To build the car of the future, the traffic scenario of the future must first be illuminated. After all, vehicles are always optimised for their use, for example for right-hand or left-hand traffic today. The aim of the EU’s Flexcrash project is to build lighter and more environmentally friendly cars, which use less material to ensure passenger safety.
The first step in the project, which runs until 2027 and brings together a total of ten research partners, is taking place at IMC Krems. Before the material mix, design and assembly instructions for future passenger cars can be devised and scaled on an industrial scale, software engineer Alessio Gambi tries to identify critical crash scenarios of the future: “We are assuming a mixed scenario with autonomous vehicles as well as those controlled by humans on the road in order to develop the cars for 2050. The goal is to understand which accident scenarios can be expected when some road users react according to programmed rules and safety concepts and others simply react the way humans are: distracted, in a hurry, overtired, unfocused, possibly tipsy, experienced or novice drivers.”
Gambi assumes there will be fewer and less dangerous accidents in 2050, but not zero. If an autonomous vehicle needs to prevent an accident, it will follow clear rules. Combined with human unpredictability, this could change the safety-relevant points on the car body.
Historical accident research and AI for driving style
Since he wants to know how artificial intelligence might respond to human driving behaviour, Alessio Gambi must clear two hurdles: data on accident histories exist only from the human past, while information about autonomous vehicles’ safety protocols tends not to be published. In Europe, only driver assistance systems are currently approved, but no autonomous vehicles (AVs). “We treat autonomous vehicles like a black box because we don’t know the exact criteria. However, we expect a typical AI driving style to emerge over the next 25 years. The challenge in the coming years will be that AI and human traffic behaviours have little compatibility.”
Therefore, in a first step, he is training methods for machine processing of language (Natural Language Processing) and images on U.S. police reports (NHTSA) that attempt to describe accident sequences with texts and sketches: “To do this, we give AI a kind of vocabulary book with important parameters for this scenario: vehicle colour and type, speed, weather, road conditions, time of day, construction site, intersection, highway, lane, manoeuvre, turn-in angle and more. In this way, important information can be read out automatically and translated into 3D representations.” To simulate the reaction of autonomous vehicles to typical accident hazards, the team is working confidentially with the system of an autonomous vehicle from the Technical University of Munich, which is designed to avoid collisions. Trajectories, i.e., possible paths of movement of objects, are the basis for this. Safety settings exclude some trajectories to avoid collisions and calculate the “best” steering angle if a collision is unavoidable. “The system is similar to the one in the robo-vacuum cleaner for the living room,” says Gambi.
Typical accident scenarios – new safety parameters
The result of these simulations with variable parameters are “typical” predefined movement patterns of autonomous vehicles, e.g., when someone suddenly changes lanes to avoid a crash. The last step to test the interaction of humans and machines without an emergency, is a “serious game” in which car drivers react to variants of critical scenarios, e.g., at different times of the day, and then learn what the others did. “My personal hope is that there will initially be lanes reserved for autonomous traffic, as there are for trucks, for example,” says Gambi.
For the time being, there will be no personalised cars for anxious, risky, defensive drivers. The task is therefore to distil typical accident scenarios from the simulated case studies. Even when exploring unpredictable behaviour, the software engineer ultimately comes to conclusions as to which components are most likely to be damaged in a future crash. The other Flexcrash partners take over from there and specifically reinforce (new) stressed areas on the car body, for more safety in the traffic scenario of the future – between artificial intelligence and human irrationality.