Do We Still Need Autonomous Driving Competitions?

Written by | Zhang Linyu

Edited by | Huang Dalu

Designed by | Zhen Youmei

Five formula racing cars, no drivers in the cockpits.

On September 5, 2026, the first European stop of the Abu Dhabi Autonomous Racing League (Abu Dhabi Autonomous Racing League, A2RL) was held at the Imola circuit in Italy.

A2RL is organized by ASPIRE, a major tech challenge organization under the Advanced Technology Research Council (ATRC), established by the Abu Dhabi government of the UAE.

Abu Dhabi hosting the race is not just to stage a driverless performance. ATRC, when announcing the event in 2023, listed three levels of goals:

Test artificial intelligence with unified racing cars and extreme speed pressure, directing research results towards road safety;

Attract global universities, research institutions, and technology suppliers with prizes, circuits, and R&D platforms, establishing a local autonomous mobility ecosystem;

Cultivate STEM talent through student projects and position Abu Dhabi as a research and development center for applied AI and autonomous driving.

In other words, A2RL is at the same time a research plan, a talent engineering project, and an urban tech brand project.

But this global mobilization did not attract the US electric vehicle and AI company Tesla.

Tesla has never participated in any similar competition as an official team.

In the lists of events that can be found, there are no enterprise teams founded or controlled by Musk in the UK robot racing series Roborace, the Indy Autonomous Challenge, or A2RL.

Musk has also been publicly invited by name.

In 2025, Markus Lienkamp, a Professor of Automotive Technology at the Technical University of Munich in Germany, called out to Musk on his personal LinkedIn account, stating that TUM's software can run on public roads as well as tracks, and asked if Tesla was willing to race in the US.

The official account of the Indy Autonomous Challenge (Indy Autonomous Challenge, IAC) subsequently suggested moving the race to the Laguna Seca circuit. Since then, there has been no public follow-up; neither Musk nor Tesla responded, nor did they form a team to compete.

Only four cars remained at the official start.

The TUM racing car from the Technical University of Munich in Germany suffered a braking failure during the formation lap and returned to the pit lane early. On lap 10, the leading Unimore car suddenly slowed down at the final left turn; the following PoliMOVE car swerved inward to avoid it but still hit the car in front.

Tim Stevens, a reporter for an American automotive magazine, directly interviewed Matteo Pini, the head of software integration for Unimore, after the race. Pini stated that the onboard switch aggregating multi-sensor signals failed during the turn, causing the vehicle to lose key perception instantly; safety logic subsequently triggered emergency braking.

Finally, only two out of five cars completed the 12-lap main race. Kinetiz, representing the UAE, slowed down before the accident zone and passed through, winning the championship; Germany's Constructor Racing took second place.

This race was easily treated as a joke: driverless taxis on roads are already carrying passengers, yet a group of European college students caused three racing cars to retire; did they use their strength in the wrong place?

Twenty-two years ago

Competitions seek answers for the industry

The first collective acceleration of the autonomous driving industry indeed came from competitions.

In March 2004, the US Defense Advanced Research Projects Agency (DARPA) held the first Grand Challenge for unmanned vehicles in the Mojave Desert.

Review data released later by DARPA showed that none of the 15 finalist cars completed the 142-mile course; the farthest, Carnegie Mellon University's Sandstorm, only advanced 7.4 miles.

One year later, the second competition increased the prize to $2 million. Stanford University's Stanley completed 132 miles in 6 hours and 53 minutes and won the championship; four other cars completed the course.

The 2007 DARPA Urban Challenge moved the test from the desert to a simulated city, requiring vehicles to obey traffic rules, pass intersections, and handle other moving vehicles; 6 out of 11 finalist teams completed the race.

The value of these three races was not in selecting a mass-producible car. At that time, the industry didn't even have an answer to "whether a machine can drive continuously for a stretch".

DARPA provided common goals, deadlines, prizes, and venues for public failure, compressing localization, perception, planning, and control research originally scattered in universities and military projects into a single vehicle that had to run.

Sebastian Thrun, head of the 2005 champion team, later participated in creating the Google autonomous driving project. In reviewing the Grand Challenge, he said this competition "advanced autonomous driving significantly".

In 2018, he suggested Indianapolis Motor Speedway launch a successor event, hoping to continue pushing the limits of machines with high speeds and multi-car competition. This suggestion later became a significant starting point for the Indy Autonomous Challenge (IAC).

China also established its own real-vehicle competition during the same period.

In 2009, the National Natural Science Foundation of China launched the Intelligent Vehicle Future Challenge (IVFC).

The first event was held in Xi'an; 6 teams participated, including Beijing Institute of Technology, Hunan University, Xi'an Jiaotong University, Shanghai Jiao Tong University, National University of Defense Technology, and the University of Parma, Italy, inspecting traffic light recognition, obstacle avoidance, and designated route driving.

It helped China's unmanned vehicle research move from single algorithms to real-vehicle integration.

These early competitions suited the technical status of the time.

How to combine sensors, how to build maps, how to locate vehicles, how to hand over recognition results to planners—different teams had their own answers. Competitions eliminated unrunnable schemes using finish lines and timers, and also brought runnable schemes into the public eye.

Technology Path Convergence

Questions Cannot Stay at the Previous Generation

Since then, autonomous driving competitions have expanded in different directions.

The UK company Roborace attempted all-electric driverless racing; German Formula Student competitions set up a Driverless category since 2017; IAC started allowing university teams to develop software on unified Dallara race cars from 2021; A2RL held its first car race at the Yas Marina Circuit in Abu Dhabi in 2024, making multi-car confrontation and human-machine lap time comparison its selling point.

The China Intelligent Vehicle Future Challenge gradually added urban roads, mixed traffic, and simulation tasks.

However, the basic technical framework for cars driving themselves is no longer a no-man's land.

In the past, the industry debated LiDAR vs. pure vision, HD maps vs. map-less, rule-based systems vs. neural networks.

Today, sensor and commercial boundaries for each still differ, but training paradigms are converging towards learning-driven end-to-end models, multimodal foundation models, Vision Language Action (VLA) models, and world models.

In the 2026 International Conference on Computer Vision and Pattern Recognition (CVPR) activity introduction, Tesla described the latest version of FSD as an end-to-end driving model, claiming it trains driving strategies using large-scale embodied intelligence datasets formed by millions of vehicles.

Tesla also placed cars and humanoid robots Optimus in the same "real-world artificial intelligence" narrative.

Musk had already called cars "robots on wheels" at Tesla AI Day back in 2021, meaning vehicles and robots must perceive, judge, and execute actions from real environments.

This also explains why Tesla had no motivation to compete.

A2RL uniformly provides 7 cameras, 4 millimeter-wave radars, and 3 LiDARs; participants develop software around the same set of hardware; Tesla chose to center on cameras, onboard computing, and large-scale fleet data.

Re-adapting a dedicated race car with LiDAR for a competition makes it difficult to directly prove the capability of its mass-produced systems with win/loss results: Losing incurs public opinion costs, and winning cannot prove the system can handle unprotected left turns in San Francisco or construction detours in Shanghai rainy nights.

Tesla's coldness indicates a distance between such competitions and the R&D mainline of top enterprises, but it cannot be said that the competitions are meaningless.

Enterprises have mass-produced fleets, user data, and commercial operation scenarios; university laboratories do not have these conditions.

For the latter, unified hardware and shared tracks remain a rare public infrastructure.

EMMA. Waymo, an autonomous driving company under Google's parent company Alphabet, is researching this end-to-end multimodal model. It utilizes Google's Gemini multimodal large model to directly convert camera input into vehicle trajectories, while explicitly studying the pros and cons of pure end-to-end solutions.

Embodied AI. The UK autonomous driving company Wayve uses this concept to summarize its route. Founder and CEO Alex Kendall proposed that the same model should be able to work across vehicles and across cities; Wayve's 2026 public materials summarized industry changes as shifting from proving end-to-end feasibility to scaled deployment.

Physical AI. Jensen Huang, founder and CEO of US computing platform company NVIDIA, categorized autonomous driving cars and robots together into this category. NVIDIA's Cosmos World Foundation Model Platform focuses on generating rare scenarios, predicting future states, conducting closed-loop simulations, and training and evaluating machines with massive synthetic data.

This does not mean autonomous driving only has one completely identical route remaining.

Sensor configuration, map dependency, model structure, safety redundancy, and commercial scenarios are far from unified. What truly converges is the R&D question: Is the data diverse enough? Can the model be transferred to unfamiliar environments? Is the system reliable when encountering long-tail events? Can it exit safely after failure?

At this stage, whether a car can run the full length on a known track is no longer the scarcest evidence.

What is truly scarce is whether it can run on a different track that hasn't been practiced, what happens if the camera is blocked by strong light or the perception data link is interrupted, whether the model still holds when placed on another vehicle, and whether all failures left reproducible data.

Did European Universities Use Their Strength in the Wrong Place?

When Tesla, Waymo, Wayve, and NVIDIA direct resources to foundation models, world models, simulation, and real-road deployment, European universities continue to tune algorithms around formula racing cars; it indeed appears there is a generational mismatch. The background of the Imola competing teams easily reinforces this impression.

Among the five competing teams, Unimore Racing comes from the University of Modena and Reggio Emilia in Italy; its core is the HiPeRT Lab for High-Performance Real-Time Systems, supported by real-time computing technology company Hipert.

PoliMOVE is an unmanned racing team under the mOve Automation and Control Research Team of the Politecnico di Milano, Italy.

TUM is the Technical University of Munich's autonomous driving racing team, having won IAC and the A2RL championships in 2024 and 2025.

Constructor Racing was co-founded by Constructor University, a private research university located in Bremen, Germany, and the Constructor technical system; the predecessor team participated in Roborace.

The champion Kinetiz is not a purely UAE commercial team, but a joint team composed of Singapore's Nanyang Technological University, UAE technology company K2, and autonomous driving software company Alp Autonomy, competing in the name of the UAE.

The bulk of these five teams are university laboratories and technical partners, not factory teams from Volkswagen, BMW, Mercedes-Benz, or Stellantis.

Comparing them directly with Tesla's data scale or Waymo's Robotaxi mileage places them on the wrong coordinates.

Universities compete first to train students and verify research.

High-speed racing amplifies latency and control errors: when a vehicle travels at 250 km/h, 100 milliseconds advances about 6.9 meters; a slight positioning deviation, slightly late braking, or a planner error in predicting an opponent's trajectory can immediately become an accident. This stress test trains talent in real-time systems, vehicle dynamics, sensor fusion, and fault handling.

Competitions can also generate public research materials.

In 2026, 8 participating teams including TUM released the A2RL Vmax open dataset, containing over 300,000 frames of LiDAR point clouds, over 380,000 frames of radar point clouds, and multi-car interaction data collected from the 2024 race, providing nearly 30,000 frames of professional annotations. For teams researching high-speed long-range perception and opponent trajectory prediction, this has more sustained value than the podium.

So, European universities are not stupid. They are working on a well-defined research question: on unified racing cars and closed tracks, push the machine's real-time perception, prediction, and limit control to higher speeds. The problem is that event promotion often states this boundary too broadly, as if racing faster by a few seconds would naturally translate to urban autonomous driving.

If a competition only allows teams to repeatedly scan the same track, build fine maps, and tune parameters for fixed vehicles, it looks more like a high-level systems engineering competition, unable to prove the model has universal driving capabilities across cities and vehicles. Universities did not go the wrong way, but competitions must honestly state where this road leads.

What Was Actually Measured by the Imola Crash?

A2RL's current rules have an important advantage: racing car hardware is basically consistent; the driver's seat, steering wheel, and pedals are replaced by computing equipment and actuators; remote driving from the pit area or real-time parameter modification is prohibited during the race. Teams can receive flag signals and telemeter back; specific steering, braking, and overtaking decisions are completed by the onboard system.

This setup puts software and whole-vehicle integration capabilities in the foreground. The Imola accident exposed at least four questions that can be further investigated.

First, having many sensors does not equal redundancy in the perception link.

MotorTrend's interview pointed the fault to the data aggregation switch; Motorsport.com only disclosed sensor failure. Currently, there is no final investigation from teams or organizers. It can be confirmed that the vehicle lost key perception instantly. The competition should disclose where the failure occurred—sensor, data link, or aggregation node—whether backup paths exist, and what state estimation capabilities the vehicle retained.

Second, minimum risk maneuvers cannot only protect the faulty vehicle.

Unimore's emergency braking after losing perception is reasonable in single-car logic but creates a sudden stationary obstacle for following cars in multi-car competition. The system needs to evaluate its own remaining capacity, oncoming vehicles behind, available escape areas, and deceleration simultaneously, not just have "continue running" and "stop in place" as options.

Third, the following car must also identify the leading car's failure.

PoliMOVE attempted to avoid but failed. Besides following distance and braking capability, competitions should check if the leading car's failure state can be quickly transmitted via vehicle lights, V2I communication, or race flags, and whether the following car can timely judge based on motion abnormalities alone when communication is missing.

Fourth, reliability changes competition strategy.

Kinetiz did not run the fastest lap time of the day, but put completing the race first and actively slowed down in the accident zone. The championship thus rewarded risk selection and system stability. But with only one 12-lap race, it still cannot distinguish whether this is repeatable reliability or a correct judgment plus an accidental result caused by the leading car's retirement.

Post-race data released by ATRC shows that the five teams conducted 9 days of real-vehicle testing before the race, experienced rain and hail, and after changing to wet-weather kits, completed a total of only 38 laps of wet driving.

Organizers also provided teams with Imola digital scans and simulators. Changing tracks increases difficulty, but this is still not "zero-shot" adaptation: teams got the track model in advance and had time to train and debug around it.

This is exactly the distance between A2RL and next-generation autonomous driving models. Wayve emphasizes bringing the same model directly to unfamiliar cities; world models emphasize large-scale generation of never-before-seen long-tail scenarios; A2RL currently still allows participants to concentrate optimization on one vehicle, one track, and one race.

The former tests generalization; the latter tests specific performance. Both are valuable but cannot replace each other.

How Should Today's Competitions Be Changed?

After technology path convergence, competitions no longer need to prove "cars can drive themselves".

It should turn the hardest-to-publicize, hardest-to-repeat failures in the industry into standardized test questions. If A2RL wants to continue to represent the forefront of autonomous driving, it can at least add four sets of rules beyond existing racing.

Write unfamiliar environments into rules.

Competitions can announce local track changes only after code freezing, no longer providing all high-precision scans in advance; Temporarily change cones, drivable areas, road surface friction coefficients, or flag positions in the final. Competitions can also require the same set of driving models to continuously complete high-speed track, low-speed pit lane, and slippery road tasks, limiting time for retraining on a single track.

Actively create controllable failures.

In closed testing, sequentially turn off one camera, lower LiDAR refresh rate, cut off satellite positioning, limit onboard computing power, or simulate sensor and data link interruptions. Scoring should not only look at whether the vehicle stopped, but also how long it took to identify the fault, what parking position was chosen, how much reaction distance was left for following cars, and whether it could exit the main track in degraded mode.

Test whether the model can be transferred.

Software from the same team should run on two sets of vehicle dynamics parameters or two tracks. Participants need to declare map dependency, training data, simulation hours, and computing resources; the competition will publish results separately for zero-shot, limited adaptation, and fully tuned states. Only then can it be seen if victory comes from general capability or from longer track-specific tuning.

Publish failure data along with the champion.

After the race, organizers should release an independent technical report detailing the trigger chains for every retirement, emergency stop, manual abort, and collision; publicize key telemetry, scenario data, and replication experiments while protecting trade secrets.

Imola's most valuable outcome should not be just Kinetiz's trophy, but should include why Unimore lost perception, where exactly the fault lay—sensor or data aggregation link—why PoliMOVE couldn't avoid it, and whether improvement schemes can pass the next fault injection test.

These four changes will make the competition look worse. Vehicles may become slower, completion rates may drop, and human-machine lap time showdowns won't always suit promotional videos. But it will make results closer to evidence truly needed by the industry: whether the same system can still provide predictable behavior under incomplete information, environmental changes, and hardware failures.

If A2RL continues to only race the fastest lap time on familiar tracks, its main value will remain in talent cultivation, extreme sports control, and Abu Dhabi tech brand display. This value is real, but insufficient to represent the next stage of autonomous driving.

If it writes unfamiliar environments, system failures, cross-platform migration, and data publicization into the rules, it can become a public impact laboratory that the industry dare not do and cannot jointly do usually. Enterprises can generate a million accidents in simulation, but rarely are willing to put the failure process, logs, and repair results of real vehicles under the same set of rules for outside comparison.

Competitions can just happen to undertake this matter. Answers on the track must ultimately return to roads for verification.

Twenty-two years ago, DARPA asked with a desert route: Can machines drive cars to the finish. Today, the ultimate exam venue needed by the industry is no longer on closed tracks, but at every morning rush hour intersection and the physically evolving real world.