Guotai Haitong has issued a research report noting that on October 21, NVIDIA (NVDA.US) announced an upgrade to its Isaac platform, launching an open-source robot learning workflow that combines ChatGPT Astra, Isaac, and Cosmos. The upgrade supports GR00T 1.7 reinforcement learning, dexterous manipulation training, and Cosmos world model post-training, further expanding the application boundaries of general large models in the embodied intelligence industry. The robotics supply chain is expected to benefit. The main points of the Guotai Haitong report are as follows.
Key changes
First, the introduction of AI Agents to automate the robot R&D process. The Agent primarily handles R&D task understanding, tool invocation, and feedback iteration. It can coordinate simulation tools to complete scenario configuration, inspection, and correction rather than replacing the robot motion model. This can reduce manual scenario construction and repeated debugging, improving R&D efficiency.
Second, the technical pathway from simulation training to real-world deployment has been improved, achieving a platform upgrade. Isaac Lab 3.0 supports robot policy training, reinforcement learning, and multi-GPU workflows. Isaac OS connects AI, sensors, control systems, and robot hardware, supporting real-time operation and simulation-to-reality transfer. Simulation and data resources have added more than 2,000 SimReady simulation assets and the Human 1M human activity dataset, providing more scenarios and data for robot training. This means NVIDIA is integrating models, data, simulation, training, and deployment tools into a more complete development platform, lowering the engineering barriers to robot R&D.
Third, the launch of Agile ONE SN to promote integrated software and hardware. Agile Robots has introduced the Agile ONE SN humanoid robot reference platform based on Isaac GR00T and Jetson Thor, integrating the robot body, edge computing, dexterous hands, and the development toolchain. It provides developers with pre-integrated reference hardware, reducing adaptation work between the body, computing power, and software, and accelerating model validation and skill deployment.
The application boundaries of general large models in the embodied intelligence industry are further expanding. In addition to task planning on the robot side, large models can also serve as R&D Agents participating in simulation environment construction, training experiments, and result validation. The firm believes this upgrade reflects that the physical AI industry is shifting from purely pursuing model capabilities to improving robot R&D efficiency, simulation training capabilities, and real-world deployment capabilities. NVIDIA is attempting to combine these capabilities with its own models, data, simulation, and computing platform to increase the automation level of the robot training process. The most important verification indicators going forward are whether the R&D cycle is shortened, whether the success rate of real tasks improves, and whether these advances can translate into actual deployment, hardware orders, and large-scale shipments.
Guotai Haitong's view
The firm believes that NVIDIA's upgrade of the Isaac platform will help the robotics industry grow. In the short term, simulation, data collection, robot testing, and R&D equipment are expected to benefit from increased development activity. In the medium term, as dexterous manipulation and whole-body control capabilities improve, performance requirements may rise for joint-level actuator units, dexterous hands, force sensors, encoders, and control systems. Looking ahead, if Sim2Real transfer continues to improve and robots meet reliability and cost requirements in real scenarios, this is expected to drive growth in industrial applications and complete machine shipments.
Risk warning
Technological development may fall short of expectations, and end demand may fall short of expectations.
Comments