Embodied AI Glossary中文Motion Tracking
运动跟踪CommonA whole-body control task where a robot reproduces a reference motion, such as human motion-capture data, in real time.
In humanoid robotics, motion tracking means giving a policy a reference motion — usually human motion-capture data retargeted into the robot's joint angles — and having the robot reproduce it frame by frame without falling over or violating its physical limits. The dominant approach follows 2018's DeepMimic: train with reinforcement learning in simulation, with a reward based on how closely each body part matches the reference pose, then transfer the trained policy to the real robot. It is the basic skill behind humanoids ‘learning human motion,’ underlying teleoperation, dancing, martial-arts performance, and many general-purpose whole-body controllers such as GMT and BeyondMimic. Don't confuse it with computer-vision object tracking or with motion capture itself — those are about ‘seeing,’ while motion tracking is about ‘doing.’
ExampleBeyondMimic trained a tracking policy on the LAFAN1 motion-capture dataset; a single set of hyperparameters let a humanoid robot perform cartwheels, spinning kicks, and sprints on the real hardware.
- Also called
- Motion Imitation, Reference Motion Tracking
- Related
- Motion Retargeting · DeepMimic · BeyondMimic · GMT · Whole-Body Control · Motion Capture
- Sources
- DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills (arXiv:1804.02717)
BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion (arXiv:2508.08241)
GMT: General Motion Tracking for Humanoid Whole-Body Control (arXiv:2506.14770) - As of
- 2025-11