فهرست نشریات علمی دسترسی آزاد ایران

Path planning with avoiding obstacles autonomously with a large of computing capabilities in an unknown dynamic environment is a difficult challenge for a mobile robot to solve. This research solves this challenge by combining deep Q-network (DQN) with cloud computing. To begin, a DQN is created and trained to predict the state-action value function of a mobile robot. The information collected from the original RGB image (pixels in the image) taken from the surrounding is fed into the DQN using a cloud computing platform, which reduces the algorithms high computation complexity; Finally, the action chosen policy picks the current optimal mobile robot action. To validate the DQN algorithm, we trained the robot in a dynamic environment with a simple and complex case. The simulation results show that, in a simple case of the environment, the DQN technique can converge to explore a path with fewer steps and higher average reward than in a complicated case and find a collision-free path with an accuracy rate of 89\% in the simple case and when the environment becomes more complex, the accuracy rate is 70 %.