<![CDATA[The increasing demand for contactless urban logistics has driven the integration of autonomous delivery robots into real world operations. This study investigates the application of Deep Reinforcement Learning (DRL) to enhance robot navigation in complex urban environments, focusing on three advanced models: MODSRL, SOAR RL, and NavDP. MODSRL employs a multi objective framework to balance safety, efficiency, and success rate. SOAR RL is designed to handle high obstacle densities using anticipatory decision making. NavDP addresses the sim to real gap through domain adaptation and few shot learning. The models were trained and evaluated in simulation environments (CARLA, nuScenes, Argoverse) and validated using real world deployment data. Evaluation metrics included success rate, collision frequency, and energy efficiency. MODSRL achieved a 91.3% success rate with only 4.2% collision, outperforming baseline methods. SOAR RL showed robust performance in obstacle rich scenarios but highlighted a safety efficiency trade off. NavDP improved real world success rates from 50% to 80% with minimal adaptation data, demonstrating the feasibility of sim to real transfer. The results confirm the effectiveness of DRL in advancing autonomous delivery navigation. Integrating domain generalization, hybrid learning, and real time adaptation strategies will be essential to support large scale urban deployment. Future research should prioritize explainability, continual learning, and user centric navigation policies.]]>
Copyrights © 2025
