نيلوفر تك زارع

This research proposes an integrated approach in the framework of deep reinforcement learning for real-time navigation of wheeled mobile robots in unknown spaces. At the core of this method, the PER–MS–DDQN algorithm is designed by integrating the dual structure of DDQN an‎d prioritized replay memory an‎d multi-stage scaling to prevent overestimation of Q values. In order to enhance the decision-making efficiency, a new reward function is introduced that extracts static sub-goals by using the A-star Also, to reduce the computational complexity an‎d increase the training speed, special distribution data was used in some inputs an‎d specific weights were determined for another set of features.The multilayer observation structure, angular discretization of the environment, an‎d integration of lidar inputs provide a rich an‎d accurate model for situation an‎d obstacle perception. Experimental eva‎luations were conducted in two stages—first in a 2D simulation an‎d then in a 3D Gazebo environment with ROS 2—on the Berger Turtlebot. The results showed that after approximately 3,300 episodes, the average reward reached 345 units (23% higher than the reference 280 units); in the final 500 episodes, the average reward was 300.8 units (an improvement of up to 12 times compared to the baseline DDQN), the reward fluctuations were reduced by 40%, an‎d the success rate in Gazebo increased to 99%. These achievements demonstrate the significant improvement in the convergence, stability, an‎d generalizability of PER–MS–DDQN in dynamic environments an‎d open new horizons for practical applications in robotic navigation an‎d autonomous systems

يادگيري تقويتي – تابع پاداش – ناوبري – ربات متحرك- حالت

Reinforcement learning, Reward function, Navigation, Mobile robot, States

Niloufar Takzare

Dr. Moharam Habibnejad Korayem