry partners worldwide to co-build a thriving data ecosystem – and accelerate the adoption of robots across every sector!

MotionDecode Homepage: https://chingmudata.github.io/MotionDecode/
MotionDecode Dataset: https://huggingface.co/datasets/CMRobot/MotionDecode
Open Data Inquiries: https://www.wjx.cn/vm/eJk1hin.aspx – fill out the form and we'll get back to you shortly.