A Visual Navigation Algorithm for Small Autonomous Robots Based on Lightweight Convolutional Neural Networks

Di Wu

Autonomous mobile robots operating in compact or resource-constrained environments increasingly rely on visual perception for safe and efficient navigation. However, conventional vision-based algorithms often depend on computationally intensive neural networks that exceed the processing capabilities of small robots equipped with low-power embedded hardware. To overcome this challenge, this paper presents a visual navigation algorithm based on a lightweight convolutional neural network (CNN) specifically designed for small autonomous robots. The proposed approach integrates scene understanding, obstacle perception, and motion command generation within a compact end-to-end framework optimized for real-time computation. A streamlined feature-extraction backbone and lightweight decision module are introduced to minimize computational overhead while maintaining effective spatial perception. Experimental evaluations conducted across diverse indoor environments indicate that the algorithm achieves stable navigation performance and demonstrates robustness to illumination changes and moderate dynamic disturbances. Trend-level comparisons with conventional CNN-based navigation methods show a clear reduction in computational demand while retaining comparable navigation accuracy, highlighting the suitability of the proposed method for embedded robotic applications.