Route Planning Using Artificial Intelligence

Abstract

Abstract : This study presents an independent Route Planning System using Artificial Intelligence that monitors small-scale environments, such as warehouses, retail stores, and car parking areas, to generate routes for ground robots. The system enables robots to navigate and avoid collisions by providing optimal paths without direct control. It employs semantic segmentation models—Fully Convolutional Network (FCN), U-Net, SegNet, and PSPNet—to transform top-view images into binary grids distinguishing free and occupied spaces. These grids are processed by route planning algorithms-A*, Dijkstra's, and Bidirectional to compute efficient routes. Implemented in Python using PyTorch, the system demonstrates robust performance across 800 training and 200 testing images from the iSAID dataset. Results show U-Net consistently outperforming other models, achieving the lowest segmentation loss (0.0699) and highest scores across multiple metrics including Pixel Accuracy (0.970), Jaccard Index (0.948), and Recall (0.993). SegNet emerged as a strong competitor, while PSPNet demonstrated notable limitations. Route planning performed most efficiently on U-Net and SegNet-generated grids, with A* achieving the fastest execution times (0.15-0.18 ms) and lowest memory usage (0.024-0.025 MB), while Bidirectional yielded the shortest paths (7.05-7.57 units). In contrast, PSPNet-generated grids led to significantly longer paths (20.8-21.8 units) and higher computational demands. The system enhances navigation efficiency, safety, and cost-effectiveness for logistics applications in small-scale environments, with U-Net and A* identified as the optimal combination for real-time performance.