中国科学院自动化研究所机构知识库

Deep reinforcement learning can endow robot agents with the ability to learn manipulation skills with promising performance. However, manipulation skills learning on complicated tasks poses great challenges due to higher exploration of state-action space and uninformative sparse rewards. To address the challenges in manipulation skill learning on multi-step complex tasks and overcome the limitation of current methods, we propose an efficient curriculum learning method called Task Auxiliary and Task Difficulty-HER (TATD-HER), which is endowed with a high-low level curriculum learning structure and combines the explicit and implicit curriculum learning mechanisms together for policy training. The experimental results show that the proposed learning paradigm has the capability of achieving satisfactory performance in manipulation skill learning on multi-step complicated task.