Multi-objective incremental path planning with learning-guided sampling and heading-constrained expansion for autonomous vehicles in dynamic occupancy grid environments

Path planning for autonomous vehicles in dynamic grid environments must balance reliability, efficiency, smoothness, safety, and rapid recovery after map changes. We propose LKSD-PRRT*, a modular framework built on APF-PRRT* that integrates learning-guided sampling (LG), multi-objective incremental rewiring (KMO), three-stage smoothing (SMO), and dynamic path repair (DPR). The contribution is a modular integration of learned sampling, Pareto-aware rewiring, geometrically screened smoothing, and tree-based repair, evaluated through ablations and external comparisons. LG combines a baseline-trained heatmap, geometric priors, and random exploration; KMO evaluates path length, curvature, an energy-consumption proxy, and obstacle safety during rewiring; SMO improves geometric continuity with collision checks and quality screening; and DPR reuses valid tree structures after local map updates. In independent-test static ablations (90 runs per configuration), the complete static method achieved a 100% planning success rate and, relative to APF-PRRT*, reduced path length, arc-length mean absolute curvature, and total turning angle by 4.88%, 46.44%, and 49.04%, respectively. Pure search time also increased by 260.31%. In dynamic tests, adding DPR to the LG+KMO backbone increased dynamic path validity from 84.44 to 93.33% and reduced mean recovery time and full task life-cycle time by 77.25% and 71.50%, respectively. The framework provides an interpretable trade-off among path quality, safety, and dynamic responsiveness in two-dimensional geometric planning. Complete static planning averaged 0.2640 s. Vehicle dynamics and closed-loop tracking remain untested.