The Complexity Behind Carnegie Mellon’s Drive Towards Autonomous Racing
What could be a more appropriate backdrop than CES 2022 for “the first head-to-head, high-speed passing competition” involving autonomous racing cars? On January 7th, 2022, the cars lined up at the Las Vegas Motor Speedway, without human drivers, for the second edition of the Indy Autonomous Challenge (IAC).
The race was won by PoliMOVE from Politecnico di Milano and the University of Alabama, taking the $150,000 grand prize, ahead of first stage winner TUM Autonomous Motorsport from Technische Universität München. After all teams raced against the clock in a time trial qualification round to determine the head-to-head race rankings, PoliMOVE faced TUM in the final, not only winning but also recording a top speed of 173 mph. PoliMOVE’s top race speed wasn’t its fastest, however. In December 2021, just days before CES, the team secured the title of “World’s Fastest Autonomous Racecar” with a top speed of 175.96 mph recorded at the Yucca Proving Ground in the Arizona Desert.
Advances in autonomous vehicle technology are such that autonomous racing is now emerging as a serious international motorsport. The inaugural IAC—which even featured a Boston Dynamics robot dog waving the chequered flag—was held at the Indianapolis Motor Speedway in October 2021; other autonomous race series include Roborace, Formula Student Driverless, and the comparatively low key Self Racing Cars.
Start your servers!
Although Carnegie Mellon University (CMU) has a storied history in autonomous technology development, its participation in the relatively new area of autonomous racing is itself relatively new. That’s why Hassan Azmat, Director of Autonomous Racing at Carnegie Mellon Racing (CMR-A), is taking a cautious and realistic approach with a team that was only established in mid-2021.
“Autonomous racing is an in-depth coding area where you have to dive deep into research first to understand how you’re going to develop it,” explains Azmat. “You can’t start from scratch and have something ready within a couple of months. It takes years of development.”
Azmat has his sights set on competing in IAC, but the CMR-A team will first compete in three other competitions: Formula Student Driverless, Roborace, and F1Tenth. All of these involve fully autonomous vehicles, but as its name suggests, the latter event, founded at the University of Pennsylvania in 2016, features 1:10 scale model vehicles.
The cars may be smaller, but lessons learned on a one-tenth scale still translate for full-sized vehicle development, notes Azmat. “Despite the smaller scale, the algorithms and concepts are the same, and you’re still racing other vehicles, so you have to deal with the same complexities of being surrounded by other cars on the track,” he explains.
Much of what CMR-A learns in F1Tenth will help the team as it develops the cars for Formula Student Driverless and Roborace, says Azmat. “The smaller scale platform makes it more manageable logistically and financially. In F1Tenth, you’re using similar sensors and algorithms to those you would use in a full-sized car, but the type of software stack and the level of simulations required for the Indy challenge mean that competing in IAC is something we’ll consider further down the line. For now, we’re developing the car for Formula Student and Roborace.”
Road and track
Like its road-going cousins, the full-sized race car developed by Azmat’s team uses sensor fusion to guide it around the track. “Our car uses lidar, camera, and radar for vision and range,” Azmat explains. “We also have inertial measurement units, or IMUs, which are typically combined with GPS for localization, and wheel sensors for odometer readings.”
Yet while the sensor suite may be similar, race car driving is a different discipline to road car driving—there’s a specific way of cornering, of accelerating and decelerating. But racing is about much more than just the ability to drive really well—it also involves a racing instinct, something that’s difficult to teach, and even harder to code. “That’s where being under Carnegie Mellon Racing really helps us out,” says Azmat. “Everyone on the team is obsessed with Formula 1 (F1), and several CMR-A team members are also involved in the Carnegie Mellon Racing team, further enhancing our expertise.”
In F1, the glamour, brands, and driver personalities create excitement and partisanship; how, then, does Azmat sell the concept of autonomous vehicle racing, especially when all the cars are the same? “In F1, the driver makes the difference, but in student autonomous competitions, the driver is the code and because not everyone’s code is perfect, it’s the code that makes the difference.”
Racing involves risk, from overtaking to pushing the vehicle beyond its optimal operating limits to get the win. At the same time, track racing involves much less unpredictability than road driving. “I think it’s more complicated to program an autonomous road vehicle than an autonomous race car because there are so many more random factors to consider, and unpredictability such as how to respond should an animal suddenly run out,” notes Azmat. “In the Indy Autonomous Challenge head-to-head races, for example, the only unpredictability you have is the behavior of the other car or cars around you. You already know the track layout, and because all teams use the same car, you can even program it to recognize the car shape and detect it accordingly.”
Leading F1 drivers are known to occasionally ignore team instructions, relying instead on their racing instincts to gain time or defend a position. But there’s no instinct or ego in code, says Azmat. “If you call the car back, it will come back, because it will follow exactly what you instruct it to do—unless, of course, you specifically code it to overrule your decisions.”
Although software now plays a major role in F1, it’s the only thing behind the wheel of an autonomous race car, and therefore the difference between success and failure. “Our software team is like the pit stop crew,” chuckles Azmat. “If anything goes wrong on the track or something’s going wrong with the car, our software team is on-hand to figure out the problem and how to fix it without losing much time. They find the issue, decode it and then send the car out again.”
The racing line
Unlike the single-car, time-trial racing format of the inaugural IAC, the Las Vegas race featured the wheel-to-wheel racing of typical motorsports, adding considerable excitement for spectators and team bosses alike. Longer term, could we see races between human drivers and autonomous cars? “I don’t know if that will ever happen, but it’d be cool to watch,” grins Azmat. “The autonomous cars would have the advantage as they’d be driving the perfect lap.”
The question is, in such a race, could experience, tactics, racing instinct, and the willingness to take risks give a human driver an advantage over the AI? “On an oval track, there’s a perfect racing line. Whether it’s speed, control, or the angle at which you turn, all of those have to be perfect to get the fastest time on the track. A human will obviously try to do that every time. But a robot will definitely get it right every time on every track, because it’s been programmed to do it. A human cannot do that.”
That’s in oval track racing, and when racing against the clock. But if human drivers raced autonomous cars in a wheel-to-wheel race, Azmat concedes that a human driver might have the advantage—for now. “If a human-driven car and an autonomous car race separately against the clock, I would say the autonomous car would win. But if there were several human-driven cars racing against several autonomous cars, there’s a strong chance that a human-driven car would win, because the autonomous car would need not only to drive the perfect lap, but do this against many other cars.”
From track to road
Azmat is realistic about the challenges facing the CMR-A team, especially given their inexperience in competition. “Even after preparing for many months, and assuming our vehicle’s autonomous functions are working great, the biggest challenge we face will be when we see how well our code performs with other cars racing against us. It’s also a relatively new field, so that creates additional challenges in obtaining proven testing and development platforms.”
From safety, to powertrain, to electronics and connected vehicle technology, the evolution of road car technology has been accelerated by advances in motorsport. Developers of road-going self-driving technology can draw two key insights from autonomous racing, believes Azmat. “The first is that it’s fast, which provides a rigorous testing platform for developing highway AI technology, as it’s intended for a similar, but less chaotic environment.”
The second insight, he says, is the experience of testing and developing technology under pressure in a thrilling and challenging environment. “When you all have the same car, your code is your driver in a neck-and-neck battle. In road-going technology, there are of course other competing companies, but you’re not racing in an actual exhilarating race.”
But autonomous racing offers one more benefit for road-going autonomous vehicle development, notes Azmat, namely, talent. “Just as you can find the best car engineers building the best race cars for competitions like Le Mans, you can find the best AI tech developers working in the field of autonomous racing.” They might be behind laptop screens rather than steering wheels, but make no mistake: the leading autonomous race car programmers of today could be among the top autonomous road car leaders of tomorrow.