Reinforcement learning–based design of orienting devices for vibratory bowl feeders

The sorting of small parts is one of the common tasks in the field of industrial manufacturing. Vibratory bowl feeders (VBF) are commonly used to accomplish this task. Nowadays, the design process of the VBF is based on a manual and expensive trial-and-error approach, in which different traps are arranged and tuned. This paper outlines a method which modifies this conventional process using Reinforcement Learning to automate the VBF design. To enable this, a software agent is used to model the placement of traps on multiple positions and measure the subsequent configuration efficiency. A physics simulation provides the characteristics of the individual traps. During the training, Q-learning is applied to determine the environmental indicators under which a certain trap should be replaced. A 3D matrix is used to store information in a problem-related representation. Due to the trial-and-error principle of Reinforcement Learning, this training is comparable with the traditional proceedings. In addition, valuable action paths are stored in a memory and the agent frequently is trained on these paths in order to remember good solutions. Additionally, a knowledge base is used to exclude inefficient sets of traps. The rules for the knowledge base are built upon knowledge from the conventional design process. In first test cases, the trained agent is able to assemble traps achieving promising configurations. The results of the agent will be validated in the next step using physics simulation.