Cathay Pacific Haitong: Nvidia releases GR00T2 world action model, core production links and layout, North American supply chain companies are expected to benefit

The Zhitong Finance App learned that Cathay Pacific Haitong released a research report saying that Nvidia released the GR00T2 world motion model to verify that cross-body skill migration and rapid task adaptation can be achieved through large-scale human data pre-training and small-scale real machine demonstrations. Supply chain companies with core production links and layouts in North America are expected to benefit. The bank believes that Nvidia simultaneously binds models, data, simulation, evaluation, and edge deployment into full-stack standards, further consolidating the underlying card position of “computing power+basic model in the robot era.” Currently, zero-sample stability, precise force control by dexterous hands, and long-term generalization are still limited. In the long run, the core competitive focus of robot companies is still who will first run the WAM+ intelligent device through real working conditions.

Event: At the CoRL 2026 conference, Nvidia unveiled GR00T 2, a next-generation robot basic model, and completed a paradigm shift from a “vision-language-action model (VLA)” to a “world-motion model (WAM)”. The model no longer only directly outputs actions based on the current image, but first jointly predicts “future video frames+robot motion sequences”, internally “imagines” how the physical world will evolve, and then drives grippers, dexterous hands, and humanoid robots to execute. The GR00T 2 and the lightweight GR00T 2 Turbo are scheduled to be open sourced in December 2026 to simultaneously open part of the Human-1M human video data and SimFoundry simulation toolchain.

Cathay Pacific Haitong's main views are as follows:

GR00T 2 achieves zero-sample skill generalization through massive human video and cross-morphological migration. Combined with reinforcement learning and hierarchical agent control, the Turbo version supports real-time deployment of edge devices.

The GR00T 2 core is an approximately 2 billion parameter diffusion transformer. It is initialized based on the Cosmos video model, inputs 0.8 seconds of multi-view video, robot body state and language instructions, and outputs future observation and action blocks to form closed loop control. There are three key upgrades:

1) “Using human empirical physics”, using 403,000 hours of human-first perspective video, 10,000 hours of UMI handheld data, and 23,500 hours of remote operation data mixed training, the proportion of human video can be increased linearly to increase the success rate of real machines;

2) X-Adapt migrates across morphology, and the skills demonstrated by human wearables, UMI, and DeXumi exoskeletons can be transferred from zero remote operation to robots that have not experienced this task. The average success rate of zero samples of 50 unknown tasks is 40%, the success rate reaches 71% after 1 demonstration data, and the success rate reaches 95% after 5 demonstration data;

3) “Pre-training+RL+ Agent” hierarchical architecture. The Turbo version can run on the Yushu G1 Jetson Thor (124ms delay). The inference agent performs task disassembly and error correction, and GR00T 2 performs contact-level execution.

GR00T 2 lowers the cold start threshold for robot skills and strengthens Nvidia's full-stack ecosystem, but zero-sample stability and fine force control are still limited.

The GR00T 2 shifts the intelligent data bottleneck from “real machine remote operation costs” to “cross-body migration efficiency,” which means that robot companies can use human videos+a small amount of demonstrations to quickly cold start new skills to accelerate the implementation of pilot projects in scenarios such as factory assembly, logistics sorting, and service operation. The bank believes that Nvidia simultaneously binds models, data, simulation, evaluation, and edge deployment into full-stack standards, further consolidating the underlying card position of “computing power+basic model in the robot era.” Currently, zero-sample stability, precise force control by dexterous hands, and long-term generalization are still limited. In the long run, the core competitive focus of robot companies is still who will first run the WAM+ intelligent device through real working conditions.

Risk warning: The development of embedded intelligent model technology falls short of the expected risk, and downstream demand falls short of the expected risk.