Deep Reinforcement Learning for Robot Batching Optimization and Flow Control

摘要

Robot batching is an optimization problem found in many industrial applications. Current state-of-the-art approaches utilize a combination of heuristic based parameters and statistical analysis. This approach necessitates many tunable parameters, which again provides challenges when delivering systems to new customers. We challenge current state-of-the-art in statistical approaches by presenting a novel application of a policy gradient method for a Deep Reinforcement Learning (DRL/RL) agent. We have developed a Unity simulation framework of an existing robot-batching cell, on which a RL agent is able to successfully train and obtain a policy for performing robot batching, using a tabula rasa approach. The trained agent is capable of packaging 47.86% of 1218 total batches within the prescribed tolerances, with a positive give-away of 8.76%. The application of DRL in performing robot batching is to the authors knowledge the first of its kind.