NVIDIA trains SONIC for general humanoid motion control

NVIDIA researchers have developed SONIC, a single controller intended to reproduce a broad range of humanoid whole body motions instead of relying on a separate control policy for each skill. The model, described in a peer reviewed Science Robotics paper, was trained on more than 100 million frames of human motion.

According to Tech Xplore, SONIC translates movement targets into coordinated commands across a humanoid robot’s body. The researchers report that the same controller can accept input from VR teleoperation, video based motion and vision language action models without retraining.

One motor model for multiple command sources

The team evaluated SONIC in physics simulation across locomotion and manipulation scenarios, including motions outside its training data. It also tested the controller on a humanoid robot, assessing its reproduction of both learned and unseen reference movements. The resulting motions were described as robust and natural, although Tech Xplore’s account does not provide quantitative tracking results or identify the robot platform used.

SONIC occupies the motor control layer rather than functioning as a complete autonomy system. In demonstrations, the researchers connected it to other components for remote teleoperation and for converting written instructions into robot actions. Language interpretation therefore comes through a higher level VLA model, while SONIC handles the physical execution.

This separation could give developers a reusable motion policy beneath different interfaces and task planners. NVIDIA researcher Yuke Zhu said the approach could support manufacturing, warehousing and logistics applications without requiring engineers to create a new controller for every task.

Real world robustness remains the next test

The evidence reported so far establishes motion breadth more clearly than production readiness. NVIDIA plans to improve SONIC’s awareness of its surroundings, with the aim of reducing collisions and supporting navigation in dynamic settings and on uneven terrain.

The researchers are also working on manipulation involving sustained or complex contact, stronger simulation transfer and greater robustness under real world conditions. These are central requirements for industrial humanoids, where a controller must manage changing loads, imperfect footing and interaction with equipment rather than simply track an isolated reference motion.

NVIDIA’s next planned step is closer integration with Isaac GR00T. The proposed stack would pair GR00T’s higher level reasoning with SONIC’s general whole body control, allowing task instructions to drive a wider repertoire of physical actions.

Source: Tech Xplore