'DribbleBot' can learn to dribble a soccer ball and could save lives

The robot begins by not understanding how to dribble the ball; instead, it is only given positive reinforcement when it succeeds and negative reinforcement when it fails. It thus attempts to determine the order in which its legs should exert forces.

"One aspect of this reinforcement learning approach is that we must design a good reward to facilitate the robot learning a successful dribbling behavior," says MIT Ph.D. student Gabe Margolis, who co-led the work with Yandong Ji, research assistant in the Improbable AI Lab. "Once we've designed that reward, then it's practice time for the robot: In real-time, it's a couple of days, and in the simulator, hundreds of days. Over time it learns to get better and better at manipulating the soccer ball to match the desired velocity," he added.

A robot faces an intriguing new set of challenges when kicking a football around. For instance, a ball's friction and drag interactions with the ground differ from how a robot's legs deal with the same terrain. So, a robot must be able to think about both what it is doing and what it is trying to kick at the same time.

“DribbleBot” improvements aren't just for fun and games; they could also have some real-life benefits for society. 

“If you look around today, most robots are wheeled. But imagine that there’s a disaster scenario, flooding, or an earthquake. We want robots to aid humans in the search-and-rescue process,” Pulkit Agrawal, a CSAIL principal investigator and director of Improbable AI Lab, said. “Our goal in developing algorithms for legged robots is to provide autonomy in challenging and complex terrains that are currently beyond the reach of robotic systems,” he added.