A2RL raced five driverless cars at Imola. Three broke, and that was useful.

On 5 September five cars with nobody inside them raced each other for twelve laps at Imola. It was the first time a multi-car autonomous race had been run in Europe, and the first time the Abu Dhabi Autonomous Racing League had raced anywhere outside Abu Dhabi.

The race ran to a finish and Kinetiz won it. Three of the five cars broke along the way, and the failures are the reason the afternoon was worth the money.

Why this event mattered

Autonomous driving is demonstrated constantly and tested rarely. A demonstration runs on a route the system knows, at a time the operator picks, with a safety driver in the seat. It proves that the software works when conditions are favourable, which nobody doubted.

Imola removed all three of those comforts at once.

None of the teams had raced there. The circuit is old and hilly, and the organiser chose it. There was no safety driver anywhere. And four other cars were on track making their own decisions at the same time, which is the condition that separates a driving demonstration from traffic.

Every car was an EAV-25, the second-generation A2RL machine built on a Dallara Super Formula chassis. Same sensors, same computer, same engine for all five. What differed was the code.

That parity is the point of the whole exercise. When one car beats another, the reason is somebody's perception, planning or control software, and there is no larger hardware budget to hide behind.

Valentino Jadriško, a senior associate at Constructor Capital, watched from the grandstand. "Imola is where Formula 1 history was made," he told us. "I think I was watching a different kind of history being made."

What happened

TUM, champion in 2024 and 2025, developed a brake problem on the formation lap. A reset from the pit lane did not hold and the car was out before the rolling start.

Unimore took pole, led, and set the fastest lap of the race at one minute forty seconds. PoliMOVE sat within a second of it from the start and recorded the highest speed of the day at 252,3 km/h. Kinetiz and Constructor Racing ran further back.

Then a sensor on the Unimore car failed. The car was blinded, its safety logic took over, and it stopped quickly on the circuit. PoliMOVE, a second behind and closing, had no room to go round. It hit the back of the stationary car and both were out.

Kinetiz inherited the lead and held it to the flag, with Constructor Racing second.

Three failures, three different kinds

The failures are not the same failure, and the difference is what makes the data valuable.

The brakes. What has been published is that TUM had a brake issue and that a reset did not fix it. Nobody outside the team knows yet whether that was hardware, control software, or the two disagreeing.

The sensor. A car that cannot see cannot plan, and Unimore's system reached the correct conclusion about its own situation: stop. The safety logic did what it is there to do.

The following distance. No fault has been reported in PoliMOVE's system. It was a second behind a car travelling at racing speed, and a second is not enough distance to route around an obstacle that appears where clear track used to be.

Put the second and third together and you have the problem that public roads are made of. Unimore's car did the right thing for itself, and any regulator reviewing that decision alone would approve of it. The same decision put a stationary object on the racing line in front of a car that had been lapping within a second of it all race.

Safe behaviour for one vehicle became a hazard for another. That is the difference between single-agent safety and multi-agent safety, and it is the harder of the two.

Why three failures make this a success

A test that surfaces no failures has told you nothing you did not already believe.

Fifteen minutes at Imola produced a mechanical fault, a perception failure with a correct safe stop, and a following-distance problem. All under load, all recorded, all in front of witnesses. Stephane Timpano, chief executive of ASPIRE, which runs the series, described the day as "generating valuable data to accelerate the next stage of R&D". On the evidence that reads as an accurate description of the day.

Simulation does not produce these, and the people building the software say so plainly. Ilia Shimchik, team principal at Constructor Racing and part of Constructor Labs, the arm that sits between Constructor University and industry, put it this way the day before the race.

"Sim to real gap, a very standard problem. Simulating the real world is very hard. Tyre behaviour differs, asphalt differs, temperature, grip. We live with the inefficient simulator we have. Whoever solves that will probably be next."

A stationary car appearing mid-corner because another team's sensor died is in nobody's test suite until it happens once.

The commercial value of the afternoon is the failure log, and it belongs to five teams and one organiser who now know something they did not know on 4 September.

Who pays for this, and what they are buying

A2RL is run by ASPIRE, which sits under Abu Dhabi's Advanced Technology Research Council. The prize money is large and the racing is the visible part of a research programme.

What that programme accumulates is not lap times. Three of the five teams on the grid carry a university's name: TUM, Unimore and Constructor. The asset being built is a body of software, a set of failure cases and a group of engineers who have now debugged autonomy at racing speed in public.

Some of it is already moving. "We have a partnership with Drako Motors, an American startup making high-performance electric cars," Shimchik said. "We help build emergency obstacle avoidance algorithms on their vehicle. We understand how to control a car at its limit."

That is the transfer path in one sentence: a team that spends its year finding the edge of what a car can do sells that knowledge to somebody who needs a car to stay inside it.

For anyone buying autonomy, Imola is a useful shopping list of questions. What does your system do when a sensor stops reporting? What following distance does it hold, and what does it assume about how fast the vehicle in front can stop? What happens to everyone else when your system fails safe?

Those questions are cheap to ask and expensive to answer, and until this month there was little public data to argue about them with.

The pace is nearly there

None of which should obscure how quick these cars now are.

A2RL says the fastest autonomous lap in qualifying came within 0,85 seconds of a benchmark lap set by Juju Noda, a current Super Formula driver, at a circuit none of the teams had raced on before.

That figure needs a caveat and here it is. The benchmark at Yas Marina last November was set by Daniil Kvyat and the gap there was 1,585 seconds. Different driver, different circuit, different session, and the Imola figure comes from the organiser's own material. Anyone stringing four measurements into a trend line is drawing through points taken in four different conditions. The direction is real. The line is not.

What can be said without qualification is that on an unfamiliar European circuit, five software teams put driverless cars within about a second of a professional driver. At an exhibition in 2024 the same series was around ten seconds off a human lap.

What transfers

Speed is close to solved. Reliability is not, and behaviour around a failing neighbour is not.

It is also worth hearing what the racing does not solve, from somebody inside it.

"We are not the first to build a driverless car in racing," Shimchik said. "Plenty of companies work on autonomous driving and they need a different approach. A car in a city at 60 km/h needs a different approach, with more focus on traffic regulation compliance. Here the conditions are limited. A race track, no pedestrians, no cyclists, no junctions. It is easier here."

That is a useful correction to anyone reading a race result as a preview of a public road. The track removes most of what makes driving hard and keeps the part that is about physics and speed.

What it leaves is the ordering. The remaining work is not making the software faster. It is making it survive a component failure, and making it behave sensibly when the vehicle in front does something no reasonable model predicted.

Imola produced three examples of exactly that in a quarter of an hour, which is why the organiser and the teams left happier than the results sheet suggests.

There is also the part that does not fit in a research plan, which is what a crowd takes home. "A use case that is clear, and visually clear to a mass audience: physical AI is relatively safe," Shimchik said. "If it can do 300 km/h on a track, it can probably manage 40 in town with me in the back."

Sources

The result, lap and speed figures and the ASPIRE comment come from the announcement issued by ASPIRE on 5 September 2026. The causes of the three failures come from Motorsport.com's report of the race. The qualifying gap to the human benchmark comes from A2RL's own material published before the event. The EAV-25 specification comes from A2RL's technical page for the car.

Ilia Shimchik spoke to Inside Deep Tech at Imola on 4 September 2026, the day before the race. Valentino Jadriško commented after the race.