NVIDIA’s Humanoid Reference Robots Use China-Linked Hands in Two Models, Report Says

humanoid.guide reported on November 3, 2026, that two of NVIDIA’s humanoid reference platforms use products from robot-hand companies with ties to China. According to the report, the latest model, Agile One SN, pairs an Agile Robots body with the WUJI Hand 2, while the earlier design combines a Unitree body with a Sharpa hand. The report said both hand suppliers have ties to Shanghai.

NVIDIA’s two reference robots combine bodies and robot hands made by different companies. For developers, they illustrate how hardware-specific features such as a hand’s degrees of freedom and tactile sensing can shape the requirements of building a physical robot, even when using development platforms such as Isaac.

The difference between the two reference robots is that NVIDIA’s robotics development platform is not represented by a finished robot from a single manufacturer; instead, the body and manipulation hardware are sourced from different suppliers. The earlier design reportedly used a Unitree body and a Sharpa hand, while Agile One SN uses an Agile Robots body and the WUJI Hand 2. At least in these two examples, the reference platform’s configuration is not tied to a specific body manufacturer. However, the available information does not show how readily software or skills can be transferred between different bodies and hands. Specifications in NVIDIA’s official announcement for Agile ONE SN say that each hand has 20 degrees of freedom, force control and tactile sensing, and that the robot combines these features with Jetson Thor computing and Isaac GR00T. This suggests a design that gives developers a robot’s manipulation capabilities as a system encompassing not only software but also hand degrees of freedom, tactile information and onboard computing. The 1kHz force-control and tactile-sensing figures, however, are specifications and should be distinguished from test results demonstrating work accuracy or durability in real-world conditions. On the supply-chain front, the report’s assertion that the two hand suppliers have ties to Shanghai is a point of interest. However, the two examples do not establish NVIDIA’s overall level of sourcing dependence or supply volumes. The development platform for a reference robot is also a separate matter from the parts sourcing and supply arrangements for mass-produced robots. A related technical issue was raised in NVIDIA Explains World Models for Interactive Simulation Using Cosmos, which identified adapting models to specific sensors, tasks and data as a challenge. Even with simulation and training methods in place, it remains unclear how much adjustment is needed for sensor data and manipulation characteristics when the physical robot uses a different hand. Issues to watch include the production scale and supply terms for each hand, pricing, interoperability with different bodies, and safety and performance testing in real-world settings. NVIDIA’s announcement said integration of Halos was planned, and that Agile ONE SN was scheduled to become available in early 2027. The materials available so far do not establish the supply arrangements or safety validation results for commercial deployment.

According to NVIDIA’s announcement dated October 21, 2026, Agile ONE SN uses the Agile ONE S body, which is 1.40 meters tall and weighs 40 kilograms. It has 31 degrees of freedom and a payload capacity of 3 kilograms per arm. The two WUJI Hand 2 units each have 20 degrees of freedom, 1kHz force control and tactile sensing. Jetson Thor delivers 2,070 FP4 teraflops of computing performance and has 128GB of unified memory. The company plans to make Agile ONE SN available in early 2027 and said it plans to integrate NVIDIA Halos safety features.