Interaction-Aware Motion Planning for Autonomous Vehicles via Estimated HMMs of Road User Decisions

Preprint, 2026

In this paper, we propose an interaction-aware predictive motion planning framework for an autonomous vehicle (AV) that must avoid collisions with stochastic and interactive human-driven vehicles (HDVs). We first model the input decision formation in a network of HDV agents as a hidden Markov model (HMM), in which the joint agent input is approximated as a multivariate Gaussian. The distribution parameters are determined by the hidden Markov chain state representing the agent decision configuration in the network. The interaction-aware model predictive controller then determines an optimal AV control input by constrained optimization over the most probable future scenarios, resulting from input ranges specified by the Gaussian parameters of state sequences predicted by a scenario tree. By constraining the AV according to the input bounds of one agent, the controller helps maintain the accuracy of the HMM’s joint input predictions. The framework is evaluated in a case study of a traffic intersection that explores how learning different interactive HDV behaviors affects an AV’s decision to pass or yield. The proposed interaction-aware controller produced feasible solutions in 93% of simulations, whereas the feasibility of comparative baseline controllers, which do not take interaction into account, was 77%.