COLLEGE PARK, Md. — A 2D blast from the past is now training the future of what 3D AI vehicles may look like on the roads: Super Mario Kart.
“It’s like our Tamagochis back in the day,” associate professor at the University of Maryland Dr. Mumu Xu said. “We’re trying to provide mathematical guardrails to show that AI systems can be safe.”
Xu studies aerospace engineering and is running the study with her students. They “teach” Mario to drive himself using artificial intelligence, by gamifying the system.
Vehicle safety isn't a game, but researchers are using one to give AI 'guardrails' for self-driven cars
“We give positive points for finishing the lap, negative points for going off the road, negative points for going backwards,” Xu explained. “[Mario] will overtime… will learn what gives him more points and what doesn’t and he eventually learns that if you finish a lap you get the most amount of points.”
It’s all in the name of safety.
“Right now AI is such a black box that there’s no formal way to show that something is safe. And so, it’s a little bit scary if you’re thinking about putting these cars on the roadway because I can’t guarantee that it’s going to perform well every single time,” Xu said.
There were a few road bumps along the way.
“It started out where he was hopping and just wouldn’t move at all. Because he didn’t want to get penalized,” Xu explained. “He ‘reward hacked’ and just basically made loops over and over again to checkpoint one and would come back to the start and come back again.”
But after some trial and error, AI Mario figured out how to finish a lap. Xu says it took six months.
“We train Mario for millions of millions of times,” she explained.
Now that he’s fully on track, they can take the data to figure out which path he takes is the safest, or what needs to change in order to do so.
Notably, Waymo has been seeking state approval in Maryland for clearance for its vehicles to drive in Baltimore.
RELATED: Waymo wants its self-driving cars in Baltimore, Annapolis lawmakers need to give approval first
The company has already been mapping out parts of Baltimore City with human drivers supervising.
The bills are unlikely to pass this session.
The company’s cars have been hitting the roads in 11 other cities, like Austin, TX and Phoenix, AZ. But, in some cases its cars have faced real life obstacles that are more serious than just banana peels or turtle shells.
“Those are all the corner cases that you can’t really hard to predict and so you want to be able to find them before you deploy these vehicles,” Xu said.
Now for the final lap of the game, Dr. Xu says they’re looking to go multi-player.
“One of the things we’re looking at is this notion of transfer learning. It’s that, if you train Mario here, without anybody else. Now we put him in an environment with other players. How’s he going to do?”
