An Online Model-Free Reinforcement Learning Approach for 6-DOF Robot Manipulators

Abstract

Controlling 6 Degrees-of-Freedom (DoF) robotic manipulators in an online, model-free manner poses significant challenges due to their complex coupling, non-linearities, and the need to account for unmodeled dynamics. This paper introduces a model-free adaptive approach for real-time control of a 6 DoF “EPSON” robotic manipulator, without requiring any prior knowledge of the manipulator’s dynamics. Initially, we lay out the framework for an optimal control solution. A performance index is introduced, leveraging error dynamics and correction control signals, offering the capability to incorporate high-order error dynamics without the need to explicitly derive error trajectories. The order of error dynamics is determined by the chosen number of error samples. We assume a kernel-based solution structure aligning with the performance index, resulting in a temporal difference equation. This equation can be optimized to formulate a model-free control strategy. Subsequently, a reinforcement learning approach is adopted to approximate the underlying strategy. Infeasible exact solutions are overcome by employing a value iteration mechanism to adapt the actor-critic structures within an adaptive critics framework. To validate the proposed approach, it is compared against a conventional proportional-integral controller. A Unified Robot Description Format file is generated to facilitate the import of the robotic manipulator into the MATLAB Simulink environment, enabling its control. Ultimately, the proposed method yields superior results in terms of the dynamic characteristics of the response, demonstrating its effectiveness over the conventional approach.