A humanoid robot waltzes with help from an AI trained on human motion capture recordings
Xuxin Cheng and Mazeyu Ji
An AI that helps humanoid robots mirror a person’s movement could allow robots to walk, dance and fight in more convincingly human ways.
The most agile and fluid robotic movements, such as Boston Dynamics’s impressive demonstrations of robot acrobatics, are typically narrow, pre-programmed sequences. Teaching robots to perform a wider repertoire of convincingly human movements is still difficult.
To overcome this hurdle, Xuanbin Peng at the University of California, San Diego, and his colleagues have developed an artificial intelligence system called ExBody2, which lets robots copy and smoothly perform many different human movements in more lifelike ways.
Peng and his team first created a database of actions that a humanoid robot might be capable of performing, from simple movements like standing or walking to more complex manoeuvres, such as tricky dance moves. The database included 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 vast amounts of human motion data already available,” says Peng. “By learning to mimic this kind of motion, the robot can quickly pick up a wide variety of human-like behaviours. This means that whatever humans can do, the robot can potentially learn.”
To teach a simulated humanoid robot how to move, Peng and his team used reinforcement learning, where an AI is given an example of what a successful movement consists of and then tasked with figuring out how to do it itself by trial and error. They first had ExBody2 learn with complete access to all the data on this virtual robot, such as coordinates of each joint, so it could mimic the human actions as closely as possible. Then, they had it learn from these movements but only using data it would have access to in the real world, such as measurements of inertia or speed from sensors on a real robot’s body.
After it had trained on the database, ExBody2 was put in control of two different commercial humanoid robots. It was able to smoothly string together simple movements, such as walking in a straight line and crouching, as well as perform trickier moves, such as following a 40-second dance routine, throwing punches and waltzing with a human.
“Humanoid robots work best when they coordinate all their limbs and joints together,” says Peng. “Many tasks and motions require the arms, legs and torso to work together, and full body coordination greatly expands the robot’s range of capabilities.”
Topics:
