Mathematical models of human behaviour are used in many societally important contexts, such as transport, economics, robotics, and epidemiology, to make predictions about the impact of new technologies or policies, or directly as part of technological solutions. These human behaviour models are typically either machine-learned from large datasets, or mechanistic models, based on assumptions about human cognition. However, there are hard constraints on the scope and accuracy of these approaches: machine-learned models can in principle account for behaviour in diverse real-world situations, but only as long as there are large amounts of real-world human behaviour data. In modelling of behaviour where data are scarce (e.g., safety-critical situations) cognitive models may account well for selected scenarios, but do not scale to arbitrary real-world situations. Furthermore, neither approach is able to generalise to entirely novel situations, e.g., human interaction with not yet deployed technologies or interventions. However, thanks to recent advances in both cognitive and machine-learned modelling, a novel approach can now be envisioned with the potential of high-fidelity behaviour emulation across both common and more inaccessible aspects of behaviour, and with high capability of generalisation to new situations. This modelling approach builds on the theory of human behaviour as boundedly optimal: maximising rewards, but under human perceptual, motor, and cognitive constraints. It is argued here that models of this nature can now be achieved for complex real-world human behaviour, by leveraging (1) large-scale integration of existing mechanistic models from fundamental computational cognitive science, to model human constraints, and (2) powerful deep reinforcement learning methods, to learn boundedly optimal behaviour under these constraints. This requires a new type of research programme, spanning state of the art methods both from cognitive science and ICT, in addition to domain knowledge from the applied contexts in question. The PI of this discipline-hop project has a world-leading track record in integrative cognitive modelling in the domains of road traffic safety and vehicle automation. A primary objective of this project is for the PI to be immersed in the relevant subdisciplines in ICT, to establish the cross-disciplinary bridge needed to pursue the envisioned research programme, primarily within his home discipline, but also with a wider range of application areas in sight. At King's College London, the PI will be hosted within a research team with cutting-edge research expertise in the required ICT domains, and with relevant cross-disciplinary experience. During this immersion, research toward a second main objective will be pursued: A proof of concept demonstration in the form of a model of safety-relevant pedestrian-vehicle interaction, with capabilities beyond what is possible with purely machine-learned or mechanistic approaches. This proof of concept will make use of an existing naturalistic dataset from two Leeds locations, in collaboration with Leeds City Council, to investigate the potential of the developed models for simulation-based design of traffic safety interventions. The developed models will have high value within their specific applied domain, but a third objective of this project is to also engage with fundamental and applied behaviour modelling researchers and ICT researchers more widely, to promote the proposed cross-disciplinary line of modelling research for use also in other domains. This will be done within international networks, but with a specific emphasis on strengthening UK capabilities for human behaviour modelling in key application areas, with potential for truly major academic and societal impact.