Research on Integrated Road Dynamic Recognition and Path Planning of Intelligent Vehicles Based on Reinforcement Learning

Abstract

The integrated research on intelligent vehicles in road dynamic recognition and path planning is based on reinforcement learning. In complex urban scenarios, intelligent vehicles face the core bottleneck problem of "disconnection between perception and decision-making, and lagging response". This paper focuses on this issue and proposes an integrated framework of "lightweight reinforcement learning path planning integrating multi-modal perception". Before that, it first analyzes the current research status and mainstream algorithms of intelligent vehicle path planning. Thus, the limitations of the current separated architecture in highly dynamic traffic environments are pointed out. In response to this limitation, this paper designs a synchronous-alignment module that can be trained end-to-end, enabling multi-source heterogeneous data such as cameras, lidars, and millimeter-wave radars to be efficiently compressed into a unified and compact state vector while retaining key interaction information. In addition, a reinforcement learning state-action-reward space for dynamic traffic flow was constructed, and the dual-objective reward function of safety and efficiency was embedded in it, thereby achieving online collaborative optimization of road dynamic recognition and path planning. The algorithm development was completed on mainstream deep learning frameworks, and then the system was verified on the CARLA/SUMO high-fidelity simulation platform and in real vehicle environments. The experimental results show that the proposed scheme is much better than the traditional separated A*/Dijkstra and single-modal RL baseline in core indicators such as navigation success rate, average travel time, and collision rate. Moreover, it reduces the number of model parameters by 42% and keeps the single-step reasoning delay within 50ms. It fully meets the dual requirements of real-time performance and robustness in the complex urban environment. This research achievement not only provides a highly adaptable and easily deployable end-to-end navigation solution for intelligent vehicles, but also offers theoretical basis and engineering practice paradigms for the deep integration of multimodal perception and reinforcement learning in the field of autonomous driving.