Hybrid Reinforcement Learning-based approach for agent motion control

The complexity of modern robotic systems poses a challenge to their control due to their dynamic properties and the nonlinear effects they present. The deployment of these systems in changing uncontrolled environments is one of the main focuses of research in the field of study. The deployment of Reinforcement Learning-based algorithms presents a very promising solution for the modelation of a robotic agents´ interaction with its environment. This paper presents an upgrade and an enhancement to the novel approach for the alternative motion control approach for robotic manipulators presented in [1] applied to an exemplar 2DOF robotic manipulator. The enhancement is achieved through a partial decoupling of the manipulators axis. A comparison of its performance with results achieved with a manipulator with fully coupled and one with fully decoupled axis is presented. The introduced hybrid approach provides an excellent compromise between the computational effort linked to the convergency speed of the algorithm and the quality of the gained solutions set by the time needed to complete tasks.