Whole Body Manipulation using Reinforcement Learning

Abstract

[eng] Recent advancements in artificial intelligence(AI) have revolutionized the field of robotics. One of the most intriguing use cases in this domain is whole-body manipulation. Whole-body manipulation combines the precision of robotic manipulators with the expanded reach of mobile platforms. This thesis explores the task of autonomous whole-body manipulation using reinforcement learning (RL). By leveraging RL’s ability to learn from experience and adapt to new scenarios, we aim to navigate and manipulate a robot jointly. First, we explore RL for navigation and manipulation separately. After developing a keen understanding of these tasks and training successful RL agents, we move towards joint navigation and manipulation. We conduct experiments using different training methods to combine these tasks under the paradigm of hierarchical RL (HRL) to achieve autonomous whole-body manipulation. The resulting RL agent is capable of successfully reaching a target location outside the operating range of the arm without collisions. In conclusion, we provide an example of the future potential of HRL for complex tasks within the domain of robotics.