Publication Date

Summer 2021

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Engineering

Advisor

Robert Morelos-Zaragoza

Keywords

deep reinforcement learning, human-like operations, robotics, six DoF, Unity 3D

Subject Areas

Electrical engineering; Robotics; Artificial intelligence

Abstract

The objectives of this research are to investigate, design, develop, and implement a DRL-based reward function enhanced robotics technique to achieve ”human-like” operations for a six DoF robot. Technical challenges for robotics applications arise nowadays with a need to achieve human-like intelligence. DRL can formulate human-likeintelligence by maximizing a long-term reward function. However, the design of states, actions, policies, and reward functions having continuous values turns out to be quite tricky. Lastly, to reach a target position, control motion behavior without calculating inverse kinematics, and update state space for actions and policy optimizations are also challenging. Inspired by Google Deepmind’s published research, we propose utilizing of SAC technique to solve the above challenges. In this research, we introduce gain factors in the baseline reward function for enhanced performance of the robotics operations. We integrated and developed algorithms and software tools on a state-of-the-art Unity 3D platform for mathematical modeling and simulations. The experiments consisting of reaching the object are conducted in a simulated environment a few times. We plot training curves and conclude that our reward function is better than the baseline reward function. We note an improvement of about 259:5% over the baseline reward function.

Share

COinS