Improving the Learning Efficiency of Robot Path Planning via Fuzzified Deep Reinforcement Learning
Suaib Al Mahmud, Azhar Mohd Ibrahim, Mazbahur Rahman Khan, Odai Saeed Baqalaql, Philippe De Wilde · 2025
Mobile robot navigation is a significant research challenge, especially in scenarios where efficient learning is crucial. Deep reinforcement learning (DRL) offers a promising alternative to traditional control strategies, yet its effectiveness is often limited by sparse reward structures and simple episode termination criteria. This work proposes two modifications to enhance DRL-based path planning. First, we revise the episode termination process to ensure that each robot accumulates rewards until it reaches its individual terminal condition, thereby preventing premature episode endings and promoting improved learning. Second-and most importantly for this study-we introduce a reward shaping mechanism that leverages fuzzy logic and the Adaptive Neuro-Fuzzy Inference System (ANFIS) to produce continuous, context-aware sub-rewards that provide detailed performance feedback to each agent throughout its navigation task. Simulation results show that the revised episode termination significantly improves learning efficiency, while the fuzzy logic-based reward shaping notably enhances reward accumulation and overall performance. The fuzzy logic method effectively encodes the knowledge and manages uncertainty, offering a flexible and interpretable framework for decisionmaking in complex environments. These findings demonstrate the potential of incorporating fuzzy logic into DRL frameworks to achieve faster convergence and superior navigation performance in mobile robot systems.