DribbleBot, MIT's Soccer-Playing Robot Pooch, Can Handle Almost Any Terrain
With a pair of NVIDIA Jetson Xavier NX modules on board, DribbleBot can recover from a spill without losing track of the ball.
Researchers from the Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL) have given a quadrupedal robot some impressive skills on the soccer pitch — which give it the ability to dribble a ball over a range of terrains with smooth recovery from stumbles.
"Past approaches simplify the dribbling problem, making a modeling assumption of flat, hard ground. The motion is also designed to be more static; the robot isn’t trying to run and manipulate the ball simultaneously," says Yandong Ji, first author on the paper. "That's where more difficult dynamics enter the control problem. We tackled this by extending recent advances that have enabled better outdoor locomotion into this compound task which combines aspects of locomotion and dexterous manipulation together."
Dubbed the DribbleBot — somewhat clumsily backronymed into "Dexterous Ball Manipulation with a Legged Robot" — the resulting robot is designed to be able to navigate a wide range of terrain, from stepped hard surfaces to sandy ground, while keeping control of a soccer ball. "We overcome critical challenges of accounting for variable ball motion dynamics on different terrains and perceiving the ball," the team explains, "using body-mounted cameras under the constraints of onboard computing."
The robot itself is a modified Unitree Go1 quadruped with two 210° field-of-view (FOV) fisheye cameras, one facing forward and the other down at the ground, feeding into a pair of NVIDIA Jetson Xavier NX modules. These high-performance embedded computers process the full-resolution video through the YOLOv7 object detection network, sending ball location estimates to a third computer running the locomotion and manipulation policy model.
This isn't the first time we've seen a quadrupedal robot playing soccer. Back in October last year a team from the University of California at Berkeley gave an MIT Mini Cheetah some impressive goalkeeping skills — developing a model which allowed it to predict and block a shot on goal with near-90 per cent accuracy.
There's more to the project than just showing off ball skills, though. "If you look around today, most robots are wheeled. But imagine that there's a disaster scenario, flooding, or an earthquake, and we want robots to aid humans in the search-and-rescue process. We need the machines to go over terrains that aren't flat, and wheeled robots can't traverse those landscapes," explains Pulkit Agrawal, MIT professor and director of the Improbable AI Lab where the research took place.
"The whole point of studying legged robots is to go terrains outside the reach of current robotic systems. 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."
More information on the project, including a preprint of the paper under open-access terms, is available on the DribbleBot website; the release of the project's training code, pre-trained models, and a sim-to-real deployment toolkit have been promised, but at the time of writing had not yet been published.