Direction aware and self-adaptive A* algorithm with PPO heuristic for UAV path planning of smart city

Abstract Path planning is a fundamental component in the development of robotics, autonomous navigation, and intelligent systems, playing a pivotal role in the functioning of smart cities. Within the realm of smart cities, where infrastructure is becoming increasingly interconnected, efficient path planning algorithms are essential for optimizing traffic flow, reducing congestion, and ensuring the seamless movement of people and goods. Among various path planning algorithms, the A* algorithm remains one of the most widely used approaches due to its completeness and optimality under consistent heuristics. However, traditional A* suffers from several limitations when applied to complex 3D environments, including uniform neighbor expansion, fixed-resolution grids, and overly simplistic heuristic functions. These drawbacks often lead to excessive computation, suboptimal paths, and failure in cluttered or large-scale scenarios. To address these challenges, we propose a direction aware and self-adaptive A* algorithm named DASA*, an enhanced A* based path planning framework. Firstly, we propose an adaptive hierarchical direction-aware neighborhood selection mechanism. Driven by a lightweight stagnation detector based on open-set growth rate, this mechanism enables seamless automatic switching of neighborhood expansion from narrow conical angles to full 26-neighbourhood coverage. This fundamentally resolves the search incompleteness inherent in traditional fixed-conical-angle directional A* algorithms at the mechanism level, significantly enhancing planning success rates and efficiency. Secondly, a resolution-adaptive search strategy dynamically adjusts planning granularity based on local obstacle density. It advances rapidly with coarse granularity in sparse areas while switching to fine granularity for precise obstacle avoidance in dense zones, achieving an optimal balance between efficiency and safety in heterogeneous environments. Moreover, the designed learnable heuristic plugin interface enables seamless integration of PPO models trained on spatio-semantic features. This provides more informative, purposeful, and forward-looking heuristic guidance, further accelerating convergence towards optimal paths. Finally, we have devised a delayed path adjustment strategy that simplifies trajectories by omitting redundant waypoints through a deferred update mechanism. Subsequently, cubic B-spline interpolation is applied to smooth the path, ensuring curvature continuity and effectively preventing abrupt turns and unnatural maneuvers. This enhances both the fluidity and kinematic plausibility of the trajectory. Subsequently, Extensive experiments in simulated 3D environments demonstrate that DASA* outperforms conventional A* and Other mainstream algorithms in terms of planning time, length of optimal path, and success rate. The proposed framework provides a practical and extensible foundation for real-world applications such as UAV navigation, mobile robotics, and autonomous inspection in complex terrains.