SLAM and deep reinforcement learning-based autonomous navigation delivery robot
Yongdong Gan, Zhen Xie · 2025
To address the challenges of path planning and obstacle avoidance for autonomous navigation delivery robots in complex dynamic environments, a Simultaneous Localization and Mapping (SLAM) with deep reinforcement learning collaborative navigation framework SLAM-DEDQN is proposed in this paper. Firstly, an innovative multi-resolution grid map construction method is designed in the SLAM module, which establishes a region feature-based adaptive resolution adjustment mechanism. In addition, a Dynamic ε-Greedy Policy Optimized DQN algorithm (DE-DQN) is proposed in the deep reinforcement learning module. DE-DQN dynamically adjusts the exploration factors with an exponential decay function based on learning progress. Finally, a positional information-based Q-value initialization strategy is proposed, which utilizes the reciprocal of the Euclidean distance between the robot's current position and the target point as initial Q-values. Experimental results in simulation environments containing dynamic obstacles demonstrate that, the proposed SLAM-DEDQN achieves a path planning success rate of 96.7%, with average navigation time reduced by 28.4% compared to traditional methods.