The Study of a Textile Punching Robot Based on Combined Deep Reinforcement Learning

Abstract

Punching of textile uppers is an important process in the shoe production line, which determines the accuracy of positioning and affects all subsequent shoemaking steps. At present, there are few researches on the 3-DOF punching robot in this field. The critical value of the punching force of the end mechanism is not clear, and the punching path planning is not adopted with a suitable algorithm, resulting in large power consumption and low work efficiency. Therefore, this paper describes the design of a punching robot to meet the action requirements of punching the textile uppers and complete the path planning to improve efficiency. The first step is to model a 3-DOF robot for textile upper punching including rail type and articulated type to meet the needs of punching movements. The second step is to gain the lowest critical value of the pressure with ANSYS by simulating deformation process when the punch punches in the vamp until breaks it. And this outcome is the theory basis for motor selection and robot design. The third step is multi-point punching path planning, where a combined deep reinforcement learning (DRL) method is proposed. The DRL pre-training is performed through asynchronous advantage actor-critic (A3C). Training data is adopted to optimize the recurrent neural networks (RNN) that parameterizes the stochastic policy. Then the inspire planning (IP) algorithm is conducted to get the optimal path. The experimental results show that this method can jump out of the local optimal solution and outperform other state-of-the-art methods.