Content area
Full text
Abstract: This paper provides a comprehensive introduction to Reinforcement Learning (RL), summarizes recent developments that showed remarkable success, and discusses their potential implications for the field of robotics. RL is a promising approach to develop hard-to-engineer adaptive solutions for complex and diverse robotic tasks. In this paper RL core elements are reviewed, existing frameworks are presented, and main issues that are limiting the application of RL for real-world robotics, such as sample inefficiency, transfer learning, generalization, and reproducibility are discussed. Multiple research efforts are currently being directed towards closing the sim-to-real gap and accomplish more efficient policy transfers methods, making the agents/robots learn much faster and more efficiently. The focus of this work is to itemize the various approaches and algorithms that center around the application of RL in robotics. Finally, an overview of the current state-of-the-art RL methods is presented, along with the potential challenges, future possibilities, and potential development directions.
Key words: reinforcement learning, robot arm, robotic vision, manipulation tasks.
1.INTRODUCTION
Reinforcement Learning (RL) has attracted a lot of attention in recent years with breakthroughs in multiple domains, including robotics.
Industry 4.0 is characterized by modularity, interoperability, and real-time capabilities. To address the custom manufacturing demands, RL is a key catalyst for turning an industrial robot which is designed for a fixed and repetitive task into a 'smart manipulator', having the capability to learn and perform a desired task without any explicit task-specific controller. For that to happen, present-day controllers must be augmented with learningbased solutions for motion planning, trajectory tracking control, collision avoidance, force control, and robotic vision.
Recent studies show exciting progress for RL in robotics. Deep Learning (DL) has been applied successfully to many important problem areas, including computer vision, robotics and RL. The current challenges entail solutions for scaling up to complex tasks for robots, designing robust policy representations, and optimizing the computing time.
Precise, collision-free, trajectory tracking with optimal control has been an active research area where progress is needed to a great extent [1, 2, 3]. A vast amount of research emerged for tracking applications using manipulator arms, from simple tasks such as pick-andplace [4], peg-in-hole [5], cloth folding [6], to more complex tasks, such as tracking a trajectory on an irregular surface in a multi-robot...