Skip to content

Tools

Robotera opens orders for 40 robot hands used in NVIDIA’s GR00T 2 demo

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. Robotera opened commercial orders for the five-fingered XHAND 1 Pro-U, with 40 hands in the first batch.
  2. NVIDIA used a wearable with matching joints to record human demonstrations for its GR00T 2 robot model.
  3. Robotera is inviting inquiries from research teams and developers. Its announcement gave no price or delivery date.
PracticalDevelopers
An articulated robot hand on a stand against a navy background
AI-generated conceptual image: an imaginary articulated hand on a stand illustrates Robotera’s launch of a research robot hand. · AI Ping

Robotera said on November 16 that it had opened commercial orders for the XHAND 1 Pro-U, a five-fingered robot hand used in an NVIDIA GR00T 2 demonstration. The first batch consists of 40 hands. Research teams and robotics developers can now inquire about buying the model that NVIDIA used to test how a robot can learn movements demonstrated by a person.

The XHAND mounts on a robot arm and handles objects during a test. A separate wearable records the human demonstration. Robotera’s launch announcement invites inquiries from researchers, developers and industry partners. It gives no unit price, delivery date or list of shipping markets.

NVIDIA introduced GR00T 2 at the Conference on Robot Learning on November 9. Its research showed how a model could learn from a person wearing a hand-tracking device and then control a two-armed robot. A week later, Robotera opened orders for the hand that carried out the learned movements in that setup.

Recording the fingers and the force they apply

The XHAND has 16 mechanical degrees of freedom, a count of the independent ways its joints can change the hand’s pose. Its joints line up with those on NVIDIA’s GEAR DexUMI wearable. Robotera says it adjusted the thumb position, palm-side touch sensing and articulation to support a range of grasps.

A person wears the DexUMI while picking up or moving real objects. The wearable uses magnetic encoders to track finger angles and a camera at the wrist to record the action. Sensors at each fingertip measure the force pressing into a surface and the force running along it. In a movement such as twisting a lid, the recording can track both the motion of the fingers and the way they apply force as they turn.

Wearing the device on both hands lets the operator demonstrate actions such as holding a container steady with one hand while opening it with the other. The recording takes place at the work surface with the objects in hand. A team can collect these demonstrations without an operator steering the robot through each action.

A human hand wearing finger straps above joint models on a wooden bench
AI-generated conceptual image: Finger straps on a human hand and joint models on a bench illustrate how recorded finger motion can map to a robot hand. · AI Ping

How the robot did on a two-part breakfast task

NVIDIA used a previously trained GR00T 2 model and adapted it using the wearable demonstrations. The team used no extra robot motion data for that stage of the experiment. It then tested the model on a Tianji robot with two arms and an XHAND 1 Pro-U on each arm.

Two parts of a single breakfast routine were scored separately. In Table 5 of NVIDIA’s technical report, the robot finished the banana yogurt sequence in 60% of trials. It opened the granola jar, served the granola and put the jar back in 80% of trials. It placed the banana on the cutting board in 95% of trials. That move was scored on its own as well as being part of the longer yogurt sequence.

Robotic fingertips approach an orange block beside a cup and cork disk on a teal mat
AI-generated conceptual image: A cup and block on a teal mat illustrate the stage at which a robot hand applies learned movements to physical objects. · AI Ping

What’s public and what’s on sale

NVIDIA made the wearable’s designs and data collection software public on November 9. The documentation includes a parts list and mechanical and electrical designs. Robotera opened orders for the robot hand that executes the learned movements. Its launch announcement gives no terms for buying a wearable with the hand.

NVIDIA also tried data from the same wearable on a Sharpa hand with 22 degrees of freedom. Its joints are arranged differently, so the team recalculated the recorded movements for the new hand. For that test, the researchers also trained on recordings of the Sharpa robot handling objects alongside the converted demonstrations. It adapted the model using the converted joint positions. The force targets calculated for the XHAND did not carry over to the Sharpa hand.

Sources