A quick look through the Cars Technica back catalog (the carchive, perhaps?) shows that autonomous driving technology and racing technology are both topics we return to quite often. But it has been a while since we covered their intersection—specifically, what’s been going on at Roborace. The series first broke cover at the end of 2015 and then wowed everybody with the Robocar a few months later.
Initially, the idea was for a driverless support series for Formula E. Roborace would supply teams with identical Robocars, and the teams would try to program a better racing AI. However, it’s fair to say that the idea of watching a grid full of AI cars race each other did not meet with universal approval. “We realized that humans are very much part of the storyline of autonomous driving technology. The machines need to learn from humans. What’s it like to take a ride in one as a passenger? These cars have to learn how to fit into a human world. Human and AI cars will share the road,” said Rod Chong, Roborace’s deputy CEO.
So the series has pivoted. Instead of being a driverless championship, Roborace wants to find its place in the future of racing through the combination of man machine. Which, when you think about it, has always been the point of the sport.
Technology + competition + entertainment
A student of the sport, Chong sees the purpose of Roborace as helping to further the development of autonomous driving technology through competition. “Look through the history of the automobile; so many new technologies first appeared on the race track. If the breakthrough didn’t happen at the track, it was popularized and developed in competition,” he explained. But he also knows that a successful race series should balance more than just competition and technology—it also has to entertain.
Up until this point, the series had run some demos at Formula E races, even pitting the DevBots against professional human drivers on two occasions (DevBots are test machines based on an LMP3 sports prototype, still with a cockpit for a human driver but with electric propulsion and all of the sensors and computing hardware from Robocar. (“In both cases the human driver—enthusiast or professional—was faster,” Chong told me.) For this year’s Berlin ePrix, Roborace decided to combine humans and machines in a time attack competition.
“The goal is to showcase where the technology is developing with relation to human drivers in controlled conditions. We wanted to look at the state of autonomous driving solutions from universities and compare them against enthusiast drivers to see where the state of the art is,” Chong said. So the AI for one team would be programmed by researchers from the Technical University of Munich (TUM), with the other coming from the University of Pisa. Each team would also work with a human driver—TUM’s AI would be joined by Acronym designer Errolson Hugh; Pisa’s would be joined by YouTuber Nick Andrew (AR12Gaming).
The format for the competition was quite simple: first the human driver sets a lap, then the AI follows. Whichever team has the shorter average of the two wins. That sounds simple, or it except for the very limited amount of track time available. Each human driver got an hour-long test session in a DevBot during the lead-up to the race—the first time either had driven anything quite as focused. “A day or two before, I sent Rod a text message asking if it was like driving a normal car,” Hugh told me. “Rod texted back ‘absolutely not,’ but by then it was too late to pull out!”
Hugh accurately describes the sensory overload of that first test. “While getting the seat made, I got my first view into how claustrophobic it is and how well those seats work. You’re fused to the car; it’s like wanting to be a cyborg,” he told me. And as he discovered, it’s often a case of “hurry up and wait” for the driver while the mechanics and engineers are busy with their thing. “You’re waiting 15 to 20 minutes to get started—if you’ve never been in that environment before, it’s super distracting, and a lot of adrenalin builds up. But when they say go and you hit the pedal that first time, it all disappears. You’re so focused on all the inputs that all the claustrophobia just vanishes. Then it’s just huge amounts of fun. I think my face hurt from smiling.”
At the Berlin ePrix, the two Roborace teams would be limited to just two 45-minute practice sessions, then the race itself. And most of the practice running would be given over to the AI. You might expect the AI teams to have been working for weeks on programming the cars, but Formula E tracks are temporary courses, and the competitors only get highly accurate lidar scans of the track a few days before. “I got an out lap, the timed lap, and an in lap. It was very intense and a very unique experience—one I won’t get to do again soon, unfortunately. My mindset was ‘don’t die, don’t bin the car.’ Otherwise, it was too much pressure to try to be fully present in the process,” Hugh told me.
Evidently that worked. Hugh set a laptime of 91.54 seconds, and, with the help of his data, the TUM AI was just a tenth of a second slower with a 91.64. “It was fascinating to see how quickly they could get the AI to adapt to the new information. Initially I had a lap time that was almost four seconds faster than my teammate’s. Then they looked at the data from the runs and saw I was braking a lot later than the AI was, so [they] used that data. By day two, the AI was just a tenth slower,” Hugh said.
Man and machine were evidently in less harmony in the other garage; Andrew drove the course in 93.23 seconds, but Pisa’s AI could do no better than 97.49 seconds.
I asked Hugh for his impression of the experience. “It’s a more positive way to look at the whole autonomous driving thing,” he said. “It’s not a case of either-or; there will be AI, there will be humans. The only real way to understand the ramifications of that, which are going to be massive, is to get out in front of them, and this was an extremely eye-opening way to do that. Everyone who was there thought the same way.”
Chong was also pleased with how the race was received. “It was a radical experience for the drivers but constrained by Formula E’s events. The Roborace teams were constantly developing software, and we had to give a fair amount of track time to the universities. But over each 45 minute session, we developed a lot of elements to play on the jumbotrons. It felt like something new was happening, and the feedback was positive. There’s always a level of risk when you’re running a live event—look at the trouble Renault had running a turbo F1 car in 1982. You have to be able to take that on board—if you’re running the very latest version technology, it may not work sometimes,” Chong told me.
As a proof of concept, then, it appears to have worked. Roborace is now preparing for its first season, which will take place in early 2019. CEO (and Audi factory racing driver) Lucas di Grassi recently laid out his vision for Roborace’s “Series Alpha,” and it’s going to require a new car. Like the DevBots, it will be based on an LMP platform (presumably LMP3, given the speeds and budgets involved) and rear-wheel drive, with a motor per wheel. Roborace is setting 1,000kg and 400hp (300kW) as its target. Oh, and there will be a cockpit for a human driver:
Yes, there will be a human professional driver inside the car driving, “teaching” the machine for part of the race. The rest will be taken over by the machine-learning algorithms, or an “AI driver.'” The winner will be the best combination of both. You can think of it as two drivers sharing a car in an endurance race, but in this case, one is the machine itself!
While I’m sure some will remain permanently unenthused by the idea of AI competing in motorsport, my sense is that both di Grassi and Chong understand the world of racing well enough to come up with a format that is suitably entertaining. Motorsport is already a rather broad church, covering everything from speed trials on salt flats and quarter-mile drag races to the twolargest annual sporting events in the world. Surely there’s room for this one.