Modelling Pedestrian-Vehicle Interaction for Autonomous Vehicles: From Computational Rationality to Adversarial Testing

As autonomous vehicles (AVs) move towards real-world deployment, their ability to interact safely and naturally with human road users has become an important challenge to address. Meeting this challenge requires accurate models of road user behaviour. Existing modelling approaches are typically either mechanistic, which offer interpretability but struggle to generalise across complex environments, or data-driven machine learning (ML) models, which provide strong predictive ability but lack transparency and overlook the mechanisms that generate human behaviour, with performance strongly dependent on data availability and representativeness. This thesis adopts the computational rationality framework, viewing human behaviour as boundedly optimal decision-making under human perceptual and motor constraints, and uses reinforcement learning (RL) to model human-like interactions in road-crossing scenarios.
The first study develops an RL model that explains pedestrian crossing decisions as boundedly optimal behaviour under noisy visual perception. The model captures the observed dependencies of crossing decisions on vehicle time-to-arrival (TTA) and speed, showing that such behaviour can be interpreted as a rational adaptation to perceptual uncertainty.
The second study extends the framework to include motor constraints by integrating a bio-mechanical representation of walking into the RL problem. This approach enables continuous control of crossing actions and reproduces human-like trade-offs between time pressure and walking effort.
To extend the framework beyond controlled experiments and capture real-world interactions, the third study introduces a multi-agent RL framework in which both pedestrians and vehicles are modelled as adaptive agents interacting under perceptual and motor constraints. The resulting interactions reproduce key behavioural phenomena such as yielding and human-like smoother speed adjustments, closely aligning with patterns observed in a naturalistic dataset.
Finally, an application study evaluates the use of a human-like pedestrian model for AV testing. In closed-loop simulation with a full AV control stack, the interaction outcomes of the rule-based CARLA pedestrian and the behaviourally realistic model are compared. The results show that the human-like model produces more realistic gap acceptance patterns and smoother vehicle decelerations, and that the generated adversarial scenarios can effectively tune AV braking behaviour to achieve safer and more efficient pedestrian-vehicle interactions.
Together, these studies represent the first application of the computational rationality framework to model and explain pedestrian-vehicle interaction. They show that boundedly optimal decision-making under perceptual and motor constraints can reproduce human interactive behaviour. Moreover, the thesis highlights the effectiveness of human-like behavioural models for AV testing, demonstrating their value in enhancing the realism of simulation-based evaluation and control design.