Map-based and Map-less Mobile Navigation in Crowded Dynamic Environments

This thesis presents a comparative analysis of map-based and map-less navigation strategies for mobile robots operating in congested dynamic environments. The map-based framework is constructed on the ROS2 NAV2 stack, utilizing Simultaneous Localization and Mapping (SLAM) methodologies and global path planning. Conversely, the map-less system utilizes a model-free deep reinforcement learning approach, specifically the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, for reactive navigation based on real-time sensor data. Both methodologies are executed on a simulated differential-drive platform called "Robot Ben" within a Gazebo environment and assessed under uniform conditions across various navigation scenarios. The trajectory optimality, navigation success rate, computational efficiency (CPU, RAM, GPU utilization), and power consumption were analysed. Experimental results from more than 4,000 trials indicate that although the TD3-based method exhibits reduced CPU utilization, it is surpassed by the map-based NAV2 configuration regarding path quality, robustness, and energy efficiency. The results highlight the practical difficulties of implementing DRL-based navigation in dynamic real-world environments and propose that hybrid architectures, which integrate SLAM with learned local policies, could provide a feasible solution.