A Belief State Planner for Interactive Merge Maneuvers in Congested Traffic

Abstract

Autonomous driving in urban environments requires the capability of merging into narrow gaps. In cases of high traffic density this becomes more complex since one must consider the interaction with other vehicles. We formulate the problem as a Partially Observable Markov Decision Process (POMDP) by including the surrounding drivers in the state space to realize interactive behavior. The problem is solved online by an anytime Monte Carlo sampling algorithm in combination with an efficient A* rollout heuristic. This combination makes a combined lateral and longitudinal optimization possible. The resulting policy is optimized regarding various future merge scenarios and approaches the most suitable gap while taking into account the uncertain behavior of the surrounding drivers. Therefore, we present a novel motion model representing the uncertain cooperation of other drivers. It is based on a logistic regression model estimating the probability for cooperative behavior of a human driver given a future scene. We demonstrate the performance of our algorithm by various simulated scenarios. The resulting behavior for approaching gaps, estimating the future cooperative behavior of surrounding drivers and performing merges in narrow gaps are discussed.