Addressing the current pain points in hardware and software integration for humanoid robots, Spencer Huang, Director of Robotics Products at NVIDIA, explained the new approach proposed by NVIDIA.Isaac GR00T Reference DesignWith the open source omnipotent world modelCosmos 3This not only further lowers the barrier to robot development, but also reconstructs the operational process of robots entering the physical world through the logic of "think first, then act".
Hardware Integration Challenges and the Isaac GR00T Reference Design
One of the biggest challenges facing robot development today is not simply software computation, but the underlying hardware integration problem of "making robots move smoothly." Huang Sheng-bin pointed out that different robot chassis and dexterous hand movements often require the team to spend several months integrating and tuning a large amount of firmware and drive layers.
To address this issue, NVIDIA launched the open-source Isaac GR00T reference design and collaborated directly with Unitree Robotics on a humanoid robot chassis, as well as with Sharpa Wave on a robotic arm with five-finger tactile feedback and dexterity. By providing a standardized hardware and software stack architecture, pre-tested drivers, and middleware, developers can directly configure computing resources on the robot itself. This allows researchers and developers to skip the tedious hardware tuning and debugging phases and focus their efforts on upper-level skill training and data generation, significantly shortening the time required to deploy AI models to physical devices.
Huang Shengbin also explained that the operation of robots usually requires the use of hands to simulate human actions such as grasping and carrying. Different arm designs are required depending on the deployment and application needs. For example, simply grasping objects may only require a simple two-finger arm, but for situations that require gentler force or accurate operation, the robot's hand movements become more important. However, training often takes time. Therefore, NVIDIA hopes to provide reference designs that can be applied immediately in the early stages so that more robots can be deployed in the working environment in a shorter time.
Cosmos 3 World Model: Complementarity and Evolution with Omniverse
At the software and model architecture level, NVIDIA's open-source, all-encompassing world model, Cosmos 3, has become a key core in reshaping the development process of entity AI. When discussing the positioning of Cosmos 3, it is essential to clarify its differences from existing Omniverse platforms (such as Isaac Sim):
• Omniverse (traditional simulator):It boasts extremely high physical realism and accurate ground truth, but its drawback is that it requires a large number of digital assets and environments to be manually built, and its scalability is limited by the speed of pre-build.
• Cosmos 3 (Neural Network Simulator):It belongs to the category of models that "learn" physical dynamics from a large amount of pre-trained data. It can provide a high degree of environmental diversity and flexibility without the need for manual environment construction, and realize the ability to predict and evaluate strategies by "thinking first and then acting".
Huang Sheng-bin emphasized that Cosmos and Omniverse are not in competition, but rather complementary technologies that grow together. As the physical models improve, Cosmos can handle highly diverse and large-scale data generation and policy testing; while Omniverse continues to serve as the highest-precision physical verification environment.
The combination of these two will effectively solve the huge bottleneck in the generation of training data for current World Action Models.
Four Applications of Robots in Real-World Scenarios
Regarding the practical application scenarios and commercialization of robots, in addition to the relatively mature industrial manufacturing and warehousing logistics sectors, Huang Shengbin also pointed out several key areas with great potential for the future:
• Data Centers:With the explosive growth in AI computing demands, data centers are becoming increasingly large and facing a severe shortage of basic maintenance personnel. Introducing robots for data center inspections and even hardware replacements could significantly improve the operational efficiency of data centers.
• Medical institutions (Hospitals):Currently, up to 70% of clinical nursing staff spend their time on non-clinical administrative and logistical tasks (such as moving supplies and guiding patients). Introducing assistive robots with basic navigation and material handling capabilities will effectively free up medical resources and improve the quality of patient care.
• Smart Home (Home Robotics):This is the area with the greatest environmental variables (such as pet toys scattered all over the floor). Huang Shengbin cited LG's exhibit at CES 2026 earlier this year...humanoid robot CLOiDFor example, it is emphasized that the design of home robots cannot only consider the software; the size, weight, and speed of the hardware itself directly involve extremely high personal safety risks.
Building a robot-friendly environment
On the other hand, Huang Shengbin also believes that building a robot-friendly environment is also relatively important. He anticipates that instead of requiring robots to have extremely precise and expensive robotic arms to mimic the human action of opening a traditional refrigerator, it would be better to create a "robot-friendly environment." For example, through IoT connectivity, when a robot approaches a refrigerator, it can automatically open the refrigerator door after determining its intention, allowing the robot to directly take items from the refrigerator, saving it the step of manually opening the refrigerator first, thereby speeding up the completion of work tasks.
Perhaps a more pragmatic approach to accelerating the adoption of home robots is to enable the robots to work together with their environment, rather than simply pursuing higher precision.
