Vision-Based Autonomous Navigation and Obstacle Avoidance in Mobile Robots Using Deep Reinforcement Learning

Abstract

This paper presents a hybrid autonomous navigation system for mobile robots that integrates vision-based deep reinforcement learning (DRL) with a visual simultaneous localization and mapping (SLAM) module. The proposed framework employs a convolutional neural network and a Proximal Policy Optimization (PPO) algorithm to learn control policies from raw RGB images, enabling end-to-end navigation and obstacle avoidance. To enhance localization accuracy and robustness, we incorporate ORB-SLAM2 as a geometric localization backbone, providing real-time pose feedback to the policy network. Extensive experiments in both simulated and real-world environments demonstrate that the combined DRL+SLAM architecture outperforms classical and vision-only navigation baselines in terms of success rate, path efficiency, and collision avoidance. The results highlight the benefit of fusing learned perception and geometry-based reasoning to achieve robust and generalizable robot autonomy in complex indoor environments.