Nvidia researchers have introduced SONIC, an AI-based universal controller designed to manage humanoid robot movements. The system acts as a foundational motor control model, converting high-level commands into coordinated full-body actions without requiring separate controllers for individual tasks.
The model was trained on a dataset containing over 100 million frames of human motion data. This extensive training allows the system to generalize its knowledge, enabling robots to perform both familiar and novel movements across various scenarios, including object manipulation and locomotion.
SONIC supports multiple input methods, such as virtual reality teleoperation, video-based motion capture, and text-based instructions. Testing in physical simulators confirmed that the controller maintains stable and natural movement patterns even when executing tasks not present in the original training set.
The development team, led by Yuke Zhu, intends to further enhance the system by improving environmental perception for better navigation and obstacle avoidance. Future plans include integrating SONIC with the Nvidia Isaac GR00T platform to combine high-level reasoning with motor control.
