Reconstructing full-body poses from hand and head movements alone

Reconstructing a person's full-body pose from sparse head and hand trajectories is essential for VR/AR telepresence, but current methods accumulate errors and produce unnatural movement. AtomicMotion breaks this problem into five functional body regions (arms, legs, torso, head), trains each to understand local movement patterns while preserving skeletal structure, and uses an attention mechanism that respects biological joint constraints. Results on AMASS show significantly better reconstruction fidelity and more biomechanically realistic motion than existing baselines.