Introduction: Enhancing Robotic Capabilities
Good hand-eye coordination is essential for robotic systems, most notably in the context of engaging with objects in tasks that involve reaching, manipulation, and/or pick-and-place tasks. This paper reviews approaches utilizing visual servoing and deep reinforcement learning (RL) to improve control of robots with a comparison of the two approaches and suggests a hybrid method for optimal control performance.
Problem Statement and Significance
Robotic tasks often involve coordination of visual perception with motion of the robot. Classic methods that rely on visual servoing can achieve good accuracy with limited training data, while methods based on reinforcement learning can generalize globally but require a large amount of training data. There is a possible synergy to bring together the best of both worlds—a hybrid approach that avoids the problems of each method and yields good accuracy, robustness, and efficiency.