Watch someone pour granola into a bowl while wearing a motion-capture exoskeleton, and a robotic hand across the room learns to do the same thing. That is the pitch behind ROBOTERA's XHAND 1 Pro-U, a five-finger dexterous robot hand the company launched commercially on November 16, 2026, built to pair directly with Nvidia's GR00T 2 foundation model and its DexUMI exoskeleton system.
The hand carries 16 degrees of freedom in a human-like configuration, with palm-side tactile sensing and a thumb geometry tuned to mirror how human hands actually grip objects. ROBOTERA is selling an initial limited run of 40 units, open to research teams, robotics developers and industry partners working on manipulation and robot learning, rather than a mass retail release.
From Human Demonstration to Robot Execution
The hand's headline feature is how it learns. A person wears Nvidia's DexUMI exoskeleton and performs a task on a real object; the exoskeleton captures hand movements and fingertip force data, which then post-trains Nvidia's GR00T 2 model. Because the XHAND 1 Pro-U's joint layout closely matches the exoskeleton's structure, those recorded movements transfer with minimal translation loss onto a dual-arm robot fitted with two of ROBOTERA's hands.
Nvidia's Jim Fan, director and distinguished research scientist, featured the hand in a GR00T 2 DexUMI demonstration at the Conference on Robot Learning. In that demo, the setup tackled three multi-step tasks: preparing banana yogurt, scooping granola, and striking a match for a fire-chemistry task, succeeding at rates of 60%, 80% and 75% respectively.
Why the Hand, Not Just the Model, Matters
Foundation models like GR00T 2 have improved quickly at deciding what a robot should do next, but execution still bottlenecks on hardware that can actually replicate fine human hand movement. A 16-degree-of-freedom hand with tactile sensing gives a model like GR00T 2 richer signal to learn from than the simpler two- or three-finger grippers common in industrial robots, and it lets demonstration data collected on a human captured through an exoskeleton map more directly onto a machine.
That compatibility is the real selling point for ROBOTERA's limited release. Rather than competing purely on hand specs, which rivals building dexterous hands have also chased, ROBOTERA is betting that tight integration with Nvidia's widely adopted GR00T training pipeline will attract labs that want to generate their own task data instead of waiting on pre-trained skills.
With only 40 units available at launch, the near-term audience is research labs and robotics developers rather than factory floors. The success rates from the CoRL demonstration, 60% to 80% depending on task complexity, suggest the hand-and-model combination is further along than fully reliable but still short of production-ready for unsupervised commercial deployment, according to Manila Times / PR Newswire.