Glossary · Biomechanics
Motion capture
Also called mocap, stock motion, motion clip, animation clip
Recording a real person's movement, usually with markers or cameras, so a computer can replay it on a digital body.
Motion capture (mocap) records how a real person moves. In the classic setup the performer wears reflective markers and a ring of cameras tracks them in 3D. Newer systems use body-worn sensors, or estimate the pose from ordinary video. The result is a clip: the angle of every joint, many times a second.
Films and games use mocap to animate characters. Companies sell libraries of ready-made clips (stock motion), and game engines such as Unreal come with their own.
Mocap in robotics and RL
Much research on simulated athletes starts from mocap. The clip becomes a reference motion, and a policy is rewarded for copying it under physics. This works well, but the policy can only be as good as the clip. A stock “punch” recorded by an actor who doesn’t fight carries the actor’s mistakes. A clip recorded on a different skeleton also has to be retargeted (translated) to the simulated body, and some detail is lost on the way.
In this project
No stock motion is used. The first techniques (the stance, the balance, the jab) were built on the muscular MuJoCo body itself, from the coach’s descriptions, corrections and photos of himself in the key positions, so no retargeting was needed.
Larger movements (the knees, the kicks, the uppercuts over distance) come from motion capture of the coach, with open-source software (Pose2Sim) and two phones filming at 60 fps from about 60–90° apart. The phones are placed by calibration, the joints found in each video by a pose estimator on a GPU and triangulated in 3D, a skeleton scaled to him is fitted to them, and the result is retargeted onto the fighter, whose legs are about 25 % longer than the coach’s relative to his trunk. The coach reviews it against his own video, frame by frame, before it becomes a reference. The whole pipeline is in Correcting the knee.