Robotics & Physical AI
ROBOTERA Opens Orders for XHAND 1 Pro-U Dexterous Robot Hand

Key PointsROBOTERA's Pro-U hand mirrors NVIDIA's DexUMI exoskeleton so human demonstrations map…
View more
- ROBOTERA's Pro-U hand mirrors NVIDIA's DexUMI exoskeleton so human demonstrations map directly to robot joints.
- GR00T 2 succeeded on all three bimanual tasks while GR00T 1.7 and π0.5 completed none.
- Openly licensed capture hardware paired with a commercial hand lets teams gather force-aware data without occupying the robot.
ROBOTERA announced the commercial launch of its XHAND 1 Pro-U on November 16, 2026, opening orders for an initial 40-unit release of the five-finger dexterous robot hand, according to its launch release.
The launch follows the 16-degree-of-freedom hand’s use in NVIDIA’s GR00T 2 research, including a DexUMI demonstration presented by Jim Fan at the Conference on Robot Learning (CoRL). ROBOTERA is targeting research teams, robotics developers and industry partners working on dexterous manipulation, robot learning and real-world robotic applications.
Matching the Hand to Human Demonstrations
ROBOTERA designed the Pro-U around NVIDIA GEAR Lab’s DexUMI exoskeleton, a wearable device that records human demonstrations without requiring a robot. The company says corresponding joints map movements directly between the exoskeleton and robot hand, reducing additional motion conversion. The redesign also adjusts thumb positioning, palm-side tactile sensing and articulation for natural demonstrations and varied grasping postures.
The GR00T 2 whitepaper says GEAR DexUMI reproduces the XHAND1 Pro-U’s joint kinematics, contact surfaces, wrist-camera mount and tactile-data format, with a one-to-one mapping onto the robot’s joints. Each GEAR DexUMI unit combines 15 magnetic encoders for finger-joint tracking, an OAK wrist camera and fingertip sensors for normal and shear force. Two units can be worn together to record bimanual demonstrations.
From Demonstration to Robot Execution
During collection, a wearer manipulates real objects directly while DexUMI records hand movements, wrist-camera video and fingertip forces. Those demonstrations are used to post-train GR00T 2, and the learned skills run on a dual-arm Tianji robot fitted with two XHAND1 Pro-U hands. For the three evaluated tasks, post-training used only the wearable demonstrations, without task-specific teleoperation or additional general robot-motion data; no operator remotely controlled the target robot during collection.
The force recordings also shape control. NVIDIA converts fingertip forces into joint torques and adds them to the robot’s action targets. During execution, predicted torque lowers each joint’s stiffness and offsets its position target, allowing the hand to apply the predicted force while accommodating positional uncertainty. That mechanism supports contact-rich actions such as slicing, unscrewing a lid and striking a match.
NVIDIA publishes the DexUMI hardware designs and documentation under CC BY 4.0, excluding vendor product photos, and its data-collection software under Apache-2.0 by default. The vendor’s XHAND robot-description files are separately MIT-licensed, while ROBOTERA’s commercial hand supplies the execution side of the same workflow.
Results Across Three Bimanual Tasks
The evaluation covered three multi-step sequences combining tool use, hand-to-hand transfers and lid handling. Banana-yogurt preparation involved drawing a knife, slicing the banana, removing the yogurt lid and moving pieces into a bowl. Granola scooping required passing a jar between hands, unscrewing its lid and using a spatula. In fire chemistry, the robot struck a match, lit a lamp beneath a flask and extinguished the match in water.
In NVIDIA’s evaluation, GR00T 2 achieved overall success rates of 60% for banana yogurt, 80% for granola scooping and 75% for fire chemistry. GR00T 1.7 and π0.5, post-trained on the same DexUMI demonstrations with the same data recipe, completed none of the three full tasks. These are results for the specified test sequences, not a general reliability figure for other tasks or operating environments.
Pairing openly licensed capture designs with a commercial matching hand gives research teams a specific route to collect force-aware demonstrations without tying up the target robot. The remaining constraint is execution: the results demonstrate transfer across all three sequences, but the unsuccessful trials leave consistent end-to-end performance as the open question for broader deployment.












