Hand Motion Dataset Dashboard

Interactive statistics for bimanual hand motion data curation
TL;DR RoMo Hands scales pure-hand text-to-motion data by an order of magnitude — ~542K in-the-wild bimanual sequences, 884 hours, 95.5M frames @ 30 fps (~10.6× the duration of the next pure-hand dataset, CLUTCH). This page benchmarks RoMo Hands against GRAB and GigaHands on durations, activity mix, per-finger and per-joint dynamic scores, smoothness, and data-quality flags. Headline take-away: a stringent dexterity filter pushes RoMo Hands to 0.2% static / 53.7% active with mean dynamic score 0.33, sitting between GRAB's choreographed 0.66 (1.1K seq, 2.86h) and GigaHands' uncurated 0.14 (12.7K seq, 16.4h) while delivering orders of magnitude more data.
Select a hand_dataset_stats.json file to begin.