Tracking both hands in video when they're hidden or out of frame

Recovering precise 3D hand motion from first-person video is essential for robot learning, but hands frequently disappear from view or get hidden by objects. StableHand uses a flow-matching generative model that predicts per-frame reliability scores for hand observations, then intelligently weighs them during reconstruction—trusting high-confidence frames while inferring missing ones from learned hand-motion patterns. On two egocentric benchmarks with heavy occlusion and missing frames, it cuts position error by 20–25% versus prior methods.