Yogiek Indra Kurniawan; Siti Shofiah; Rosalina Yani Widiastuti; Teguh Arifianto; Ribut Julianto
Background: The rapid growth of warehouse automation and autonomous mobile robots has increased the need for adaptive navigation systems capable of operating safely and efficiently in dynamic industrial environments. Classical path planning algorithms such as A* and RRT perform well in structured settings but exhibit limitations when handling moving obstacles and environmental uncertainty. Objective: This study aims to develop and evaluate a reinforcement learning based navigation framework integrated with sensor fusion to improve path efficiency, collision avoidance, and robustness in dynamic warehouse scenarios. Method: An experimental research design was implemented combining high-fidelity simulation and real-world warehouse prototype testing. Deep Q-Network and Proximal Policy Optimization models were developed and trained using multi-sensor inputs from LiDAR, camera, and inertial measurement units. Performance was evaluated using path efficiency, collision rate, computational cost, and robustness metrics, with benchmarking against classical algorithms. Results: The results demonstrate that the Proximal Policy Optimization model achieved the highest path efficiency and lowest collision rate while maintaining stable computational performance under dynamic conditions. Reinforcement learning models significantly outperformed classical planners in adaptability and robustness, confirming their suitability for scalable industrial warehouse automation.