The grandstands erupted with shock as engineers clutched their heads in the paddock. One driverless race car, produced by the company PoliMove, slammed into Unimore’s slowing vehicle in an event held by the Abu Dhabi Autonomous Racing League (A2RL).

At the famous racing track in Imola, Italy, Autodromo Internazionale Enzo e Dino Ferrari, both cars crashed in an area called Rivazza. This is the circuit’s final sequence, made up of corners that are notoriously difficult to navigate. Formula 1 champions have met a similar fate here. But this time, the impact posed no danger to the drivers because there weren’t any.

Launched in 2024, A2RL sought to introduce self-driving cars to the motorsport to test the boundaries of their current capabilities. And this year’s race at Imola revealed those edges, demonstrating of the gap between the computational abilities of autonomous vehicles and their capacity to respond when something goes wrong.

Only two of the five cars in the final completed the race. The UAE’s Kinetiz finished first, followed by Germany’s Constructor Racing. PoliMove clocked in at third place, although its battered entrant had to be replaced for the podium ceremony. Two-time winner TUM withdrew its car after it developed a brake problem during the formation lap.

Participating teams had just nine days of physical testing before the race, in conditions that included rain and hail. This also happened to be the league’s first international competition away from its usual testing ground at Abu Dhabi’s Yas Marina Circuit.

So why choose one of motorsport’s most unforgiving tracks? “Because everybody can do ‘easy’, right?” says Nicola Palarchi, engineering director at Aspire, the company that founded A2RL. “We have to show we go where it matters.”Palarchi made it abundantly clear that autonomous racing did not intend to compete with human racing, but rather stay its own motorsport. It could even support the latter.

Because autonomous racing is pitched as more than spectacle: It’s a laboratory with guardrails. Its proponents describe it as a way to test perception, prediction, and vehicle control at extreme speeds without endangering a human driver. These are all factors in the safety of driverless cars in the real world, and so measuring the limits at which these vehicles succeed—or perhaps more importantly fail—can inform future improvements in the technology.

The first hurdle is perception. The vehicle’s sensors identify its location and surroundings, informing planning software that then chooses a trajectory. Finally, the control system translates that decision into steering, acceleration, and braking. Racing makes each of these stages more challenging, since at high speeds, vehicles have less time to interpret incomplete information, predict another car’s movements and execute an evasive maneuver.

“The car will be blind for a certain point because the track drops off,” Alexander Winkler, A2RL’s head of sporting, had said before the final. If one vehicle stopped beyond that drop, he explained, the car behind could have less than a second to react.

That warning proved prescient. Unimore said its car, Gianna, initiated a safety stop after losing all data from its lidar and radar sensors. GPS alone could not locate the car accurately enough to continue racing at speed. The team said future systems could use cameras and an inertial measurement unit to continue at a slower pace until precise localization returned. In other words, the car stopped because it lacked a reliable fallback.

PoliMove’s vehicle, Eva, detected Gianna ahead, according to the team, but could not avoid hitting it. Gianna had braked with a force of approximately 1 g with Eva just 1.5 seconds behind it.

PoliMove said its system identified the danger and began an evasive maneuver. But perception, decision-making, trajectory replanning, actuator response, and vehicle dynamics all introduce delays. Once Gianna braked that severely mid-corner, the team said, a collision was physically impossible to avoid.

Even F1 legends like Fernando Alonso and Carlos Sainz have crashed at this part of the track. However, the A2RL teams’ explanation exposes a problem that faster processors alone cannot solve: an autonomous vehicle may recognize a hazard and still be unable to respond before physics takes over.

TUM’s failure was less algorithmic but equally instructive. The team said its car, Hailey, retired after the rear-left brake became stuck at low-to-medium pressure during the formation lap. Even sophisticated autonomous software cannot compensate for mechanical faults.

Kinetiz’s car, Sparkz, took the lead all the way to a first-place podium finish with Constructor AI’s vehicle as the runner-up. But almost any result can be considered progress, when the failures produced useful edge cases: sensor loss, compromised localization, sudden braking, limited visibility, and interaction between autonomous systems making separate decisions.

Beyond crash data, Palarchi says these motorsport environments provide the opportunity to develop more resilient hardware. Temperatures in cockpits can get near 170 degrees F, and—while hopefully your robotaxi is not that hot—radars and lidars that can withstand high heat can be useful in our warming cities.

Chee Kiong Ong, deputy team principal at Kinetiz, also observes that predictive algorithms developed on the track could eventually help civilian vehicles respond to emergencies. “You could probably apply it to make the car stop itself in a safer manner or control it at the limits so that you can save lives,” he says.

One caveat is that unlike other lab testing, the companies aren’t required to share the precise data collected from their vehicles. The data from A2RL’s EAV 25, the racing car developed by the league based on modified Dallara SF23 cars, currently stays within the competition.

The next A2RL race returns to Yas Marina, where teams have spent years developing their software. Organizers are already considering making the competition harder by shortening testing periods, increasing the grid to eight cars, or restricting access to GPS.

Those changes could reveal whether the systems are becoming more adaptable. In the meantime, we’ll wait to see how this information transfers into real-life scenarios. This story originally appeared in WIRED Middle East.