SE-DQN: local path planning method of mobile robots based on security exploration

As a classical deep reinforcement learning method, Deep Q-Network (DQN) is widely used in path planning of mobile robots. However, in the complex unknown environments, due to lacking the guidance of prior knowledge and seeking the optimal path at the cost of frequent trial and error, DQN-based path planning methods have the problems such as difficult model convergence, low exploration efficiency and poor real-time obstacle avoidance ability. To address these issues, this paper proposes a new path planning method named security exploration DQN (SE-DQN). In SE-DQN, the developed “expert experience” module combined with security exploration strategy guides the action selection to achieve collision-free safe exploration and improve exploration efficiency. In addition, an intensive reward function is meticulously designed so that the agent get timely feedback from the environment after performing each action. To verify SE-DQN’s performance, the convergence, generalization, and obstacle avoidance tests are designed and conducted. The experimental results demonstrate that the proposed SE-DQN algorithm exhibits excellent convergence performance, environmental adaptability, and obstacle avoidance capabilities.