Collaborative Multi-Agent Reinforcement Learning Control of Parallel Robots

In this paper, a novel algorithm based on Multi-Agent Reinforcement Learning for controlling parallel robots has been suggested. The dynamic models of parallel robots are complex and full of uncertainties, and deriving them requires deep knowledge of the mechanism of the robot. Therefore, the proposed algorithm is designed model-free to be independent of prior knowledge about the system from the outset. Moreover, this algorithm comprises two primary components, making it efficient in training and convergence. The proposed algorithm takes each loop or limb in parallel robots as a separate agent. These agents then learn to collaborate to fulfill the robot’s defined task by producing appropriate control signals from a decentralized point of view. For studying the performance of the proposed algorithm, a 3-DOF parallel robot called Agile Eye is taken into account as a case study which is simulated in CoppeliaSim simulation environment for the task of reference tracking. Two other controllers, including the classic Proportional Integral Derivative (PID) controller and the single-agent counterpart of the suggested algorithm, have been implemented for a better performance comparison of the proposed algorithm. Using the Root Mean Square Error (RMSE) index, the recommended algorithm with an RMSE value of 0.0553 is superior to its single-agent counterpart with an RMSE of 0.1105. On the other hand, the proposed algorithm is inferior to the PID controller with an RMSE of 0.0275, mainly due to the fact that the PID Controller is in velocity control mode, while the proposed algorithm manipulates the robot in torque control mode, which is less stable.