Humanoids · November 17, 2026 · 5 min
The glove came first. Now Robotera is selling the robot hand built to match it
Robotera opened orders on November 16 for the XHAND 1 Pro-U, a 16-joint hand designed as a mirror image of NVIDIA’s DexUMI exoskeleton. Only 40 are in the first batch, and the price is still a secret.
For most of robotics history, the hardware came first and the training data was an afterthought. You built a hand, then figured out how to teach it. Robotera’s newest product flips that order. The XHAND 1 Pro-U, which the Beijing company put up for commercial order on November 16, was designed backward from a wearable data-capture glove, so that whatever a person does in the glove, the robot hand can do with the same joints.
A hand shaped by its teacher
The glove in question is DexUMI, an exoskeleton from NVIDIA’s GEAR Lab. A person straps it on and goes about a task with real objects. It logs the angle of every finger joint and the forces at the fingertips, so it records not just where the fingers went but how hard they pressed. The Pro-U has 16 degrees of freedom arranged to line up with that exoskeleton joint for joint, which means a recorded demonstration can play back on the robot without being translated from one hand shape to another. Robotera says that translation step is where fine finger detail tends to get lost.
To get there, the company reworked its earlier XHAND 1. It moved the thumb, added touch sensing on the palm side and changed how the hand articulates, so a person wearing the glove can move naturally and cover more kinds of grip.
What the demo actually proved
The Pro-U is the hand NVIDIA used when its robotics director Jim Fan showed GR00T 2 at the Conference on Robot Learning in Austin. A two-armed robot fitted with two of the hands made a bowl of banana yogurt, scooped granola, and in the showiest test struck a match, lit a lamp under a flask and dropped the match in water. The model was post-trained on glove demonstrations alone, with no one steering the robot remotely for those tasks and no extra robot practice data. Reported success rates were 60 percent for the yogurt, 80 percent for the granola and 75 percent for the match sequence.
Two things are worth keeping straight. Those are results on three specific sequences, not a reliability rating for everyday work. And according to the GR00T 2 whitepaper, as reported by Unite.AI, GR00T 1.7 and Physical Intelligence’s π0.5, trained on the same glove data with the same recipe, did not complete any of the three full tasks. That gap is the real headline for NVIDIA. For Robotera, the headline is simpler: its hand was the one on stage.
Why the glove matters more than the hand
Collecting robot training data is expensive mostly because it usually needs a robot, a skilled operator and hours of remote control. Take the robot out of the collection step and you can gather demonstrations anywhere a person can wear a glove. We have watched this idea spread all year. Japan’s AIRoA-0 learned two-handed chores from people holding grippers, and Panasonic’s T-Rex work argued that fingertip touch is what finishes a grasp. DexUMI pushes both ideas into a full five-finger hand: human motion plus contact force, captured without a robot in the room.
The catch has always been the body at the other end. If the robot’s hand does not match the human’s, the data loses value on the way over. Robotera’s pitch is that it has closed that gap by building the matching body and selling it.
The open parts, and the closed ones
Much of this stack is already public. NVIDIA has posted the DexUMI exoskeleton’s parts list, mechanical and electrical design files and assembly guide on GitHub, with the hardware designs under a CC BY 4.0 license and the software mostly under Apache 2.0. Fan has said GR00T 2’s weights and its cross-embodiment training recipe will be released in December. That follows Runway’s decision to open the weights of its Praxis-1 robot model earlier this month.
The hand is the piece you cannot download. Robotera is taking orders through its marketing team for a first run of 40 units and has not published a price, weight, grip force or delivery schedule. That makes this a research-lab product for now, aimed at teams working on dexterous manipulation and robot learning, not something a hobbyist will add to a cart.
It also fits a pattern in the hand market. A report we covered this year found Chinese firms shipped 96 percent of dexterous humanoid hands in the first half of 2026. Robotera, which raised about $280 million in a round led by SF Group in April, now has a hand tied directly to an American lab’s flagship model, which is a stronger selling point than a spec sheet.
Our take
Forty units is a tiny batch, and an unlisted price usually means an expensive one. But the idea underneath is bigger than this product. Once the glove designs are free and the model weights land in December, a lab with the matching hand could try to repeat NVIDIA’s setup on its own bench. If that works outside NVIDIA, “designed for the data glove” could become a standard feature on robot hands. Follow Robotera and NVIDIA for the next step, compare the bodies these hands attach to in our humanoids guide, and find manipulation and robot-learning roles on our jobs board.
