中国科学院自动化研究所机构知识库

Deep reinforcement learning (DRL) has achieved remarkable results in many high-dimensional continuous control tasks. However, the RL agent still explores the environment randomly, resulting in low exploration efficiency and learning performance, especially in robotic manipulation tasks with sparse rewards. To address this problem, in this paper, we introduce a simplified Intrinsic Curiosity Module (S-ICM) into the off-policy RL methods to encourage the agent to pursue novel and surprising states for improving the exploration competence. This method can be combined with an arbitrary off-policy RL algorithm. We evaluate our approach on three challenging robotic manipulation tasks provided by OpenAI Gym. In our experiments, we combined our method with Deep Deterministic Policy Gradient (DDPG) with and without Hindsight Experience Replay (HER). The empirical results show that our proposed method significantly outperforms vanilla RL algorithms both in sample-efficiency and learning performance.