Driverless racing is real, terrible, and strangely exciting

ABU DHABI—We live in a weird time for autonomous vehicles. Ambitions come and go, but genuinely autonomous cars are further off than solid-state vehicle batteries. Part of the problem with developing autonomous cars is that teaching road cars to take risks is unacceptable.

A race track, though, is a decent place to potentially crash a car. You can take risks there, with every brutal crunch becoming a learning exercise. (You’d be hard-pressed to find a top racing driver without a few wrecks smoldering in their junior career records.)

That's why 10,000 people descended on the Yas Marina race track in Abu Dhabi to watch the first four-car driverless race.

Test lab

The organizers of the Abu Dhabi Autonomous Racing League (A2RL) event didn’t brief me on what to expect, so I wasn't sure if we would see much car movement. Not because the project was likely to fail—it certainly had a lot of hardware and software engineering behind it, not to mention plenty of money. But creating a high-speed, high-maneuverability vehicle that makes its own choices is an immense challenge.

Just running a Super Formula car—the chassis modified for the series—is a big task for any race team, even with an expert driver in the cockpit. I was ready to be impressed if teams got out of the pit lane without the engine stalling.

But the cars did run. Lap times weren't close to those of a human driver or competitive across the field, but the cars did repeatedly negotiate the track. Not every car was able to do quick laps, but the ones that did looked like actual race cars being driven on a race track. Even the size of the crashes showed that the teams were finding the confidence to begin pushing limits.

Each of these Dallara Super Formula cars has been modified by its team to operate without a human driver onboard or in control.

Credit: A2RL

Is it the future of motorsport? Probably not. But it was an interesting test lab. After a year of development, six weeks of code-jam crunch, 14 days of practice, and one event, teams are going home with suitcases full of data and lessons they can use next year.

The track and the cars

A2RL is one of three competitions being run by Aspire, the "technology transition pillar" of Abu Dhabi's Advanced Technology Research Council.

Yas is an artificial island built as a leisure attraction, housing theme parks and hotels alongside the circuit, with an influencer photo opportunity around every corner. The island was the focus of the Emirate restyling itself for tourism, and its facilities now play secondary host to another image makeover as a technology hub. An F1 track is now finding a second use as a testing lab, and it's probably the only track in the region that could afford the kind of excess that two weeks of round-the-clock, floodlit, robotic testing represents.

Although the early ambition was to use Formula 1 cars to reflect Yas Marina's purpose as a circuit, the cost compared to a Super Formula car was absurd. Plus, it would have required eight identical F1 chassis. Even in the days of unrestricted F1 budgets, few teams could afford that many chassis in a season.

So Aspire’s Technology Innovation Institute (TII) went to the manufacturer Dallara, which supplies almost every high-level single-seater chassis, including parts of some F1 cars, but also every IndyCar, Super Formula, Formula E, Formula 2, and Formula 3 car, plus a whole array of endurance prototypes. Dallara was also involved in the 2021 Indy Autonomous Challenge via the IndyNXT chassis.

TII in Abu Dhabi was also involved in the Indy Autonomous Challenge as part of a university’s team, so it got to see how the cars had been rapidly adapted to accommodate a robotic “driver.”

The computer that controls the driving and interprets the sensor stack, situated in the cockpit—almost like a human driver.

Hazel Southwell

The computer that controls the driving and interprets the sensor stack, situated in the cockpit—almost like a human driver. Hazel Southwell

The Meccanica42 actuators that operate throttle, brake, and steering onboard the adjusted SF23 chassis.

Hazel Southwell

The Meccanica42 actuators that operate throttle, brake, and steering onboard the adjusted SF23 chassis. Hazel Southwell

The computer that controls the driving and interprets the sensor stack, situated in the cockpit—almost like a human driver. Hazel Southwell

The Meccanica42 actuators that operate throttle, brake, and steering onboard the adjusted SF23 chassis. Hazel Southwell

L-R: The robotic array that sits lower in the car's cockpit for the actuators to operate the car, and the computer that sits above it for maximum ventilation.

Hazel Southwell

L-R: The robotic array that sits lower in the car's cockpit for the actuators to operate the car, and the computer that sits above it for maximum ventilation. Hazel Southwell

A look at one of the car's sensor pods.

A2RL

A look at one of the car's sensor pods. A2RL

L-R: The robotic array that sits lower in the car's cockpit for the actuators to operate the car, and the computer that sits above it for maximum ventilation. Hazel Southwell

A look at one of the car's sensor pods. A2RL

TII’s Najwa Aaraj then designed a sensor array made of three lidars, four radars, and five cameras, whose data was fed through a computer separate from the car’s onboard electronic control unit. The decision to use a mechanically operated robotic array complicated the job, as it required highly responsive movements to the brake, acceleration, and steering controls, not just commands-by-wire. It did mean, however, that the driving computer could be wholly discrete from the car’s systems, just as a human driver is.

The sensor array was coupled to the robotics (Meccanica42 actuators in a Danisi-engineered system) via the computer, and the whole platform was integrated into the car by a team led by TII's Giovanni Pau, who oversaw the engineering of the whole project. Pau led the team in developing TII’s car until a level of reliability and a base-functioning platform (hardware and software) were built.

Aspire then recruited teams and started the logistically grueling process of arranging eight groups of experts into an intensive, six-week crunch for programming, followed by two weeks of time with the cars on track. Although the championship took responsibility for maintaining the cars (and rebuilding any of them after a crash), every element of what made them run (other than the base software) was up to the teams to develop. No remote control was allowed—just a kill switch.

And as Pau told me, no one wants to hit the kill switch.

Is it artificial intelligence?

No. The cars are also only limitedly autonomous; they run a very clever statistical program that processes inputs extremely quickly (this is arguably how the human brain works). Its capabilities are very impressive and there’s obvious capacity for it to go further, but it’s not an AI. Added capabilities came from programming, not a car putting its experience into practice.

What’s the point of driverless racing?

Ask a motorsport devotee why they care about racing and you'll get a few different answers. The heroism, the iconic (sometimes bitter) rivalries, the high levels of driver performance, the melding of human and machine, the roaring noise.

A2RL takes away one key element: the drivers.

But for at least a century or so, top-level motorsport hasn’t been about one person piloting a machine on their own. In fact, much of the compelling nature of something like Drive To Survive is specifically due to inter-team rivalries.

If you're purely interested in drivers as the protagonists of motorsport, that’s absolutely fine. Sports are about competition, and connecting with the humans behind the wheel can be powerful. It’s why spec series remain compelling, with every driver in (theoretically) equal machinery and a chance to win.

All motorsport involves a mix of sporting competition, entertainment, and technology development. A2RL definitely has the last of those in full effect.

Credit: A2RL

But motorsport isn’t just about onboard cameras showing the frantic steering-wheel-scrabble of adjustments around a Monaco qualifying lap. Even in a spec series, it’s a competition between teams of people working on cars to find advantages.

And the kind of technologies developed at Yas Marina could end up in road cars. The ability to stop or back out of situations where the risk is unacceptable, assistive grip in difficult conditions, the ability to predict and pre-empt a loss of control—these are the sorts of driver assist functions we could see more of.

The tech could even be used in autonomous armored vehicles—one team told me candidly that it sees defense as a possible area of development. A less morally ambiguous use case is the potential for maritime and aviation collision avoidance, something the logistics industry is very interested in.

The contenders

A2RL had one of the key characteristics of any motorsport: It was deeply unfair. Three teams arrived with a wealth of expertise in autonomous racing: TUM, Unimore, and Polimove.

Polimove beat TUM and Unimore to the Indy Autonomous Challenge prize, and TUM and Polimove were involved in Roborace, an early forerunner to A2RL. Unimore also has a direct partnership with TII via its Indy Autonomous Challenge bid.

Another team, Constructor, from a private university of the same name in Bremen, Germany, won Roborace’s Season Beta. Although the series never progressed beyond test events that ranged from endearingly baffling to outright dysfunctional, Constructor beat the house team by several thousand points.

Unsurprisingly, these four teams were the fastest and best-resourced. They had frameworks for data analysis that other teams were still developing, so they could take more risks. The remaining teams—which were mostly based in technical institutions and came at the challenge from an academic background—were the underdogs.

Racing teams normally focus on cooling the rear of the car, where the engine and brakes get hottest, but teams at A2RL instead protected the sensors and wiring at the front of the car from the midday Abu Dhabi sun.

Credit: Hazel Southwell

There weren't any commercial autonomous vehicle companies competing in A2RL, though. These companies' parameters are to avoid risk, but by margins equivalent to their processing and sensor capabilities. They absolutely cannot rag it around a dusty, hot circuit at over 150 miles per hour.

After just a six-week crunch of frantic coding, the smaller teams still relied primarily on GPS to allow their cars to compute what they ought to be doing. The more advanced teams could use better risk calculations from their wider sensor arrays.

Approaches between the teams naturally differed. Non-academic American team Code19 admitted that its car largely relied on GPS to understand its positioning on the track, while Polimove turned off GPS to get better perception during fast laps. Every team is still working on integrating all the data from the stack to process it in real time.

The prize money was potentially big, too. The largest single payout was $611,875 for the winner of the multi-car race, but teams taking home less than $100,000—even those that didn't take part in competitive events—shared a pool of $111,250. That's not a lot for eight weeks’ crunch, but it's something.

What was it like?

There’s something pretty breathtaking about walking into a pit lane and seeing a car with no driver launch onto the track.

The cars line up at the end of the pit lane, the light goes green, and off they go. It’s at once an incredibly familiar sight but also completely jarring as the cars just skim along, hitting the end of the pit lane speed limit before going out of sight through the first corner. And not every run was successful; the cars sometimes crashed into a wall or needed to be pulled back from the exit.

First was a pre-qualifying process to establish which teams had a car capable of racing; each was benchmarked on its lap time and the car’s ability to cleanly perform an overtake.

The fastest team in pre-qualifying, Polimove, set a lap time of 1 minute, 57.854 seconds, with a top speed of over 156 mph. That's nothing compared to a Super Formula car’s potential at Yas Marina with a professional human racing driver behind the wheels, but it's not bad, either. The TII-developed car lapped more than 15 seconds quicker, at close to 1:42, showing how much potential remains untapped for the faster teams.

I thought overtaking might be too big a hurdle. Roborace managed an autonomous overtake for the first time during a trial race at Monteblanco in Spain in 2019 when its championship car overtook TUM’s. It was a very pedestrian move, but in autonomous terms, it was monumental—one car calculated the risk of overtaking and figured it was worth it; the other realized it was pointless to defend and risk a collision.

Several teams engaged in overtaking, though some may have gotten a little lucky, as they were still dependent on GPS data. Five teams qualified on the basis of completing overtakes. That meant Code19 got an upset entry into the next stage, a four-way competition of attack/defend duels. This was an entirely separate competition, one that didn't affect the finale, with a small chunk of the total $2.5 million prize pool at stake. Polimove took first place despite a collision with TUM in the first round—the stewards deemed TUM at fault and disqualified its car from the duel.

Not everything went to plan.

Credit: A2RL

Time trials followed, with only one car on track at a time. Polimove also won that competition, and the same top four teams made the cut for the grand finale.

Finally, we got to the race between TUM, Unimore, Polimove, and Constructor. For reasons I can't fully explain, the opening few laps felt unbearably tense. Polimove’s car led the pack in a rolling start after running behind a pace car. A second later, TUM’s car stopped on track.

TUM hadn’t crossed the line yet, and Constructor had to crawl behind it as it stop-started until it finally got enough grip to get going again. Polimove restarted the race for a second time on the start-finish straight, roaring through the first sector of the track before quickly locking brakes and spinning, tires still cold from so many slow laps.

In fairness to the car, it tried to turn around and get going again before the team stopped it from reversing onto the track and into the path of the Constructor car, which promptly stopped on track, causing a full-course caution. This left Unimore and TUM unable to pass the stricken Constructor car, their programming (rightly) forbidding overtaking under such circumstances.

Was watching all that very silly? Absolutely. But it was still exciting to be there while computers tried to work out four-way autonomous racing for the first time.

Going down to a final-lap shootout, the car that started all the chaos—TUM’s—took the checkered flag (shades of the 2021 F1 finale, which also took place at Yas Marina).

Can it race a human?

As part of the event, A2RL organized another first: a human driver versus a computer. Former F1 and current Lamborghini Hypercar driver Daniil Kvyat agreed to be the guinea pig and race alongside the TII car. That’s no small thing, as the week prior had seen Pau frantically planning what margins would be safe for Kvyat to take with the TII car.

Kvyat is a highly skilled driver, so he played cat and mouse with the championship car, lining it up for an overtake and then letting himself slip behind to do it again on the start-finish straight in front of the packed grandstand. He was obviously much faster than the TII car, especially as it was lapping in "high-perception" rather than "high-speed" mode, but messing around with a driverless car going over 100 miles per hour was still pretty gutsy.

The cars are impressive, but it will be a long time until they're competitive against human drivers.

The tech being tested here could drive a stint at Le Mans one day.

Credit: A2RL

When I pitched the idea of a Garage 56 entry with two humans and a driver-swappable computer to Dallara CEO Andrea Pontremoli, his eyes lit up (Garage 56 is a slot at the 24 Hours of Le Mans reserved for cars trying to break technical boundaries).

Dallara did well at Le Mans 2023, Pontremoli was keen to point out, supplying the winner's chassis (Dallara built the Ferrari 499P), plus the bodywork for the fan-favorite Garage 56 NASCAR entry. “With that, they said we couldn’t do it, it wouldn’t be possible—but it was. So you have to imagine this in the future.”

So what now?

This was A2RL’s single event this year. Teams now have time to work on their programming, and the championship has time to take stock and plan for next time.

For instance, would anyone have thought, “Will the race stop if there’s a car on track during a full course yellow?” For a start, the car’s compliance with the rules was on teams, not the championship, so someone could have programmed it to still overtake if the car ahead was stationary. If race teams are known for anything, it's pushing the rules as far as possible in search of an advantage.

Autonomous racing is not quite there yet. But racing is more about going in circles than it is about the endpoint. Who knows what we'll see when next year rolls around?