A Comparative Analysis of Reinforcement Learning-Based Navigation for Autonomous Mobile Robot

Abstract

Mobile robots have been widely used in many industries including manufacturing, healthcare and warehouse automation. To ensure efficiency and safety of the robots, it is crucial to design effective control strategies that can adapt to changing environments. This paper presents reinforcement learning (RL) algorithms including Q-learning, Deep Q-Learning (DQN), and Double Deep Q-Learning (DDQN) for autonomous navigation using the Turtle-Bot3 Waffle Pi in a Gazebo-simulated environment. Three progressively complex training stages were designed to evaluate the algorithms: (1) static obstacles with predefined goals, (2) randomized goals with static obstacles, and (3) dynamic obstacles with moving goals. Performance metrics, including success rates, collision avoidance, and reward stability, were analyzed to compare algorithm effectiveness. Key results highlight DDQN’s superiority in handling complex navigation tasks. In the most challenging stage, DDQN achieved a 100% success rate and zero collisions, outperforming DQN, which attained an 88% success rate with higher collision rates. Q-learning performed well only in simple environments, as it cannot easily handle continuous state spaces. This study demonstrates the scalability of RL-based navigation systems for autonomous mobile robots. The findings provide a foundation for future advancements in dynamic and real-world robot navigation.