Challenging Conventions Towards Reliable Robot Navigation Using Deep Reinforcement Learning

Effective indoor navigation in the presence of dynamic obstacles is crucial for mobile robots. Previous research on deep reinforcement learning (DRL) for robot navigation has primarily focused on expanding neural network (NN) architectures and optimizing hardware setups. However, the impact of other critical factors, such as backward motion enablement, frame stacking buffer size, and the design of the behavioral reward function, on DRL-based navigation remains relatively unexplored. To address this gap, we present a comprehensive analysis of these elements and their effects on the navigation capabilities of a DRL-controlled mobile robot. In our study, we developed a mobile robot platform and a Robot Operating System (ROS) 2-based DRL navigation stack. Through extensive simulations and real-world experiments, we demonstrated the impact of these factors on the navigation of mobile robots. Our findings reveal that our proposed agent achieves state-of-the-art performance in terms of navigation accuracy and efficiency. Notably, we identified the significance of backward motion enablement and a carefully designed behavioral reward function in enhancing the robot’s navigation abilities. The insights gained from this research contribute to advancing the field of DRL-based robot navigation by uncovering the influence of crucial elements and providing valuable guidelines for designing robust navigation systems.