Autonomous driving: almost a reality

Driverless cars will soon be on our roads, too, and they still raise questions and concerns about safety and responsibility. We explored the topic in an interview with Prof. Paolo Tonella, Professor at USI Faculty of Informatics and an expert in software testing methodologies, techniques and tools.

"Relax and enjoy the ride": in a few years, a voice assistant will repeat this phrase every time we get into a car. No more hands on the steering wheel or pedals to press: it will be the era of autonomous driving, the future of mobility. The first versions of level 4 vehicles (high automation) are being tested. For level 5 (fully autonomous driving), the timeline is longer – optimists talk about 2035 – both because a large part of the existing legislation will need to be revised (one example above all: in the event of an accident, who will be responsible if the car has no driver?), and because the technological challenges that the development of fully automated vehicles require are not easily solved. To better understand our current position in this final area, we consulted Paolo Tonella, a professor at the USI Faculty of Informatics. He specialises in methodologies, techniques, and tools for testing software, which involves verifying that software behaves as expected and meets specific quality standards. Over the past seven years, he has shifted his focus to programs that incorporate components based on artificial intelligence, such as those utilised in autonomous driving.

Professor Tonella makes one thing clear from the start: "I'm not involved in the development of self-driving cars, but in developing methods to test these vehicles in extreme conditions, to see what happens when these occur. Self-driving vehicles are a very interesting sector for those involved in software testing, as they operate in complex environments where it is difficult to predict all possible scenarios".

Are the tests conducted on simulators or the road?

"Testing on the road is costly, risky, and challenging, so we rely on simulators. We started in 2018 with Udacity, an important step because it allowed us to test various options. Today, the most accurate simulator available in the world of research is Carla. This very advanced system allows you to test cars in an urban environment, with buildings, different vehicles, cyclists and pedestrians crossing the road, so it is very close to reality and its unpredictability."

So, do we still not have a real test?

"No. To move towards the real world, we took 1:20 scale models and tested them on a laboratory circuit. Everything closely resembles reality. For several years, we have engaged students from across Switzerland—high school students, our Bachelor's and Master's students, and even some PhD candidates—in an initiative called Formula USI. This competition for autonomous driving is based on these models. We teach students how the training algorithms work, how to collect data, how to proceed with the training, how to test the cars in the field, and then on the last day, the actual competition takes place on a carpet-like surface with a printed road circuit. There are two competitions: in the first, the participants have to programme their vehicle to travel along the track at the maximum possible speed without leaving the lane; in the second, they compete in a direct elimination challenge. This year, we are planning to introduce an important innovation."

Which one?

"Until now, the circuits have been physically printed, which meant that the circuit used on race day differed from the one used during training. This year, we will use a projector to display the circuit on the floor with high definition and brightness. We have conducted some tests, and everything is functioning well. This development will greatly simplify our process, as we will be able to easily modify the circuit, shapes, and other components both during training and competitions. Additionally, Formula USI has been proposed to participate in Phänomena, the Swiss technology expo scheduled for 2026, which will take place from March to October in Dietikon (Canton Zurich). We are also considering ways to engage the audience."

On a practical level, how far have we come in terms of autonomous driving?

"Level 4 autonomous vehicles are currently being tested. Waymo, a U.S. company, launched a regular autonomous taxi service in San Francisco and Phoenix in April 2024. These services operate in urban areas, which are more complex to navigate than suburban areas, which are generally simpler. Waymo has succeeded with this initiative, partly because robotaxi rides are less expensive than those with human drivers. However, the safety data is collected by them, so we have to take it with a pinch of salt, even if their vehicles have travelled millions of miles in almost a year. That said, it appears that robotaxis are safer than human-driven cars; they have fewer accidents, and when they do have accidents, they mostly cause material damage (scratches, bumps, etc.). In short, they improve safety, even if the problem is not this but the perception of it. This is because people, often wrongly, tend to trust their perception more than the numbers. This means that even if it were proven by independent studies that self-driving cars are safer, a serious accident with wide media coverage would be enough to cast doubt on this and make people view autonomous cars with suspicion and fear."

How do robotaxis work?

"With an app similar to Uber: I see where the available ones are, I call, and the closest one arrives. There are still controls in the car, but there is no driver; however, there is a remote control, an operations centre with humans who follow the vehicles and can intervene in critical situations. The users are very happy, but I am unsure how representative these reviews are. Typically, when a new technology is introduced, the first adopters are often enthusiasts who are more willing to accept disruptions and inconveniences. I'll try it the next time I visit."

From an informatics point of view, what is the difficulty of moving to levels 4 and 5?

"Exceptional situations and unforeseen events, which are by nature rare, require sufficient data to adequately train software. To address this, we need to implement automatic anomaly detection systems that can recognise abnormal situations and respond appropriately. Managing these cases from an informatics perspective presents a challenge we are actively tackling. We aim to determine when it is appropriate to activate protective systems or switch to a lower level of autonomous driving so the human driver can take control."

We are experimenting with level 4. Prospects for 5?

"Waymo robotaxis are not that far away. The challenge for achieving level 5 autonomy lies in gathering enough data to ensure safety. While it's unrealistic to expect a zero accident rate, reducing the probability of serious incidents significantly is feasible. For instance, it could be ten times lower than the rate of serious accidents associated with human drivers. For a company like Waymo, reputation is crucial; if they fail to maintain it, they risk losing investors and customers, which could jeopardise their business."

Theoretically, it is possible to hack any programme. Is security guaranteed in this field?

"The primary issue in this field pertains to car sensors, which can serve as potential entry points for attackers. For instance, we have observed that self-driving cars can detect images, possibly displayed or projected on billboards, that may disrupt their functioning, even if these images are harmless to the human eye. We are actively working on addressing this concern. Overall, while it is important to recognise that absolute security does not exist, we can assert that similar to traditional computer viruses, effective countermeasures against attacks on self-driving cars are available. However, implementing these measures requires awareness and specialised skills in computer security."

News: Switzerland, a pioneer in Europe

From 1 March, automated driving will be allowed in Switzerland. The Federal Council approved the regulation in December. This method "can increase road safety and improve traffic flow, as well as opening up new opportunities for the economy and transport services."

The regulation provides three possibilities: autonomous driving on motorways, where level 3 will be possible; vehicles without drivers but remotely monitored by an external operator in urban areas on authorised stretches (under the jurisdiction of the cantons, level 4); and automated parking.

So far, FEDRO, responsible for examining requests, has authorised two experiments: a parcel delivery system in Bern and an electric minibus for public transport in Schaffhausen. A third project, a robotaxi service proposed by SBB in collaboration with the Canton of Zurich, is in the approval phase.

Good to know: the six levels of autonomous driving

Level 0: the driver is fully responsible and in full control of the vehicle. Assistance systems such as ABS or ESP may be present, but these do not intervene directly in driving.

Level 1: Driver assistance systems such as Cruise Control maintain the set speed or the safety distance from the vehicle in front. The driver remains responsible for controlling the car.

Level 2: Advanced assistance systems such as lane-keeping assist and automatic speed control are available. However, the driver must be ready to take back control anytime.

Level 3: the vehicle can manage traffic situations autonomously, leaving the driver to intervene only in certain situations. The system can also take total control in specific sections, such as on motorways.

Level 4: the car can operate autonomously in defined geographical areas or specific conditions, such as cities or motorways. The drivers can dedicate themselves to other activities.

Level 5: the car is driverless and requires no human intervention. There is no steering wheel or pedals, and the vehicle can face any road condition or unexpected event on its own.