Many people think that technology is the biggest challenge in the development of self-driving cars. In fact, the most complex problem is the city itself: where people and cars, bicycles and scooters, buses and trams share the traffic, a single decision depends on dozens of factors. Researchers at HUN-REN SZTAKI’s Systems and Control Laboratory (SCL) are working on algorithms that can make quick and safe decisions even in the most complex situations.
One might think that the urban environment is the ideal terrain for self-driving vehicles: lower speeds, a well-developed road network, public lighting, road signs, and traffic lights. However, the reality is quite the opposite. Cities are unpredictable, crowded, and constantly changing, which poses a serious challenge for self-driving systems.
“Urban traffic is incredibly complex. No two situations are the same, and there are many different types of road users present at the same time,” says Szilárd Aradi, lead researcher at HUN-REN SZTAKI SCL and ARNL. One moment a scooter is whizzing down the middle of the lane, the next a pedestrian steps off the sidewalk without looking around. Sometimes a car pulls over and the passenger suddenly opens the door, or the lanes are being repainted due to roadworks and traffic cones divert traffic.
One of the biggest challenges for self-driving systems is understanding what other road users want to do. A cyclist looks to the side – does that mean they are going to turn? A pedestrian is standing at a crosswalk – are they just talking on the phone, or do they want to cross? Human drivers make quick decisions in such situations based on their experience or instincts. Machines, however, must make all decisions based on algorithms, which means they must be able to “understand” complex situations.
Who hasn't been in a situation where the route planner wanted to turn into a street that had been closed a few days earlier? The navigation system didn't know about the diversion, the traffic signs weren't clearly visible, and the road markings were only half done. In such cases, we look around, replan, and instinctively adapt to the situation. However, self-driving vehicles have to decide in real time what constitutes a valid route and how to safely correct their movement. This requires not only perception, but also precise motion planning: where should the car move, along what arc, at what speed – all while remaining stable and not endangering others. This is where the control theory approach plays a key role.
What are Hungarian researchers working on?
At SCL, in addition to describing the theoretical background, they focus on the most effective cooperation between different approaches and systems. "Machine learning-based methods are well suited for pattern recognition, but they have limited functionality in situations that the system has not encountered before. In these cases, control theory models provide guidance by describing how a vehicle behaves under certain conditions based on the laws of physics," explains Szilárd Aradi.
For reliable operation, it is also crucial that the system can efficiently combine data from various sensors, such as cameras, radars, and laser scanners. This is known as sensor fusion: the vehicle can make good decisions if it is able to form a clear picture of its surroundings from multiple sources of information, even if this information is incomplete or contradictory.
"One of our research areas is the safe planning of emergency evasive manoeuvres – for example, when a tree branch suddenly falls onto the road during a storm and the car is traveling at high speed with poor visibility," says Szilárd Aradi. "In such cases, the system must quickly calculate the optimal trajectory and speed in real time and reliably predict how other road users are likely to react."
Another interesting example is rear monitoring. People only glance at the rear-view mirror from time to time, as we basically look where we are going. However, the self-driving vehicle system can continuously monitor what is happening behind it. If it detects that another car is approaching too quickly from behind, the system can even intervene by accelerating, thereby avoiding a collision or at least reducing the force of the impact – provided, of course, that the space in front allows this.
"In our view, true autonomous driving and the ability to respond appropriately to unexpected situations in urban environments do not depend on a single algorithm or perfect sensors, but on the reliable, transparent, and predictable operation of the entire system," says the researcher. This system-level thinking enables autonomous vehicles to integrate step by step into the real traffic environment – today only on closed tracks, but eventually also in urban traffic.
The above press release received numerous online appearances in Hungarian professional and other media, such as gyartastrend.hu, vezess.hu, pcworld.hu, lepesmagazin.hu, trademagazin.hu, digitalhungary.hu.