From Flips to Dance Moves: New AI Trains Humanoid Robots in Record Time

University of California Berkeley researchers have developed an AI system that trains a humanoid robot in more than 129,000 movements in just three hours using a single GPU. 

Known as SuperMimic, the system has the potential to reduce much of the computing resources that would otherwise be necessary for instilling complex manoeuvres in humanoids, be it dance routines or flips and rolls. The technology is detailed in a research paper submitted on October 20, and which will be presented at the 2026 Conference on Robot Learning; it works off a virtual model of the robot in order to speed up the training process. 

Ordinarily, one needs substantial computing power to have a humanoid replicate large numbers of movements. The usual recourse is reinforcement learning, with robots in simulation making repeated attempts and being given feedback on how they did. SuperMimic does not follow that course. It learns what the robot’s body is like and from that makes predictions as to the outcome of various movements, so the control software can be improved in a virtual environment without the need to run costly physics simulations over and over. 

In three hours on an NVIDIA H100 GPU, SuperMimic displayed a better performance score than an extant motion-tracking system for humanoids, the researchers report. To get there, the SONIC system of the earlier days had to be trained for seven days across 128 GPUs. 

Of the 129,514 movements tested, 25 were left uncompleted by SuperMimic after its first three hours. But once the researchers put two GPUs to work and let the training go on for 9.8 hours, the system got through every movement and none of the test’s failure conditions were met. These are results taken from the training data, so they are indicative of the system reproducing what it has been taught as opposed to something wholly new. 

The team was able to port the controller to an actual Unitree G1 and have the physical robot execute a variety of movements with no separate tweaking of the system. All in all, the research indicates that developing advanced humanoid movement systems may require considerably less computing infrastructure than previously demonstrated approaches.  

This might leave sophisticated motion-control work to developers and labs with limited computing resources. 

Then again, getting a robot to replicate thousands of movements is not the same as having it carry out useful tasks on its own. SuperMimic is about whole-body motion tracking, not object recognition or following instructions. And the researchers have yet to determine how well the system stands up to the challenges of an unpredictable workplace over time. 

Nonetheless, the work hints at an important change in the way humanoids are developed. If the demands on the computer can be cut, there is no need to call on big GPU clusters to explore more complicated movements.