End-To-End Mapless Navigation in Environments with Mixed Obstacles

Achieving mapless autonomous navigation in unknown environments is a key challenge for mobile robots, particularly in the presence of dynamic obstacles and complex environmental conditions. This paper proposes an end-to-end deep reinforcement learning approach based on Proximal Policy Optimization (PPO) to enable robots to perform path planning and obstacle avoidance without prior map information. The proposed policy network takes raw 2D LiDAR scan data and the relative position of the target as inputs and directly outputs continuous linear and angular velocity commands, eliminating the need for map construction or manual feature extraction. To effectively guide policy learning, a composite reward function is designed to balance target approaching, orientation alignment, motion smoothness, and obstacle safety, allowing the agent to achieve an effective trade-off between exploration and exploitation. Extensive training and testing conducted on the Gazebo simulation platform show that the trained policy achieves a task success rate exceeding 95% in complex unknown environments with both static and dynamic obstacles, while demonstrating good convergence, training stability, and crossenvironment generalization. This method provides a practical and efficient solution for mobile robots to navigate autonomously in real-world unknown environments and lays a foundation for future research on multi-goal navigation and dynamic environments. The source code of our method is publicly available at https://github.com/wyh010731/amr_rl_ws.git.