Nvidia Releases SONIC Model to Train Humanoid Robots

Nvidia has made its lightweight foundation model, SONIC (supersizing motion tracking for natural humanoid control), publicly available. Published in Science Robotics, the open-source model allows developers to train humanoid robots to walk, run, crawl, dance, and manipulate objects using a single system. Nvidia previously released a checkpoint in July for teleoperation and vision-language-action driven control.

Traditionally, roboticists had to build entirely new controllers to teach a robot a new skill. SONIC bypasses this limitation by training on more than 100 million motion-capture frames, which represent roughly 700 hours of human movement. The model scales across compute and size dimensions, with training configurations ranging from 1.2 million to 42 million parameters. This allows the controller to track diverse movements and adapt to physical scenarios it did not encounter during its training phase.

For practitioners, SONIC represents a shift from hand-designing individual controllers to utilizing a general motion foundation. In testing, the system successfully guided robots through tasks like picking up and placing a drill in a box, discarding a soda can, and handling delicate objects like carrots, sponges, and apples. The model even enabled robots to mimic a human demonstrator performing kung-fu and crawling in real time.

The model integrates with Nvidia’s broader physical AI ecosystem, including the Isaac and Cosmos platforms. While Isaac GR00T determines what a robot should do, SONIC translates those high-level intentions into coordinated physical joint movements. However, challenges remain regarding contact-rich interactions and the gap between simulation and real-world deployment. Nvidia plans to address these limitations through richer training data, improved simulation, and domain randomization.