Reward Function Analysis for Deep Reinforcement Learning Agent in Goal-Oriented Navigation

Designing a reward function is never a trivial task when implementing Reinforcement Learning (RL) agents. Instead, it is one of the most crucial steps for ensuring a stable and robust learner. This paper explores the impact of different reward function designs on the performance of a Deep Reinforcement Learning (DRL) continuous control model. We investigate three distinct classes of reward functions: “Binary Class”, “Continuous Class” and “Sparse Class”, implemented to navigate a differential drive robot through unknown environments toward a designated goal without prior knowledge or pre-made maps, while avoiding obstacles. The designed reward functions are tested in two increasingly complex environments using the Gazebo simulator integrated with the Robot Operating System (ROS). Our study conducts a comparative analysis of these reward structures within the DRL framework to evaluate their influence on learning efficiency, reward variance, and agent stability. The findings reveal significant differences in learning behaviors and performance metrics across the reward types, providing insights into optimal reward structuring for enhancing DRL implementations in robotics. This work not only contributes to the theoretical understanding of reward design in RL but also offers practical guidance for developing more effective DRL systems in robotic applications.