Transferring Deep Reinforcement Learning from a Game Engine Simulation for Robots

Lillelund, Christoffer Bredo

4. term

Medialogy, Master

2018

2018-05-31

30

This thesis examines whether image-based reinforcement learning can be trained in a realistically rendered game engine simulation and transferred to a physical robot. Many existing robot simulators lack realistic graphics, despite the importance of image data for tasks such as autonomous driving, detection, and grasping. We propose using modern game engines like Unity3D to simulate a Turtlebot2 with an RGB camera, connect the simulation to ROS in real time, and train a Deep Q-learning model to find and drive into a blue ball. The setup makes the simulated and real robots interchangeable so the same model and communication can control both. In simulation the robot completed the task, and the trained model could control the physical Turtlebot2, but it did not accomplish the task in the real environment. These results suggest game engines are a promising option for image-based reinforcement learning in robotics, while more realistic and varied simulated environments are needed to improve transfer from simulation to reality.

Dette speciale undersøger, om billedbaseret forstærkningslæring kan trænes i en spil-motors realistiske simulation og overføres til en fysisk robot. Problemet er, at mange eksisterende robotsimulatorer ikke kan gengive realistiske billeder, selv om sådanne data er vigtige i opgaver som autonom kørsel, genkendelse og gribning. Vi foreslår at bruge moderne spil-motorer som Unity3D til at simulere en Turtlebot2 med RGB-kamera, forbinde simulationen i realtid til ROS og træne en Deep Q-learning-model til at finde og køre ind i en blå bold. Designet gør den simulerede og den virkelige robot udskiftelige, så den samme model og kommunikation kan styre begge. I simulationen løste robotten opgaven, og den trænede model kunne kontrollere den fysiske Turtlebot2, men den fuldførte ikke opgaven i det virkelige miljø. Resultaterne indikerer, at spil-motorer er lovende som realistiske robotsimulatorer for billedbaseret forstærkningslæring, men at mere realistiske og varierede simuleringsmiljøer bør afprøves for at forbedre overførsel fra simulation til virkelighed.

[This apstract has been generated with the help of AI directly from the project full text]

Machine learning ; Reinforcement learning ; ROS ; Robots ; Turtlebot2 ; Unity3D ; Simulation

Download PDF
View record in AAU Student Projects