Robot Learning From a Human Expert Using Inverse Reinforcement Learning: A Deep Reinforcement Learning Approach for Industrial Applications

Andersen, Rasmus Eckholdt ; Hansen, Emil Blixt ; Madsen, Steffen

4. term

Manufacturing Technology, Master

2019

2019-06-03

109 pages

The need for adaptable models, e.g. reinforcement learning (RL), have in recent years been more present within the industry. However, the number of commercial solutions using RL is limited, one reason being the complexity related to the design of RL. Therefore, a method to identify complexities of RL for industrial applications is presented in this thesis. It was used on 15 applications inspired from four industrial companies. Complexity was especially identified in relation to the reward functions. Thus two Linear Inverse RL (IRL) algorithms in which the reward function is represented as a linear combination of features, was tested using expert data. Some of the tests indicated a visual better result than tests carried out using RL. The process of designing features shared similarities with the process of designing a reward function. The added complexity of implementing Linear IRL and constructing expert data is thus not always a simpler approach. The IRL method GAIL, which requires no feature construction, was furthermore tested showing potential.

reinforcement ; learning ; inverse reinforcement learning ; reinforcement learning complexity ; RLC ; ur5 ; imitation learning ; IRL ; RL ; industrial ; hand-eye calibration ; simulation ; robot ; gazebo ; TCP Simulation ; Master Thesis ; vr system ; HTC vive

Download
View record in AAU Student Projects