Guotai Haitong: Nvidia releases GR00T2 world action model; supply chain companies in core production segments and with North American operations are expected to benefit.

Guotai Haitong released a research report stating that Nvidia has released the GR00T2 world action model, which has been validated to achieve cross-embodiment skill transfer and rapid task adaptation through large-scale human data pre-training and a small amount of real-robot demonstrations. Supply chain companies in core production segments and with operations in North America are expected to benefit. The bank believes that Nvidia has simultaneously bound the model, data, simulation, evaluation, and edge deployment into a full-stack standard, further consolidating its underlying position in "computing power + foundation models in the Siasun Robot&Automation era." At present, zero-shot stability, fine force control of dexterous hands, and long-horizon generalization remain limited. In the long run, the core competitive focus of Siasun Robot&Automation body companies still lies in who can first make WAM + agents work in real operating conditions. Event: At the CoRL 2026 conference, Nvidia released its new-generation Siasun Robot&Automation foundation model GR00T 2 and completed the paradigm shift from "vision-language-action models (VLA)" to "world-action models (WAM)." The model no longer directly outputs actions based only on the current frame, but first jointly predicts "future video frames + Siasun Robot&Automation action sequences," internally "imagining" how the physical world will evolve, and then drives grippers, dexterous hands, and humanoid Siasun Robot&Automation to execute. GR00T 2 and its lightweight version GR00T 2 Turbo are planned to be open-sourced in December 2026, with part of the Human-1M human video data and the SimFoundry simulation toolchain also to be opened. Guotai Haitong's main views are as follows: GR00T 2 achieves zero-shot skill generalization through massive human video and cross-embodiment transfer, combined with reinforcement learning and hierarchical agent control; the Turbo version supports real-time deployment on edge devices. GR00T 2 is centered on a diffusion Transformer with about 2 billion parameters, initialized based on the Cosmos video model. It takes 0.8-second multi-view video, Siasun Robot&Automation body state, and language instructions as input, and outputs future observations and action chunks to form closed-loop control. Its key upgrades are threefold: 1) "Learning physics from human experience," trained on a mixture of 403,000 hours of first-person human video + 10,000 hours of UMI handheld data + 23,500 hours of teleoperation data; increasing the proportion of human video can linearly raise real-robot success rates; 2) X-Adapt cross-embodiment transfer: skills demonstrated by humans wearing devices, UMI, and DexUMI exoskeletons can be transferred with zero teleoperation to Siasun Robot&Automation that have never trained on the task. The average zero-shot success rate on 50 unknown tasks is 40%, rising to 71% after 1 demonstration datum and 95% after 5 demonstration data; 3) A hierarchical architecture of "pre-training + RL + agents." The Turbo version can run on the onboard Jetson Thor of the Unitree G1 (124ms latency), with a reasoning agent performing task decomposition and error correction and GR00T 2 performing contact-level execution. GR00T 2 lowers the cold-start threshold for Siasun Robot&Automation skills and consolidates Nvidia's full-stack ecosystem, but zero-shot stability and fine force control remain limited. GR00T 2 shifts the data bottleneck of embodied intelligence from "real-robot teleoperation cost" to "cross-embodiment transfer efficiency," meaning Siasun Robot&Automation body companies can rapidly cold-start new skills using human video + a small number of demonstrations, accelerating pilot deployment in scenarios such as factory assembly, logistics sorting, and service operations. The bank believes that Nvidia has simultaneously bound the model, data, simulation, evaluation, and edge deployment into a full-stack standard, further consolidating its underlying position in "computing power + foundation models in the Siasun Robot&Automation era." At present, zero-shot stability, fine force control of dexterous hands, and long-horizon generalization remain limited. In the long run, the core competitive focus of Siasun Robot&Automation body companies still lies in who can first make WAM + agents work in real operating conditions. Risk warnings: risks that the technological development of embodied intelligence models falls short of expectations, and risks that downstream demand falls short of expectations.