An Experience Owner-Aided Training method for compound actions validation in an autonomous precision assembly task

Deep reinforcement learning's efficacy in tasks involving compound actions is limited by its reliance on detailed reward systems, particularly in uncertain environments with sparse rewards. Thus, a hybrid training method called “Experience Owner-Aided Training” (EOAT) is proposed which enables actor-critic-based reinforcement learning models to efficiently learn under sparse rewards, avoiding complex reward design. The method combines two structurally identical reinforcement learning algorithms, each specializing in understanding local and global compound actions within complex tasks, respectively, and promoting their mutual fusion by a dynamic Q function. The method was experimentally validated in a robotic autonomous precision assembly task, showcasing exceptional learning speed and robustness, surpassing conventional methods. This achievement offers a novel approach to efficient learning in uncertain environments with sparse rewards and injects new vitality into the development of robotic reinforcement learning.