The editor of Downcodes learned that the NVIDIA research team has made a major breakthrough in the field of robot control. The neural network system HOVER developed by it has achieved efficient control of humanoid robots with extremely low parameters, and its performance exceeds that of specially designed control systems. The HOVER system can handle complex robot motion control with only 1.5 million parameters, which is in sharp contrast to large language models that often have hundreds of billions of parameters, reflecting its exquisite design.
This HOVER system, which requires only 1.5 million parameters, can handle complex robot motion control. In contrast, common large language models often require hundreds of billions of parameters. This amazing parameter efficiency highlights the sophistication of the system design.
HOVER's training is conducted in NVIDIA's Isaac simulation environment, which can accelerate the robot's movements 10,000 times. Nvidia researcher Jim Fan revealed that this means that a year's training in the virtual space can be completed in only 50 minutes of computing on a GPU.
A highlight of the system is its excellent adaptability. It can be directly transferred from a simulation environment to a real robot without additional tuning, and supports multiple input methods: head and hand movements can be tracked through XR devices such as Apple Vision Pro, and full-body position data can be obtained through motion capture or RGB cameras. , joint angles are collected through the exoskeleton, and can even be controlled using a standard gamepad.
Even more surprising, HOVER performed better on every control method than systems developed specifically for a single input method. Lead author Tairan He speculates that this may stem from the system's deep understanding of physical concepts such as balance and precise limb control, allowing it to transfer knowledge between different control methods.
The system is developed based on the open source H2O & OmniH2O project and can control any humanoid robot that can run in the Isaac simulator. Currently, NVIDIA has disclosed examples and code on GitHub, bringing new possibilities to the field of robotics research and development.
The breakthrough progress of NVIDIA's HOVER system demonstrates the huge potential of artificial intelligence in the field of robot control. Its efficiency, adaptability and ease of use have opened up new directions for future robot research and applications. The editor of Downcodes believes that this will promote the maturity and popularization of robotics technology faster.