Researchers from the University of California, San Diego have designed an AI-enabled robot that can perform a Waltz simply by mirroring the moves of its human partner. As far as we can tell, the robot was even able to pull off the ballroom dance without stepping on its partner’s toes.
To make their dancing robot, the team first designed an AI model trained on human motion capture videos and then integrated it into two bipedal Unitree G1 robots. Using another model, those robots were then able to analyze the motions of humans in front of them and mimic those movements themselves. The result was a humanoid robot able to seamlessly walk, dodge, squat, and dance by copying a human.
[ Related: The 11 weirdest things humans did to robots in 2024 ]
Robotics researchers aren’t shy about having their machines perform over-the-top stunts. From intricate dance dance routines to impressive parkour moves, the internet is awash with robots behaving bizarrely. But most of those examples typically involve pre-programmed, highly choreographed scripts. In this case, researchers wanted to see if they could teach a robot to learn movements more organically.
To do that, they first needed a powerful brain with a good baseline of knowledge. The researchers obtained a database of human motion capture data and used it to build out an AI model they called ExBody2. They trained the model in a simulation environment using reinforcement learning until it had a broad understanding of how to articulate and move many parts of a robot body. They then implemented ExBody2 into a pair of robots that were equipped with another model that sucked in real-world physical data captured by the robot’s cameras. The robot was able to quickly compare the movements it was “seeing” in front of it against its baseline training.
Videos of the robot in action show it pacing outdoors, sidestepping like an athlete (or attempting a foxtrot?), drawing circles with its arms, squatting, and throwing punches. Some of the demonstrations show one of the researchers besides the robot performing the initial movements. After just around a second delay, the robot dutifully repeats the motion. Researchers say this approach is useful because it means the robot is more adaptable than others that require new training sets every time it needs to perform a new movement.
Mirroring robots could speed up training and drive down costs
This isn’t the first example of a copycat robot. Last year, researchers from Stanford used a similar AI model to teach a 5’9 humanoid robot how to box and play rudimentary tennis all by “shadowing” human movements. Like ExBody2, the robot also utilizes an AI trained on motion capture data.
[ Related: Researchers tortured robots to test the limits of human empathy ]
“By mimicking humans, humanoids can potentially tap into the rich repertoire of skills and motion exhibited by humans, offering a promising avenue towards achieving general robot intelligence,” the authors wrote in a paper detailing their results.
This approach to robot movement isn’t just about novelty. Researchers believe mirroring or shadowing could cut down on the need to repeatedly train robots, which could help speed up their development and drive down costs. At the very least, the technology shows a rudimentary robot dance partner is within the realm of possibility.
