Master's lecture in Mechanical Engineering - Michael Danner
Veröld - Hús Vigdísar
Room 227
Master's student: Michael Danner
Title: QBot: Quadrupedal Ambulation via Reinforcement Learning
___________________________________________
Faculty: Faculty of Industrial Engineering, Mechanical Engineering and Computer Science
Advisor: Magnús Þór Jónsson, Professor
Also in the masters committee: Steinn Guðmundsson, Associate Professor
Examiner: Tómas Philip Rúnarsson, Professor
Abstract
Ambulation is a valuable form of locomotion for robots which must operate in spaces designed for human foot traffic or over uneven terrain. However, traditional approaches to robotic ambulation are laborious to implement. Recent advances in deep reinforcement learning have made it a promising alternative, but previous attempts have relied on detailed physics engine modeling for training in simulation. This project has developed a program which can teach a quadrupedal robot how to walk in real time regardless of its dimensions or configuration. The strategy presented in this project can be applied to larger, more robust quadrupeds which might serve some practical purpose. Additionally, a new reinforcement learning algorithm was discovered in the course of this research which may find applications across a wide variety of reinforcement learning problems.