Robotic Manipulator Control Using CNN and Deep Q-Network Algorithm
Abdullah Al-Zabt, Tarek A. Tutunji · International Review of Automatic Control (IREACO) · 2022
Convolution Neural Networks are used with a Reinforcement Learning algorithm to control a robotic arm. The objective is to move the end-effector to a desired position using minimum actions without prior knowledge of the robot’s structure nor its mathematical model. The proposed methodology processes camera images from the environment and inputs this information as raw pixels to the agent. Then, Deep Q-Network algorithm trains the agent to provide the appropriate actions based on a value function that estimates future rewards. An experience replay feature is used to improve the agent’s performance by memorizing the highest rewarded-sequence of actions episode and replays it when needed. Simulation runs, using the graphical robotic simulator CoppeliaSim, validate the proposed algorithm as the agent provides the optimal number of actions to move the robot’s end-effector to the desired target using camera images as inputs. Four different cases are presented and the results show that the proposed algorithm can reduce the number of required actions significantly.