Optimizing Robotic Arm Learning: Curiosity-Driven Deep Deterministic Policy Gradient

Applied Statistics, University of Toronto, 27 King's College Cir, Toronto, Canada

^* Corresponding author: ljr.liu@mail.utoronto.ca

Abstract

This study explores the application of the Reinforcement Learning (RL) in training robotic arms, particularly using the Deep Deterministic Policy Gradient (DDPG) algorithm enhanced by a curiosity- driven mechanism. Robotic arms have various real-life applications, such as in the surgeries and assistive technologies. However, collecting the large- scale real-world data is costly and impractical, making simulation environments essential for optimization. The DDPG, well-suited for continuous action spaces, was employed to improve the robotic arm’s precision and adaptability. Integrating a curiosity mechanism allowed the system to explore and learn more efficiently, significantly improving the training time and success rate. The results demonstrate a 12% reduction in training time and an 18% increase in the success rate when using curiosity- driven exploration. These findings suggest that the enhanced DDPG algorithm not only accelerates learning but also enables better task execution, offering a promising approach for the real-world robotic applications.