Autonomous Navigation Framework for Mobile Robots in Highly Dynamic and Complex Indoor Environments

Abstract

Autonomous navigation in complex, dynamic indoor environments is one of the fundamental challenges in intelligent automation and robotics engineering. Conventional Simultaneous Localization and Mapping (SLAM) approaches heavily rely on geometric arrangements, making them vulnerable to sensor performance degradation, localization drift and navigational crash in busy environments populated by dynamic objects or humans. The main contribution of this work is the introduction of a novel, robust hybrid semantic-geometric navigation framework specifically envisioned for mobile service robots to overcome these fundamental limitations. This work proposes an architecture that fuses real time Semantic Segmentation networks and Deep Reinforcement Learning (DRL)-powered local planners into a system able to parse spatial geometry together with semantic contextual environments. The navigation stack classifies static structures, predictable dynamic agents and erratic obstacles in indoor features into three isolated subspaces that optimize trajectory generation while retaining the localization. Experimental validation in simulation environments with dense layout of obstacles and real-world indoor environments show that the proposed framework achieves 14.5% improvement in path efficiency, with 18.2% less computational time for trajectory compared to baseline ROS NavFn and dynamic window approaches (DWA). These results suggest that this combination of high-level semantic awareness with low-level geometric tracking is a robust and scalable framework for future indoor autonomous systems.