Development of ROS-based Reinforcement Learning Environment for Single-arm Robot Training
Open Access
Author:
Kim, Jinhoo
Area of Honors:
Engineering Science
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Charles E Bakis, Thesis Supervisor Lucas Jay Passmore, Thesis Honors Advisor
Keywords:
Reinforcement Learning Robot Operating System (ROS) Single-arm Robot
Abstract:
During the COVID-19 pandemic, the demand for robots has been stronger than at any time.
However, the application of robots in the real world has been limited due to the lack of adaptability. Reinforcement learning (RL) combined with the deep neural network has shown successful results in games. As the principle of RL is to train behavior by rewarding events, robotic engineers can solve complex robotic locomotion or task problems more intuitively using RL. While OpenAI-Gym provides RL frameworks in the MuJoCo simulation environment, the robot operating system (ROS) – the most-widely used platform in robotics communities – does not support MuJoCo. Therefore, a ROS-based RL environment is developed. In this article, a 6-degree-of-freedom (6-DOF) single-arm robot Indy7 is trained to reach a randomly generated goal location without collision using the state-of-the-art RL algorithms - soft actor-critic (SAC) and hindsight experience replay (HER). This thesis demonstrates from the basic hardware/software setup to ROS RL environment development to SAC and HER implementation to the results.
