Dynamic Rewards in Reinforcement Learning for Robotic Navigation

Abstract

The paper presents a new reinforcement learning method, called Adaptive Q-learning with Dynamic Reward (AQDR), for efficient route planning of mobile robots operating in a partially known and unknown environment. Traditional Q-learning techniques are often limited in their adaptability due to slow convergence in dynamic environments. To overcome these limitations, AQDR combines an adaptive reward mechanism that adjusts in real time based on the distance of the robot to the obstacle and the target position. This dynamic feedback allows for better informed decision making and reduces unnecessary exploration. The proposed algorithm was evaluated against the Q-learning method based on static rewards by performing simulated experiments in different environmental configurations. The results show that AQDR consistently exceeds the baseline in terms of convergence speed, path efficiency, and adaptability. These results highlight the potential of dynamic reward design to improve learning performance and robustness of reinforcement learning navigation systems.