Noitom Robotics has publicly released HiPHI-Hand, the motion-capture dataset it used to help train Nvidia's EgoHand model, a system that reconstructs 3D hand movement from a single first-person camera. The dataset went up on Hugging Face on November 9, timed to a live demonstration at the CoRL 2026 robotics conference, where attendees can wear a headset and watch their own hands get reconstructed in 3D in real time.
Nvidia's EgoHand model has already logged more than 40,000 downloads on Hugging Face in the roughly two weeks since it launched, a sign of how much demand there is for tools that let robots and AI systems understand exactly what a human hand is doing, down to individual finger joints.
How do you teach a computer to see a hand?
Hands are a notoriously hard problem in computer vision because fingers constantly occlude each other and move fast, and a single video frame rarely shows every joint clearly. To solve this, Noitom captured hand-object manipulation using an optical motion-capture system with micro-markers glued onto the fingers, giving precise, ground-truth 3D positions for every joint. That marker data was synchronized with ordinary first-person video recorded through Noitom's SOL ego-vision headset, worn the way a person would naturally use their hands while cooking, assembling parts or handling tools.
The result is 11 hours of paired data: video on one side, exact 3D hand geometry on the other. Training on video paired with precise ground truth, rather than video alone, is what lets a model like EgoHand learn to infer the full 3D shape and pose of a hand from camera footage that normally hides half the story. Nvidia built EgoHand on its Isaac Capture pipeline specifically for this kind of egocentric, first-person hand-tracking work.
Why give the data away?
This isn't a one-off for Noitom. The company had already released a related dataset, ITW-Ego Plus, earlier in the week, adding full-body motion capture to its egocentric data library. Releasing training data openly, rather than keeping it proprietary, is a deliberate positioning move: it establishes Noitom as the default data provider for hand- and body-tracking research, the same way certain image datasets became the default benchmark for computer vision a decade ago.
For the robotics industry, precise hand-tracking data solves a real bottleneck. Humanoid robots and robotic arms increasingly learn manipulation skills by imitating human demonstrations captured on video, but that imitation only works if the underlying model can accurately read what the human hand actually did. Blurry or ambiguous hand data has been a known source of failed grasps and dropped objects in robot learning pipelines. A freely available, precisely labeled dataset lowers the barrier for any lab or startup trying to train its own dexterous-manipulation models, rather than relying solely on Nvidia's pretrained EgoHand.
Noitom's live CoRL demo, running November 9 through 11 at booth 1105 on the JW Marriott's third floor, doubles as a sales pitch: watching your own hand appear in 3D on a screen in real time is a more convincing argument for the technology than a benchmark chart. With EgoHand's download numbers already climbing fast, the open dataset behind it is likely to show up in robot-hand research well beyond Nvidia's own models.