Decentralised adaptive learning-based control of robot manipulators with unknown parameters

This paper studies motor joint control of a 4-degree-of-freedom (DoF) robotic manipulator using learning-based Adaptive Dynamic Programming (ADP) approach. The manipulator’s dynamics are modelled as an open-loop 4-link serial kinematic chain with 4 Degrees of Freedom (DoF). Decentralised optimal controllers are designed for each link using ADP approach based on a set of cost matrices and data collected from exploration trajectories. The proposed control strategy employs an off-line, off-policy iterative approach to derive four optimal control policies, one for each joint, under exploration strategies. The objective of the controller is to control the position of each joint. Simulation and experimental results show that four independent optimal controllers are found, each under similar exploration strategies, and the proposed ADP approach successfully yields optimal linear control policies despite the presence of these complexities. The experimental results conducted on the Quanser Qarm robotic platform demonstrate the effectiveness of the proposed ADP controllers in handling significant dynamic nonlinearities, such as actuation limitations, output saturation, and filter delays.