PS-CC: Planning by Simulation With Continuous Action Space Optimization and Adaptive Cost Learning for Human-Like Autonomous Driving

Abstract

Ensuring safe and human-like decision-making is a critical component of autonomous vehicle decision systems. However, conventional approaches often simplify the action space into discrete domains for policy formulation and rely on handcrafted cost functions for policy evaluation, which limits their adaptability and realism. In this study, we propose a novel Monte Carlo Tree Search (MCTS)-based planner integrated with cost learning in continuous action space. Specifically, we employ kernel regression (KR) and Bayesian optimization to extend action sampling into continuous space, enabling more nuanced decision-making. Additionally, we introduce parallel backpropagation to facilitate human-like pruning during the search process. Furthermore, we advance the traditional static cost function by optimizing it into a learnable, context-aware cost function, incorporating interactive trajectory prediction for enhanced adaptability. The proposed framework is rigorously validated using the real-world Argoverse 2 dataset within the MetaDrive simulation environment. Experimental results demonstrate that our approach significantly reduces information loss, accelerates the search process, and enhances adaptability by dynamically evaluating costs based on contextual scenarios, ultimately leading to safer and more human-like driving behaviors.