CONTROL SYSTEM REDUCING ENERGY CONSUMPTION OF MANIPULATOR DRIVES BASED ON REINFORCEMENT LEARNING.

Determining the optimal trajectory of a manipulator is a mathematically complex part of control due to the time-consuming determination of inverse kinematics, the complexity of which increases as the manipulator's degrees of freedom increase. The control of manipulator actuators is based on algorithms focusing on the shortest possible control time and zero position deviation. Values such as power consumption, which affect the growth of the carbon footprint, are not taken into account. For this reason, alternative control methods are being investigated, some of which can be used to determine the optimal trajectory in an uncertain and variable environment paying attention to instantaneous power consumption. This paper examines the capabilities of an artificial neural network algorithm for determining the trajectory of a four-axis manipulator with special emphasis on power consumption. The chosen method for learning artificial neural networks was reinforcement learning. Unlike the classical approach to determining inverse kinematics, the solution presented in the paper is based on model output signals. The artificial neural network determines the control policy based on the angular position, angular velocities, power consumption of the manipulator's actuators, and feedback, which is the reward function. The network was trained to reach a desired point from a fixed and random initial state of the robotic arm in optimal time and effective power requirements in such a way as to maintain a set distance from periodically moving obstacles. A three-dimensional representation of the manipulator designed in a CAD environment and mathematical models of the manipulator's drives were used for training. This approach accelerated the learning process of the algorithm, and the steps taken guarantee very similar, if not identical, performance on a real object. [ABSTRACT FROM AUTHOR]