Motion Profile Optimization in Industrial Robots using Reinforcement Learning

Abstract

Path tracking problems are challenging with the absence of dynamic models and information about robot controllers. This paper presents a method of optimizing a motion profile constructed using a set of pre-defined motion primitives and a speed command to track a spatial trajectory with high accuracy, speed, and uniform motion using industrial robots. We use a bi-level optimization approach that optimizes execution accuracy using reinforcement learning and execution speed using bi-section search. We train and evaluate the reinforcement learning policy in simulation for an ABB robot. Experiment results demonstrate that the learned policy reduces the optimization cost to achieve the desired specifications. Additionally, the trained policy can generalize to trajectories not included in the training set.