Towards Safe and Efficient, Interactive Planning with Learning-based Model Predictive Control

Licentiate thesis, 2026

Planning has emerged as a fundamental component in modern autonomous systems (AS), spanning a wide range of applications from manufacturing and robot manipulation, to the focus of this thesis: autonomous vehicles. A key remaining challenge for these systems is operating in environments shared with other actors, such as, humans or other robotic systems. To find a plan that is both safe and efficient, the AS needs to predict the plan of other actors. Crucially, this includes how the AS and other actors influence each other, or in other words, how they interact.

This thesis investigates the intersection of learning- and optimization-based control approaches. In particular, we investigate optimal control over finite, receding horizons, commonly referred to as Model Predictive Control. Three Model Predictive Control problems are investigated, each with a fundamentally different treatment of the predicted plan of other actors: Nominal, considering only the most probable future outcome; Stochastic, considering a distribution of the future outcomes, and Distributionally Robust, considering a set of distributions of the future outcomes.

The results provides steps towards interactive planning with guarantees on an optimal tradeoff between safety, and efficiency. In particular this includes, theoretical, algorithmical and applied developments, that introduce a close coupling between the planner for the AS and the predicted plan of other actors. Simulation studies show that the AS can obtain human-like reasoning, exploiting interaction for necessary performance increases, while acting cautiously when faced with uncertainties to not jeopardize safety.