Crashing cars

Hayley Everett

How advanced simulation software is benefiting the design of future autonomous vehicles.

Autonomous cars have been identified as potential solutions to reducing the number of traffic accidents on our roads for some time, however at present they remain involved in crashes, some with fatal consequences. Improving the safety of such vehicles is therefore of high importance within the transport sector, but optimisation remains difficult as simulating crash scenarios of mixed traffic situations is hampered by a scarcity of relevant data.

An innovative machine learning-based software capable of simulating future car crash scenarios is currently under development at IMC Krems – University of Applied Sciences in Austria. At the heart of the project is developing realistic crash scenarios and simulations of future traffic scenarios involving autonomous vehicles with the ultimate goal of making them safer for passengers, other road users and pedestrians.

“Currently, just like humans, autonomous vehicles have misjudgements and might react in unexpected ways,” says Professor Alessio Gambi, project leader at the Department of Science and Technology at IMC Krems. “For this reason and the current prevalence of mixed traffic, autonomous vehicles must meet the same safety and crash requirements as conventional vehicles.”

Simulation for safety

Part of a wider project funded by the European Union (EU) called Flexcrash, Gambi’s work will contribute to developing safety mechanisms for autonomous cars that reduce accident-related consequences. The first step of the project involves extracting driving scenarios from publicly available databases and feeding them into specialised driving simulations. Following this, an optimisation process based on state-of-the-art search algorithms will create novel driving scenarios with increased criticality and severity.

“We are designing a platform to enable running large-scale studies in simulated mixed traffic scenarios,” Gambi explains. “This platform will enable us to study the live interactions between different, and possibly incompatible, driving styles by co-locating human drivers and automated driving agents in the same driving simulations. Consequently, those studies let us identify potentially critical situations that humans, autonomous vehicles, or both cannot safely handle. Additionally, those studies will help us understand how well autonomous vehicles can cope with unpredictable (and sometimes unsafe and erroneous) human manoeuvres. For instance, could autonomous vehicles avoid collisions caused by distracted human drivers departing their traffic lanes.”

Gambi notes that there is currently only one globally available openly accessible database that records the level of autonomy of car involved in a crash – California’s DMV autonomous vehicle collision report database. This is a very limited base for simulating future car crash scenarios, and the project’s aim is to widen that base.

“The ability to virtually simulate various crash scenarios and configurations will allow future generations of vehicles to be designed and engineered for more realistic, non-laboratory crash scenarios,” he adds. “Doing so can improve overall occupant protection and compatibility.”

Harnessing machine learning

Using data from existing sources, the team will select a set of reference driving scenarios as a base for the next project step to feed the BeamNG tech driving simulation. Additionally, the team will develop an online, open simulation platform that follows multi-player paradigms, such as those used in video games where remote players interact with other, and artificial intelligence (AI). Using the platform, the researchers will study virtual live interactions between human drivers and simulated autonomous vehicles in order to generate an additional set of traffic scenarios that are not based on previous accidents, but real interactions.

Leveraging these two sources, specialised search algorithms will calculate virtual crash scenarios that anticipate possible actions by autonomous vehicles.

“We actually develop these algorithms in-house and hence we can easily increase the virtual criticality and severity of the simulated crashes,” says Gambi. “This will expose car structures chiefly involved in severe crashes and predict their behaviour in such situations.”

Improving future vehicle design

Not only is the project aiming to improve the safety of future autonomous vehicles, but it is also tackling challenges surrounding pollution and manufacturing costs.

“Vehicle safety in general, and occupant protection in particular, is at odds with the ecological lightweight of safety mechanisms as these requirements conflict with each other: increasing occupant protection usually implies increasing a vehicle’s mass, which requires using more material and increases production costs and pollution,” offers Gambi. “In the future, hybrid manufacturing and hybrid material technologies will make it possible to develop lighter, safer, and more cost-effective crash structures through virtual optimisation and simulation.”

IMC Krems’ work is the basic work package of the overall Flexcrash project, the results of which will provide data for improving the design of future autonomous cars. The final goal of the wider project is to use hybrid manufacturing technology for applying surface patterns using additive manufacturing onto preformed parts, a technique that the participants hope will contribute to greatly reducing accident-related fatalities, injuries, pollution and manufacturing costs in future.