arslan219/MotionDecode

ChingMu 1000-Hour Embodied Motion Dataset

High-precision optical motion capture data for humanoid robots, dexterous hands, embodied AI, and virtual production.

Duration	1000+ hours @ 120 Hz
**Scenarios **	15+ real-world scenes
**Tasks **	500+ standardized tasks
Objects	200+ tracked props (6D pose)
Modalities	Skeleton · Finger · Object 6D · Video · Labels
**Formats **	BVH · Retargeted CSV · NPZ

✅ Access note: This dataset is fully open and publicly accessible.

---

Key Features

• Optical ground truth – sub-mm accuracy, 120 fps, no estimation errors.

• Dexterous hands – 20+ DoF per hand, synchronized with object 6DOP pose.

• Robot-ready – pre-retargeted to Unitree G1; custom retargeting available.

• Real-world diversity – 15+ scenarios, 500+ tasks, 200+ objects.

• Multi-modal – full-body skeleton, finger motion, object pose, multi-view video, semantic labels.

• Quality assured – every take passes automated cleaning + manual inspection; quality flags provided.

---

Dataset Summary

ChingMu 1000H is an optical motion capture dataset designed for training and validating embodied AI and humanoid robot controllers. It covers full-body skeleton, finger articulation, object 6D pose, multi-view video, and semantic labels across 15+ real-world scenarios (industrial, household, retail, healthcare, logistics, agriculture, performance). All data is cleaned, quality-assessed, and robot-retargeted.

---

Data Format Specifications

Component	Format	Details
Raw motion	.bvh	Y-up, 120 fps, ZYX rotation, cm, 47–67 joints
Retargeted trajectories	.csv	Root position (m), quaternion, joint angles (rad)
Object 6D pose	.csv	Position (m) + quaternion, 120 Hz
Multi-view video	.mp4	4–8 cameras, co-registered
Semantic labels	.jsonl	Task, scenario, action, object

🎥 Preview Video

Watch a short demonstration of the motion capture data in action:

Demonstration of full-body motion capture with real-time skeleton overlay and object tracking.

Intended Uses

• Imitation learning / motion policy training for humanoids

• Dexterous manipulation datasets (hand-object interaction)

• Motion generation & retrieval (text/motion cross-modal)

• Sim-to-real validation (MuJoCo via retargeted trajectories)

• Virtual production & animation reference

---

Full Taxonomy (abridged)

• Locomotion → walk, jog, crouch-walk...

• Manipulation (whole-body) → shelf-pick-place...

• Dexterous Hand → pinch, precision-grasp...

• Tool Use → screwdriver, wrench...

• Object Interaction → door-open/close...

• Social / Contact → handoff-object...、

• Performance → dance, martial-arts...

👀 Try it live: Use the Dataset Preview panel at the top of this page to filter and explore the actual index table. Select the metadata config to browse available takes.

ℹ️ The full index with all rows is best viewed locally. Download metadata/index.csv to open in Excel or pandas for complete filtering.

---

🖥️ Interactive Showcase

Visit our dedicated showcase website for interactive demos, comparison videos, and detailed visualizations:

Includes: trailer video, modality breakdowns, robot retargeting comparisons, and more.

---

🆕 Open-Source Release: Unitree G1 Retargeted Data

We are releasing 100 hours of robot-ready motion trajectories retargeted to the Unitree G1 humanoid. All data is provided in CSV format under the samples/ directory. Please indicate the source of the data when using it: from Chingmu.

Quick Start

pip install huggingface_hub

from huggingface_hub import hf_hub_download

repo_id = "CMRobot/MotionDecode"
file_path = hf_hub_download(
    repo_id=repo_id,
    filename="samples/1.1.Basic_Movement_Category/1.1.1.High_Dynamic_Movement/1.1.1.1.Standing_High_Jump/BM_Standing_High_Jump_00001.csv",
    repo_type="dataset",
    local_dir="./robot_samples"
)
print(f"Downloaded: {file_path}")

---

Quality & Limitations

Quality controls: marker swap correction, gap-filling (≤6 frames), foot skating detection, manual review. Flags: pass, warning, fail.

Accuracy: joint error <1mm, object pose ±2mm / ±0.5°, temporal sync <1 frame.

Limitations: performer age skew (20–35), object accuracy varies with marker cluster size.

---

Get Full Dataset

The entire dataset is publicly available here. If you have any questions about the dataset or would like to know more information, please contact us through the following channels:

• For Chinese users: Scan the QR code below to contact us via WeChat, and include in the remarks the name of your organization, your name, and the main purpose.

For international users: Join our Discord community

Alternatively, you can click the "Request access" button on the right side of this page to automatically gain download permissions for the complete dataset.

Or email us at: MotionDecode@chingmu.com

We look forward to collaborating with researchers and industry partners!