Knowledge-based Implementation of Deep Reinforcement Learning Agents in Assembly

Robotic systems based on Deep Reinforcement Learning have shown great potential to enable assembly systems with higher flexibility and robustness. This paper presents a concept of a Case-Based Reasoning system to automate the implementation process, based on the assumption that similar assembly tasks have similar solutions as used as heuristics in the current manual procedure. For retrieving similar cases a digital description of the assembly task and a method to measure the similarity is introduced. The retrieved cases are then used to warmstart a Bayesian Hyperparameter Optimization. The approach is evaluated on two simulated robot task.