Intention-Guided Heuristic Partially Observable Monte Carlo Planning for Off-Ramp Decision-Making of Autonomous Vehicles

The Partially Observable Monte Carlo Planning (POMCP) leverages Monte Carlo Tree Search (MCTS) and Particle Filtering (PF) to enhance the computational efficiency in solving large-scale Partially Observable Markov Decision Processes (POMDPs), allowing for updates of the belief state and effective adaptation to evolving uncertainties, which has been widely studied in autonomous driving. However, this approach faces two limitations when applied to planning for autonomous vehicles: chaotic branch expansion in the belief tree reduces computational efficiency, and particle deprivation hinders the accurate estimation of the dynamic intentions of surrounding vehicles. To this end, an intention-guided Partially Observable Monte Carlo Planning with a Heuristic-based Double Progressive Widening (POMCP-HDPW) approach is proposed to facilitate efficient decision-making for autonomous vehicles. We propose an enhance resampling method of PF that accounts for the driving intentions of surrounding vehicles, maintaining particle diversity and thereby improving estimation accuracy. Additionally, we prune the action and observation spaces by leveraging human driving experience and collision risk assessment, enabling the expansion and exploration of high-value belief nodes and preventing chaotic expansion. Three different methods are employed to drive the motion of surrounding vehicles, validating the robustness of the proposed model: intelligent driving model control, offline driving using the exiD trajectories, and driver-in-the-loop validation. Notably, experimental results on the exiD dataset demonstrate a success rate of 96.88% in off-ramp scenarios.