Authors:
Vladimir Samsonov
1
;
Chrismarie Enslin
1
;
Hans-Georg Köpken
2
;
Schirin Baer
2
and
Daniel Lütticke
1
Affiliations:
1
Institute of Information Management in Mechanical Engineering, RWTH Aachen University, Aachen, Germany
;
2
Siemens AG, Digital Factory Division, Nuernberg, Germany
Keyword(s):
Reinforcement Learning, Soft Actor-Critic, Supervised Learning, Industrial Manufacturing, Process Optimisation, Machine Tool Optimisation.
Abstract:
Modern manufacturing is increasingly data-driven. Yet there are a number of applications traditionally performed by humans because of their capabilities to think analytically, learn from previous experience and adapt. With the appearance of Deep Reinforcement Learning (RL) many of these applications can be partly or completely automated. In this paper we aim at finding an optimal clamping position for a workpiece (WP) with the help of deep RL. Traditionally, a human expert chooses a clamping position that leads to an efficient, high quality machining without axis limit violations or collisions. This decision is hard to automate because of the variety of WP geometries and possible ways to manufacture them. We investigate whether the use of RL can aid in finding a near-optimal WP clamping position, even for unseen WPs during training. We develop a use case representing a simplified problem of clamping position optimisation, formalise it as a Markov Decision Process (MDP) and conduct a
number of RL experiments to demonstrate the applicability of the approach in terms of training stability and quality of the solutions. First evaluations of the concept demonstrate the capability of a trained RL agent to find a near-optimal clamping position for an unseen WP with a small number of iterations required.
(More)