A deep reinforcement learning-based optimization method for vibration suppression of articulated robots

摘要

To solve the vibration problem of articulated robots, a deep reinforcement learning-based optimization method for input shaping (RLIS) is proposed. The optimization method applies the arbitrary time-delay filter as the model and includes two main stages: determining the RLIS initial parameter ranges; and constructing and improving the RLIS optimization strategy. In the first stage, a post-adaptive method based on the recursive least squares algorithm is used to determine the RLIS initial parameter ranges. In the second stage, the input shaping optimization problem is modelled as a Markov decision process, and the RLIS optimization strategy is constructed using a deep reinforcement learning algorithm. Subsequently, a selection mechanism based on state value and a fuzzy reward system is proposed to increase the learning efficiency and simplify the design process. Finally, several groups of experiments are designed to demonstrate the effectiveness and stableness of the proposed method in residual vibration suppression.

关键词