Self-driving revolution hampered by a lack of accurate simulations of human behavior

Algorithms that accurately reflect the behaviour of road users — vital for the safe roll out of driverless vehicles — are still not available, warn scientists.

They say there is “formidable complexity” in developing software that can predict the way people behave and interact on the roads, be they pedestrians, motorists or bike riders.

To improve the modelling, a research team led by Professor Gustav Markkula from the Institute of Transport Studies at the University of Leeds has developed the first-ever simulation of how people behave on the roads based on key cognitive theories.

Those separate theories were integrated into a larger, single psychological model that would “describe behaviour in more complex, real-world tasks.”

During computer tests, the model accurately reproduced various well-known but not previously understood behaviours of pedestrians and drivers in common road scenarios. The model also predicted how real-life human subjects would behave when facing interactive situations in a virtual reality simulator.

Professor Markkula said: “These findings suggest that everyday road user behaviour relies on a number of complex underlying cognitive mechanisms, which may be part of the reason why it has been more difficult than expected to create self-driving vehicles.”

“Our research shows that it is possible to integrate separate theories from psychology into combined theories for applications such as simulating the way people behave in traffic, which is something which has been called for but rarely achieved.”

The researchers’ findings — Explaining human interactions on the road by large-scale integration of computational psychological theory — are published today (Tuesday, June 20) in the scientific journal PNAS Nexus.

Algorithms needed to unlock self-driving revolution

The development of automated vehicles could have a major impact on the UK economy.

In a vision statement, the UK Government has said driverless vehicles will launch a £42 billion industry and create 38,000 new jobs. The aim is to see the start of the safe roll out of driverless vehicles by 2025.

But writing in the scientific journal PNAS Nexus, the researchers argue that work towards driverless vehicles has been “hampered by a lack of models of how human road users interact.”

Accurate models are needed to run simulations necessary in both development and testing of driverless vehicles and their control systems, for example to demonstrate that the vehicles remain safe when confronted with a range of human behaviour on the road.

Up to now, most computer models of road user behaviour have been statistically based, with predictions of how people might behave based on analysis of large datasets, but typically without analysing those models at a detailed behavioural level.

The research by Professor Markkula and his team has instead focused specifically on the details of human behaviour and key concepts in human psychology.

Road user behaviours and theories

The researchers looked at several typical human behaviours that exist on the road, such as hesitation in unclear situations, or implicit communication using vehicle or body movement to assert priority or to encourage someone else to go first.

The model predicts how people will behave by reference to key cognitive theories. For example, one is “theory of mind,” where people will form beliefs about what someone else is doing and how their own behaviour may affect decisions being made by the other. This relates also to “behavioural game theory,” explaining how people consider the combined effects of their own behaviour and the behaviour of others when deciding what to do.

Another theory incorporated in the model describes imperfect human perception, requiring people to take time to assess and understand what is going on in their environment.

Testing with human participants in the laboratory — including the HIKER pedestrian simulator at the University of Leeds Virtuocity facilities — revealed that the new psychological-theory based model could also make correct predictions about driver-pedestrian interaction scenarios studied in the experiments.

Professor Markkula, who holds the chair in Applied Behaviour Modelling at Leeds, added: “Our research has shown that, by taking a number of existing but separate mathematical theories about human psychology and behaviour, and putting these together, we can model — in much more detail than previously possible — how humans interact in road traffic, for example as drivers or pedestrians, including phenomena such as hesitation and interpretation of others’ intentions.”

In the paper, the researchers say that much work remains to be done in the development of psychological based models of road user behaviour.

The overall aim, say the researchers, is to develop computer models that better reflect the human dimension to behaviour on the roads.

The authors of the paper — Explaining human interactions on the road by large-scale integration of computational psychological theory — are Gustav Markkula, Yi-Shin Lin, Aravinda Srinivasan, Jac Billington, Matteo Leonetti, Amir Hossein Kalantari, Yue Yang, Yee Mun Lee, Ruth Madigan, and Natasha Mera. Matteo Leonetti is from Kings College London — the others are based at the University of Leeds.