CRAR: Coherent Risk-Aware Regulation for DRL-Based Mapless Robot Navigation

Navigating without prior maps poses significant challenges for robots, especially in complex environments populated with pedestrians and obstacles. In such settings, achieving a balance between safety, socially compliant behavior, and path efficiency is critical. To address this, we propose a novel deep reinforcement learning-based approach, termed Coherent Risk-Aware Regulation (CRAR), tailored for Unmanned Ground Vehicles (UGVs). Our method incorporates a risk prediction network that enables the robot to anticipate potential hazards arising from pedestrians or obstacles. Furthermore, by integrating risk assessment and correction mechanisms, CRAR facilitates rapid learning of avoidance strategies, thereby reducing training duration. Consequently, the robot can identify navigation paths that are both safer and more efficient. Simulation results demonstrate that, in open environments, CRAR achieves an average success rate improvement of 25.33% and reduces navigation time by 10.85% compared to baseline methods. In dense environments, these benefits are amplified, with a success rate increase of 30.46% and a navigation time reduction of 21.75%. Additionally, CRAR ensures greater safety margins from obstacles and smoother trajectories, underscoring its efficacy in navigating dense and dynamic settings. Finally, we validate the performance and practicality of the proposed method through real-world experiments. A video demo is available athttps://youtu.be/j0-pEd7_gr8.