Optimal and Smooth Mobile Robot Path Planning Using GAN, A*, and Cubic Spline Interpolation

Abstract

This paper presents a trajectory-planning method for mobile robots that integrates a Generative Adversarial Network (GAN) with grid-based A*. The GAN generator samples obstacle coordinates while enforcing a 1.0-unit clearance and masking forbidden regions around the start (1,1) and goal (14,14). The workspace is discretized into a 30×30, 8-connected lattice; A* with an admissible and consistent Euclidean heuristic returns the globally optimal lattice path, which is subsequently converted into a smooth geometric trajectory via cubic-spline interpolation. On a 15×15 maze with 60 obstacles, the proposed GAN&A* pipeline achieves a path length of 19.26 units, improving over a Sparrow Search Algorithm baseline (21.8 units). To assess scalability, we further evaluate a 20×20 maze with 120 obstacles. Under identical collision models and smoothing, GAN&A* attains 28.81 units, outperforming two sampling-based planners RRT (30.63 units) and PRM (30.51 units). These results indicate that learned environment synthesis coupled with optimal lattice search yields reliable, short, and smooth trajectories, whereas RRT/PRM require substantially larger sampling budgets (RRT*/PRM*) to approach comparable quality.