QUDMI an open, lightweight imitation-learning model that turns sparse VR/XR/AR tracker signals into believable full-body avatar motion in real time

Every consumer headset gives you three tracked points: the head and two hands. Full-body avatars still need to fake the rest (spine, hips, legs, feet), and most apps either skip legs entirely or use hand-tuned inverse-kinematics that looks stiff. Qudmi learns that mapping from motion-capture data instead of hand-coding it, and ships it as a small (~5-20M parameter) model any engine can run locally at real-time latency.