Robotic Hand-Eye Coordination Fusion

Abstract

Hand-eye coordination is crucial for performing tasks like reaching for objects in both humans and robots. This paper investigates two methods—visual servoing and deep reinforcement learning (RL)—to achieve hand-eye coordination in robotic systems. Visual servoing typically leverages video tracking and online Jacobian learning to control the robot based on camera-robot geometry, while RL uses neural networks to learn a global visuomotor policy. I conducted experiments using the WAMVisualReach environment to compare these methods in terms of sample complexity and task performance. Despite the reliance on simulation, my findings suggest promising avenues for real-world applications, particularly when combining both methods to enhance sample efficiency and robustness. Future work will focus on validating these methods on physical robots and exploring their performance in more complex tasks like pick-and-place.