Humanoid robot learns to waltz by mirroring people’s movements

An AI that helps humanoid robots mirror a person’s movements could allow robots to walk, dance, and fight in more convincingly human ways.

The most agile and fluid robotic movements, such as Boston Dynamics’ impressive displays of robot acrobatics, usually follow narrow, pre-programmed sequences. Teaching robots to perform a wide range of human activities is still difficult.

To overcome this obstacle, xuanbin peng at the University of California, San Diego, and his colleagues have developed an artificial intelligence system called ExBody2 that lets robots copy and smoothly perform many different human activities in more lifelike ways.

Peng and his team first created a database of actions that a humanoid robot might be able to perform, ranging from simple movements like standing or walking to more complex maneuvers, like tricky dance moves. The database contained motion capture recordings of hundreds of human volunteers collected in previous research projects.

“Since humanoid robots share a similar physical structure with us, it makes sense to take advantage of the large amounts of human motion data already available,” says Peng. “By learning to imitate these types of movements, the robot can instantly understand a variety of human-like behaviors. This means that anything humans can do, robots can potentially learn.”

To teach a simulated humanoid robot how to walk, Peng and his team used reinforcement learning, where an AI is given an example of what a successful movement involves and then tasked with figuring out how to move. That’s how to do it through trial and error. They first learned ExBody2 with full access to all the data of this virtual robot, such as the coordinates of each joint, so that it could mimic human actions as closely as possible. Then, they learned it from these activities, but only using data it had access to in the real world, such as measurements of inertia or motion from sensors on the actual robot’s body.

After being trained on the database, ExBody2 was put in control of two different commercial humanoid robots. It was able to easily string together simple movements, such as walking in a straight line and bending, as well as performing complex moves such as following a 40-second dance routine, punching, and twerking with a human.

“Humanoid robots work best when they coordinate all their limbs and joints together,” Peng says. “Many tasks and motions require the arms, legs, and torso to work together, and whole-body coordination greatly expands the range of the robot’s capabilities.”