Racing with AI: With Great Speed Comes Great Responsibility
Auto racing has existed ever since the second car was built, and it’s become a staple of the sports landscape, from NASCAR and IndyCar here in the United States, to Formula 1 overseas. While motorsports are team sports, with numerous engineers, mechanics, and other personnel employed to work on each vehicle, the drivers are the stars of the show. They, more than anyone else involved, can spell the difference between victory and defeat. A simple lapse in judgment can lead to a horrifying incident. While motorsports have become incredibly safe in the modern era, these drivers still put life and limb on the line every time they strap in prior to a race.
Some, then, have wondered, “What if we took the driver out of the equation?”. The answer, necessarily, is self-driving race cars. Several racing series have emerged to make that happen. One, known as Roborace, has constructed a fleet of race cars that race themselves, as Craig Scarborough discusses in this article in Motorsport Technology.
First emerging in 2015, Roborace is now on its second generation vehicle and has completed two preliminary seasons, known as Alpha and Beta, with its first proper racing season slated for this year. While past seasons only had one car on track at a time, completing various challenges and avoiding obstacles and even ghost cars, the cars will finally race against each other this year.
As Scarborough explains in his article, the new second-generation car uses an array of instruments, including LIDAR, AI cameras, and GPS. LIDAR is a system of lasers that maps out the surfaces within “view” of its beam. The cameras are paired with the LIDAR system to allow the AI to “read” signs and other markers around the track, such as yellow “caution” flags which warn of trouble ahead. The GPS used is a specialized local version which allows for hyper-accurate location detection. All of these measurements, along with the typical in-car sensors that measure ride height, acceleration, and other factors all inform the AI, which uses the data as it brakes, accelerates, and steers. Also innovative is the ability for human drivers to pilot the car if desired, opening up the possibility of AI-versus-human competition in identical cars.
Some, then, have wondered, “What if we took the driver out of the equation?”. The answer, necessarily, is self-driving race cars. Several racing series have emerged to make that happen. One, known as Roborace, has constructed a fleet of race cars that race themselves, as Craig Scarborough discusses in this article in Motorsport Technology.
First emerging in 2015, Roborace is now on its second generation vehicle and has completed two preliminary seasons, known as Alpha and Beta, with its first proper racing season slated for this year. While past seasons only had one car on track at a time, completing various challenges and avoiding obstacles and even ghost cars, the cars will finally race against each other this year.
 |
| Courtesy of Motorsport Technology |
 |
| Courtesy of the Indy Autonomous Challenge |
That’s unlike the Indy Autonomous Challenge cars, which raced at Indianapolis Motor Speedway this past fall, at speeds of over 150 MPH. According to Nathan Brown of the Indianapolis Star, students from 21 universities organized into nine teams to program open-wheel cars to navigate the massive oval. The event was full of missteps and problems, with one team clocking an average lap speed of 85 MPH, and several others suffering spins and crashes as their AI-piloted vehicles inexplicably speared off track. The winning average speed was over 135 MPH over two laps, a pace reached by TUM Autonomous Motorsport. While the event shows that we still have a long way to go in terms of autonomous motorsports, it also shows how far the field has come. Brown describes off-road racing events less than two decades ago where incredibly few cars finished. While the teams at Indy had issues, they were leagues beyond what was thought possible just a few short years ago.
The role of AI in motorsport extends beyond the real world, too. In this Ars Technica article by Jonathan Gitlin, he describes an AI trained by Sony to compete in their racing game, Gran Turismo. The AI, known as GT Sophy, learned to race through deep reinforcement learning. On its own, training with the same car on the same track, the AI learned quite quickly, outranking the vast majority of players in just a day or two. After a week and a half, it was among the fastest real-life racers. As the AI developed, it began to beat humans consistently in head-to-head competition, finding racing lines even the best human racers hadn’t considered. However, the AI has not been optimized for race conditions quite yet. The intricate racecraft of knowing when to overtake, how to defend, and how aggressive to be against opponents are sure to be its most difficult challenges yet.
So AI seems to have found a place in the motorsports world, both real and virtual. However, the amount of variables in real-life racing are incredibly numerous, as seen by the competitors in the Indy Autonomous Challenge. A video game AI might be able to beat most human competitors, but a real-life AI has an incredibly steep learning curve to crest before it could do the same. It has potential to, as well. While a human driver has feel and instinct, an AI has raw data, which means it knows exactly what the car is doing at all times, something a human can only dream of. As a motorsports fan, I’m interested in this space, but wary all the same. Drivers are what define motorsports as I know it. The personalities and driving style associated with each driver are what keep me coming back. At the same time, as a computer scientist, it’s incredibly fascinating to see how much progress is being made in this field, and exciting to see what AI racing will morph into in the future.
