High‑fidelity hand‑motion capture launched

GenRobot has launched a hand-worn data collection system called Gen DAS Dex, and the important part is not the hardware alone. The company is selling it as a way to manufacture the raw material that embodied AI still lacks: dense, synchronized records of how human hands actually move, touch, and manipulate objects in the world. On its product page, GenRobot describes the device as a “high-precision multimodal data collection system” for robotics, with 23 degrees of freedom, 1 millimeter tactile spatial resolution, 0.05 newton tactile sensitivity, millisecond-level synchronization, and a wrist camera with a 150-degree field of view. (genrobot.ai) That framing matters because robotics has a data problem, not just a model problem. Vision-language models feasted on the internet. Robot manipulation models cannot. There is no web-scale archive of human grasping, finger pressure, wrist rotation, slip correction, and object contact. If you want a robot to learn how to pick up a cable, twist a cap, or shift a mug in its fingers without dropping it, you need demonstrations that capture more than video. You need trajectories, joint angles, contact, and timing, all lined up well enough to train on. GenRobot’s broader pitch is exactly that pipeline: capture, upload, govern, and turn interactions into training data. The company says its platform already spans more than 50 scenarios and more than 10,000 hours of embodied-intelligence data. (genrobot.ai) Gen DAS Dex is built to fill the hardest part of that pipeline, which is the hand. The listed specs suggest a lightweight exoskeleton-style device rather than a bulky laboratory rig. GenRobot says it weighs 210 grams, fits a range of hand sizes, can be put on in five seconds, and streams several modalities at once: fingertip trajectories, 3D tactile data, joint angles, hand spatial localization trajectories, and high-definition hand imaging. It also claims 200 hertz output, 1 millisecond signal latency, and factory calibration. Those details point to a product designed for repeated collection sessions by many operators, not a one-off research prototype. (genrobot.ai) The inclusion of touch is the sharpest signal about what GenRobot thinks the next bottleneck is. In dexterous manipulation, vision tells a robot where an object is. Touch tells it what is happening right now. Tactile feedback is what lets a hand notice slip, modulate force, and keep manipulating when the camera view is poor or the object is partly hidden. Recent research has pushed this point hard. A 2024 paper in *Nature Communications* showed that multimodal tactile sensing fused with vision improved robotic decision-making in contact-rich tasks, especially where slip and fine force control mattered. Columbia engineers made a similar case when they demonstrated tactile-visual systems handling fragile objects and in-hand adjustments better than vision-only approaches. (nature.com) That is why GenRobot is not treating Dex as a standalone gadget. The company says the new hand system pairs with its earlier ego-centric capture tools to form a “head + hand” collection setup, combining what a person sees with what the hand does. Its public dataset pages already reflect that logic. One sample ego dataset includes household tasks with ego-SLAM pose information, while the company’s open dataset page advertises large-scale real-world clips for embodied AI training. The pattern is clear: record the scene from the human point of view, record the hand at high fidelity, then use both to train policies that can be deployed through GenRobot’s controller stack. (711btc.com) There is still a gap between collecting human data and producing robot skill. Human hands do not map cleanly onto robot hands. Researchers behind systems like DexUMI have had to build explicit hardware and software bridges to reduce that embodiment gap. GenRobot is plainly betting that richer capture narrows the problem enough to make the translation worthwhile. The company’s own numbers show how far it is pushing on fidelity: sub-millimeter trajectory reconstruction, 3-millimeter magnetic encoders, and tactile sensing at 1-millimeter spatial resolution across the fingers, all packed into a 210-gram wearable with a three-hour battery. (dex-umi.github.io)