Super Mario Gets New Job: Driving Instructor

FOLLOWING LAST YEAR’S RELEASE of Mario Kart World alongside the Nintendo Switch 2, the video game sold nearly 6 million units in two months, the franchise’s most successful launch yet. Now UMD engineering researchers are using the mustachioed cartoon plumber who races a cart around a track as a model for—wait for it—responsible driving.

Funded by the U.S. Naval Air Warfare Center Aircraft Division, the researchers are using artificial intelligence (AI) to train a computer to “play” Mario Kart on its own—in this case, the inaugural 1992 Super Nintendo version—circling the track as fast as possible while hewing to actual laws of physics and, most importantly, avoiding risk. Eventually, the UMD team hopes to provide regulators with a roadmap to certify AI technologies in autonomous vehicle fleets that are increasingly popping up in cities across the U.S. and around the world.

The research is led by aerospace engineering Associate Professor Mumu Xu, who studies federal safety certification for planes and cars. She recently became interested in the disconnect between long-established vehicle safety norms and the wildly complex algorithms used in autonomous cars and trucks, making it almost impossible for engineers to fully understand the systems’ driving and safety decisions. “With AI, all the old mathematics we use to show that a system is safe no longer hold because no one knows what’s under the hood,” Xu says.

So Xu and her colleagues sought out a meeting point between humans and machines that everyone could understand—and it happened to wear a bushy mustache. In a study published last spring in IEEE Xplore, they demonstrated how Mario Kart can be used to teach an autonomous simulator to avoid collisions through an AI method called deep reinforcement learning.

The 33-year-old Super Mario Kart, with its pixelated landscapes and A-B-A-B button-pushing combos, pales in comparison to the lavish 3D graphics and gyroscopic controllers of today’s version, but its codes are simpler and available to the public. When Xu’s team reconfigured them so a computer could take the wheel, it added a reward system that offered points for whizzing past checkpoints and subtracted them for spinouts and slowdowns. The computer then raced Mario around the track millions of times. At first, Xu says, it “drove stupidly,” hopping up and down and bouncing into walls. But eventually it learned to zip through the course without a dent.

The researchers’ next goal is to ensure that Mario’s maneuvers are mathematically feasible: Can he really carve a turn at 40 mph without running off the road, for example? Those answers will help fine-tune their algorithms in a second Navy-funded study.

The research applies not just to self-driving vehicles but to any consumer product relying on algorithmic software, Xu says.

“For anything that uses AI, whether it’s a medical device or a gadget like Amazon Echo, we hope that their algorithms are safe,” she says. “This is our attempt to figure that out.”