--- dataset_info: license: other license_name: chingmu-terms license_link: LICENSE language: ["en", "zh"] pretty_name: "ChingMu Robot Motion Dataset" tags: - motion-capture - humanoid-robotics - imitation-learning - optical-mocap - bvh - dexterous-hands - whole-body-control size_categories: 1M High-precision **optical motion capture** data for humanoid robots, dexterous hands, embodied AI, and virtual production. | | | |---|---:| | **Motion Capture Duration** | **1000+ hours** @ 120 Hz | | **Scenarios** | 15+ real-world scenes | | **Task categories** | 500+ standardized tasks | | **Interactive objects** | 1000+ tracked props (6D pose) | | **Modalities** | Full-body skeleton · Finger DoF · Object 6D pose · Multi-view sync video · Task labels | | **Output formats** | BVH · Retargeted joint trajectories · NPZ · CSV | 🔒 **Access note:** Metadata, samples, and documentation are publicly browsable. Full raw captures and retargeted datasets are released under a **gated license** — click *Request access* to agree to the license terms and receive download instructions. --- ## Dataset Summary ChingMu 1000H is a high-precision, multi-modal motion capture dataset built for training and validating embodied AI and humanoid robot controllers. Data is captured using **optical mocap systems** (sub-mm precision, 120 fps) and covers: - **Full-body skeleton motion** (23–67 joints, BVH-compatible hierarchies) - **Finger & hand motion** (per-hand 20+ DoF, glove + marker hybrid) - **Object 6D pose tracking** (rigid-body markers → position + quaternion @120Hz) - **Multi-view video** (4–8 synchronized cameras, co-registered with mocap timeline) - **Semantic labels**: task name, scenario tag, skill category, quality flag Scenarios span industrial assembly, household service, retail interaction, healthcare assistance, logistics handling, agricultural work, and staged performance. ## ✨ Data Highlights - **Optical Ground Truth** – Sub-millimeter accuracy (<1mm), 120 fps, no estimation errors. - **Dexterous Hands** – 20+ DoF per hand, synchronized with object 6D pose for fine-grained manipulation. - **Robot-Ready** – Pre-retargeted to Unitree G1; custom retargeting available on request. - **Real-World Diversity** – 15+ scenarios (industrial, household, retail, healthcare, logistics, agriculture, performance), 500+ tasks, 1000+ objects. - **Multi-Modal** – Full-body skeleton (BVH), finger motion, object pose, multi-view video, semantic labels. - **Quality Assured** – Every take passes automated cleaning + manual inspection; quality flags (pass/warning/fail) provided. ## 📁 Data Format Specifications | Component | Format | Details | |---|---|---| | Raw motion | `.bvh` | Y-up, 120 fps, ZYX rotation, cm units, 47–67 joints | | Retargeted trajectories | `.csv` | Root position (m), root quaternion, joint angles (rad) | | Object 6D pose | `.csv` | Position (m) + quaternion per object, 120 Hz | | Multi-view video | `.mp4` | 4–8 synchronized cameras, co-registered timeline | ## ✅ Quality Assurance All data undergoes a rigorous pipeline: 1. **Marker swap correction** – automatic detection and repair. 2. **Gap filling** – cubic spline interpolation for gaps ≤6 frames; longer gaps flagged. 3. **Foot skating detection** – flagged when static foot drifts >2cm/s. 4. **Manual review** – each take reviewed by trained annotators. 5. **Quality flags** – `pass` (clean), `warning` (minor issues), `fail` (do not use). Typical accuracy: - Joint position error: <1mm (optical system limit) - Object 6D pose: ±2mm translation, ±0.5° rotation - Temporal sync between modalities: <1 frame ## 📸 Sample Visualization Below is a snapshot from our motion capture studio showing a subject performing a box-moving task, with real-time skeleton overlay and object tracking: ![Sample](assets/logo.png) *Figure: Optical mocap data visualized with skeleton (blue) and tracked object (red bounding box).* ## 🎥 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 / Isaac Sim via retargeted trajectories) - Virtual production & animation reference --- ## Task & Scenario Taxonomy (How to Filter) All motion takes are indexed in `metadata/index.csv`. The key **filter columns** you'll use most: | Column | Values (examples) | Use | |---|---|---| | `scenario` | `industrial`, `household`, `retail`, `healthcare`, `logistics`, `agri`, `performance` | Filter by scene | | `task_category` | `locomotion`, `manipulation`, `dexterous_hand`, `tool_use`, `interaction` | Broad category | | `task_label` | `walk_carrying_box`, `screw_with_driver`, `pinch_grasp_bottle`, `open_fridge_door` … | Specific task | | `has_finger_data` | `true` / `false` | Needs hand DoF? | | `has_object_6d` | `true` / `false` | Needs object tracking? | | `quality_flag` | `pass` / `warning` / `fail` | Skip bad takes | | `retarget_available` | `g1` / `soma` / `none` | Which robot formats exist | ### 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`](https://huggingface.co/datasets/ZIHLING/Chingmu-RobotData/resolve/main/metadata/index.csv) to open in Excel or pandas for complete filtering. --- ## Data Format Specifications | Property | Value | |---|---| | Up axis | Y-up | | Frame rate | 120 fps | | Rotation order | ZYX (BVH standard per-joint) | | Units | cm for position, degrees for channels | | Joint count | 47–67 depending on finger inclusion | --- ## Quality & Limitations ### Quality controls applied - Marker swap detection & auto-correct (per-take) - Rigid-body jitter filter (median + threshold) - Foot-skating metric computed (flagged when >2cm/s static-foot drift) - Gap-fill ≤ 6 frames via cubic spline; longer gaps tagged `quality_flag=warning` ### Known limitations - Performer population: currently skewed toward 20–35yo (planned expansion in v2) - Finger data precision depends on glove calibration per session - Object 6D pose accuracy: ±2mm translation, ±0.5° orientation ### Ethical & Privacy - All performers signed informed-consent & appearance release - Faces are **not** included in skeleton/metric data - No biometric identifier is retained in the released features --- ## Get Full Dataset Only a subset of the dataset is publicly available here. If you need **full access** to the entire 1000-hour dataset, please scan the QR code below to contact us via WeChat: ![WeChat QR Code](assets/logo.png) Alternatively, you can click the **"Request access"** button on the right side of this page to automatically gain download permissions for the complete dataset. *Note: Approval is automatic, but we kindly ask you to provide your contact information for licensing purposes.* Or email us at: **dataset@chingmu.ai** We look forward to collaborating with researchers and industry partners! ## Contact - Issues / Requests: [Discussions](https://huggingface.co/datasets/ZIHLING/Chingmu-RobotData/discussions) - Email: dataset@chingmu.ai