MotionDecode is an optical robot motion capture dataset released by ChingMu, focusing on high-precision optical motion capture for embodied intelligence and humanoid robots. It is applicable to fields such as humanoid robots, dexterous hands, embodied intelligence, and virtual production. This dataset contains over 1,000 hours of robot motion trajectory data, covering the entire skeleton, finger joints, 6D object poses, multi-view videos, and semantic tags. It covers more than 15 real-world scenarios, including industry, home, retail, healthcare, logistics, agriculture, and performance, involving more than 200 trackable objects and over 500 standardized tasks. All data has undergone automated cleaning and manual quality inspection.

Dataset composition:

• Raw motion data: BVH format, Y-axis upward, sampling rate 120 frames/second, rotation expressed in ZYX Euler angles, displacement in centimeters, containing 47 to 67 skeletal joints.
• Robot redirection trajectory: CSV format, including root position (in meters), quaternion rotation, and joint angles (in radians).
• Object 6D pose: CSV format, including position (in meters) and quaternion rotation, sampled at 120 Hz.
• Multi-view video: MP4 format, shot simultaneously by 4 to 8 spatially and temporally registered cameras.
• Semantic tags: JSONL format, used to annotate task category, scenario type, specific actions, and interaction objects.

Citation

If you use the dataset in your research, please credit the data source: ChingMu.