Autonomous Parking via Kinematics-Aware Hybrid A* Path Planning and Learning-Driven Trajectory Control

Advances in autonomous driving have made automatic parking systems a practical means of mitigating urban congestion. This work investigates how an autonomous parking system can consistently reconcile global, kinematically feasible trajectory generation with locally robust, smooth, realtime feedback control under dynamic disturbances. To address this question, we present a closed-loop framework that coordinates Hybrid A* planning, Bézier–B-spline smoothing, reinforcement-learning control (PPO/DDPG), and geneticalgorithm-based global refinement to improve parking accuracy and efficiency. Hybrid A* first produces a kinematically feasible path; PPO subsequently adapts the control policy in dynamic environments, and DDPG further refines steering and throttle commands to achieve smoother motion. The GA periodically revises the global trajectory and the associated control parameters, reducing the risk of convergence to suboptimal solutions. A Bézier–B-spline hybrid smoother is finally applied to eliminate residual curvature discontinuities. Experimental evaluations indicate that the proposed system attains a 98.7% parking success rate, an average curvature variation of 15.3°Bém −1 , and a mean completion time of 2.1 s—substantially surpassing state-of-the-art baselines. These findings suggest that explicitly coordinating structured planning with learning-based feedback control is a practical route to reliable, real-time autonomous parking.