Humanoid robots can now be taught to walk, run, crawl, dance and manipulate objects using a single Nvidia foundation model.

Nvidia's SONIC — short for "supersizing motion tracking for natural humanoid control" — is designed to give humanoid robots whole-body control, enabling them to coordinate their joints, maintain balance and adapt their movements in real time.

The open-source lightweight foundation model is now publicly available, with Nvidia releasing a checkpoint in July for applications including teleoperation and vision-language-action (VLA) driven control. The research behind SONIC was also published this month in Science Robotics as part of Nvidia's push to bring the model to more engineers.

A different approach to robot control

While language and vision models have rapidly expanded to incorporate billions of parameters trained on countless datasets, the models used to control humanoid movement have remained relatively small — a handful of GPUs tuned to a limited set of behaviors. Adding a new skill or movement to a robot's repertoire has typically required building an entirely new controller, which helps explain why humanoid systems have been slower to generalize across tasks than software-based AI models.

SONIC is designed to change that. The model draws on more than 100 million motion-capture frames — representing around 700 hours of human movement — to create a single robot training model.

"Whole body control requires every joint to coordinate while maintaining balance, handling contacts and adapting to changing motion goals in real time," Yuke Zhu, director and distinguished research scientist at Nvidia, told AI Business.

How it works

SONIC has been demonstrated across three dimensions: model size, training data and compute, with training models ranging from 1.2 million to 42 million parameters. The result is a controller that can track a range of movements while also adapting to movements it has not encountered during training.

For Zhu, that represents a change in how humanoid robots can be programmed.

"Instead of hand-designing controllers for individual skills, SONIC learns a general motion foundation through large-scale motion tracking," he said. "The same policy can be driven by VR, video, or VLA models and generalizes to behavior beyond those seen during training."

That flexibility could become increasingly important as humanoid robots move from controlled demonstrations toward real-world, unpredictable environments, where coordination, balance and recovery matter as much as the target task itself.

In tests, the team showed SONIC performing a range of tasks, including picking up and placing a drill in a box, dropping a soda can into the trash, and handling objects such as a carrot, a sponge and an apple. The robot was also shown performing kung-fu and crawling on the floor while copying a human demonstrator in real time.

Remaining challenges

Despite the progress, SONIC is still a research system, and some of the hardest problems in humanoid control remain unresolved. Zhu said contact-rich interactions and highly constrained movements are particularly challenging, especially where there is a gap between simulation and real-world performance.

"Nvidia is improving these through richer training data, better simulation and domain randomization, while some limitations ultimately depend on the robot's physical capabilities," he said.

The bigger test will be whether developers can translate the design from controlled demonstrations to physical deployments, where repeatability and safety become central requirements rather than secondary considerations.

"The next step is demonstrating long-term robustness, reliability and safety across diverse real-world environments," Zhu said. "This requires broader validation, improved sim-to-real transfer, and production-grade safety systems integrated across the robotics stack."

SONIC within Nvidia's framework

SONIC is part of Nvidia's wider campaign to build the underlying technology stack for physical AI. The chipmaker has increasingly positioned robotics as an extension of its AI platform, with its Isaac and Cosmos ecosystems spanning robot foundation models, simulation, synthetic data and tools for deploying AI models on physical machines.

While Nvidia's Isaac GR00T platform is designed to interpret instructions and determine what a robot should do, SONIC translates those intentions into coordinated physical movement.

"SONIC is the motion foundation that converts high-level intent into coordinated whole-body Motion," Zhu said. "It connects simulation and foundation models, including VLA models, to real-time robot control, providing a reusable whole-body policy rather than requiring a new low-level controller for every task."