The expensive learning cost has become a serious problem in robotic arm movement planning using reinforcement learning method. A significant amount of training is required before convergence, reducing the effectiveness of its application. To address the problems above, we propose a guided attenuation reward shaping (GAR) method that reshapes the reward of the robotic arm movement planning environment. It gradually attenuates from a slightly dense distance reward to a sparse binary reward to speed up the Actor-Critic based reinforcement learning algorithm’s convergence speed, allowing the robot arm to learn the optimal path to reach the target swiftly. We verify the method in the built robotic arm operating environment and conduct comparative experiments. The result shows that our proposed GAR method significantly increases the convergence speed and obtains the shortest path to the target.
Faster Robotic Arm Movement Planning via Guided Attenuation Reward Shaping
Lect. Notes Electrical Eng.
International Conference on Guidance, Navigation and Control ; 2022 ; Harbin, China August 05, 2022 - August 07, 2022
2023-01-31
9 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
ITERATIVE REWARD LEARNING FOR ROBOTIC EXPLORATION
TIBKAT | 2020
|Iterative Reward Learning for Robotic Exploration
AIAA | 2020
|British Library Conference Proceedings | 2006
|BASE | 2020
|