NVIDIA Uses Tianji Intelligence’s Marvin Pro in GR00T 2 Demonstration; Company Unveils 21-DOF Dexterous Hand

Tianji Intelligence announced at CoRL 2026 that its dual-arm robot platform, Marvin Pro, was used in a demonstration of NVIDIA’s GR00T 2. Under its GENTO brand, the company also unveiled a self-developed 21-DOF dexterous hand at the event. According to the company’s materials, GR00T 2 was presented for transferring skills from human demonstrations to robots, and Marvin Pro was used to demonstrate bimanual tasks.

The bimanual-task demonstration using GR00T 2 combined not only the model’s motion prediction but also physical execution by Marvin Pro and dexterous hands. However, the publicly unveiled 21-DOF hand was a prototype shown as a static display, so its published specifications should be assessed separately from its operating performance in a demonstration.

The key distinction in this announcement is that Marvin Pro, used in the GR00T 2 demonstration, and the 21-DOF hand unveiled at the same event were not presented as parts of the same physical configuration. The company’s materials say Marvin Pro was a demonstration robot equipped with two dexterous hands, while explicitly describing the hand prototype as a static display at the booth. Therefore, the materials do not establish that the 21-DOF hand was used in the GR00T 2 demonstration or integrated with the model. The configuration’s significance is that the demonstration involved not just the robot’s movements but also the hand and force control needed to manage contact with objects. The company’s materials say GR00T 2 transfers skills from human demonstrations and describe bimanual tasks using hand movements recorded with a wearable exoskeleton. The published features of the hand, meanwhile, include a structure that allows its joints to move under external force and hybrid force-position control — mechanical and control design features for contact. Model-based motion generation and the mechanisms and controls for touching objects are separate elements; the demonstration appears to test how they work together. The distinction between a model’s announcement and validation on a physical robot, highlighted in our related coverage, humanoid.guide’s report on NVIDIA’s GR00T 2, remains relevant here. A model’s ability to handle human demonstrations is not the same as its ability to reliably manipulate objects with different shapes and contact conditions on a physical robot. The use of Marvin Pro in the demonstration is confirmed, but the company’s materials alone do not establish the results of tasks performed with the 21-DOF hand or its performance during repeated operation. From a business perspective, Tianji Intelligence’s effort to add dexterous hands to its product lineup alongside arms has significance beyond unveiling a standalone end effector. The focus will be how far the arms’ torque-sensing and control specifications can be coordinated with the hand’s force-position control as a single operating system. However, the 0.02 N•m and 160 Hz figures in the materials apply to the arms, not the hand. Questions to watch include the hand model installed on the GR00T 2 demonstration robot, the relationship between the unveiled prototype and the demonstration robot, task success rates and durability in real-world operations, and how force control is connected to model output.

The unveiled dexterous-hand prototype has five degrees of freedom in the thumb and four in each of the other fingers, and uses a linkage mechanism. It weighs 750 g excluding cables and measures 200 × 96 × 45 mm. Its specifications list a maximum gripping diameter of 100 mm, rated fingertip force of 10 N, a minimum opening-and-closing time of 0.5 seconds, repeat positioning accuracy of ±0.05 mm and a no-load gripping life of 1 million cycles. It has an industrial Ethernet interface and operates at DC 36–60 V. The prototype was shown as a static display at the booth. The company’s materials describe Marvin Pro as a fixed dual-arm platform. In the GR00T 2 demonstration, it was fitted with two dexterous hands and shown performing bimanual tasks using skills based on hand movements recorded with a wearable exoskeleton. The materials describe GR00T 2 as a model that simultaneously predicts future video frames and robot actions. Tianji Intelligence’s force-control arms are specified as having a detectable torque of 0.02 N•m and a torque-control loop bandwidth of 160 Hz. These figures are specifications for the arm platform, not the hand.