GENTO Shows Humanoid Robots and 21-DOF Dexterous Hand at CoRL 2026

GENTO showcased its humanoid robots Luna and Marvin Pro, along with a 21-DOF dexterous hand prototype, at CoRL 2026, held Nov. 9-11, 2026, in Austin, Texas. roboticsandautomationnews.com reported on the GENTO exhibit on Nov. 13, citing the technology foundation of Tianji Intelligence, which starts with torque sensors inside the joints. Marvin Pro was used in a demonstration of NVIDIA’s GR00T 2, performing a task involving handling objects with both hands.

GENTO’s announcement detailed the hand’s fingertip force and positioning accuracy, as well as the torque-sensing performance of Tianji Intelligence’s force-control arm platform, providing specifications for supporting operations involving contact with objects. The hand prototype was displayed as a static exhibit, while Luna used a different hand configuration in its demonstration, so the specification figures should be distinguished from demonstrated performance.

The exhibit showed that GENTO is presenting the robot body not merely as an endpoint for AI model output, but also highlighting components for contact-based movement, including joint torque sensing, force-control arms and hand mechanisms. The hand prototype has 21 degrees of freedom, with specifications including 10 N of fingertip force and positioning accuracy of ±0.05 mm. However, the hand was a static exhibit, and a different hand configuration was used in Luna’s demonstration. The published specifications and the performance of robots operating at the event therefore need to be assessed separately. In the exhibit, Tianji Intelligence’s force-control arm platform supported robot movement, while Marvin Pro was used in a demonstration of two-handed manipulation using GR00T 2. The announcement described the model as simultaneously predicting future video frames and robot actions, and transferring human hand movements to a robot. The demo showed the model’s capabilities on a physical robot handling objects, rather than only through on-screen output. However, the demonstration does not establish the performance of production robots or repeatability across a range of work environments. A previous article by Physical AI News, “Welcome, GR00T 2!” published by humanoid.guide, said the information provided did not allow verification of physical-robot performance or deployment status. The latest announcement shows an example of manipulation using Marvin Pro, but does not provide data for a quantitative assessment of performance. There is also a connection to the topic of motion adaptation discussed in NVIDIA introduces SOMA Retargeter to adapt human motion for robots, but the latest materials alone do not establish a technical relationship or shared approach between the two. The industrial focus is how control specifications from sensors to the end effector connect with model-generated actions. The 0.02 N•m sensing threshold and 160 Hz bandwidth in Tianji’s materials apply to the arm platform and are different from the hand prototype’s specifications. Future evaluation will need to address how grip force and positioning accuracy are maintained in actual work, and how teleoperation and data-collection systems are used in training and evaluation. The published materials do not provide quantitative results from the demonstration, long-duration operating results or details of the processes from data collection to model updates.

According to GENTO, the hand prototype uses a linkage mechanism with five degrees of freedom in the thumb and four in each of the other fingers, and supports hybrid force and position control. It weighs 750 g excluding cables and measures 200 × 96 × 45 mm. Its maximum gripping diameter is 100 mm, rated fingertip force is 10 N, minimum opening-and-closing time is 0.5 seconds, and repeat positioning accuracy is ±0.05 mm. Its no-load gripping life is listed as 1 million cycles. It uses an industrial Ethernet connection and operates at DC 36–60 V. The prototype was displayed statically at the booth, and Luna used a different hand configuration in its demonstration. According to Tianji Intelligence’s technical materials, its force-control arm platform has torque sensing down to 0.02 N•m and a torque-loop bandwidth of 160 Hz. The announcement described GR00T 2 as a model that simultaneously predicts future video frames and robot actions. It also presented a demonstration in which human hand movements were recorded with a wearable exoskeleton and transferred to a robot. Marvin Pro is a fixed-base, dual-arm platform that demonstrated selecting and grasping objects in response to spoken instructions in English.