Robust Room Contour Extraction From 2-D Home Metric Maps Under Cluttered Layouts and Map Noise for Domestic Robots

Room area extraction from occupancy grid maps of home environments is critical for enabling robotic room-level spatial awareness of environmental structures, serving as a foundation for semantic-aware navigation and decision-making. However, most existing methods rely on geometric heuristics derived from human abstraction, which tend to exhibit unstable performance across maps with heterogeneous home layouts and significant noise. Learning-based approaches offer the potential to capture generalizable room features beyond the reach of human abstraction, but their application on grid maps is severely limited by the scarcity of real-world map data. To address this, we propose a simulation-based learning framework for the accurate and robust extraction of room contours from real-world home grid maps in the presence of cluttered home layouts and map noise. Specifically, we propose a realistic simulation method to generate a large-scale synthetic grid map dataset by simulating diverse household obstacle layouts, introducing multi-source mapping uncertainties, and replicating mapping traces using a reinforcement learning-based exploration controller. In addition, we propose an efficient semi-automatic annotation method to produce high-precision room contours for the extensive simulated maps. A deep learning network is employed to learn room features and predict room contours. Experimental results on real-world home maps demonstrate that our method extracts more accurate and complete room areas than existing state-of-the-art methods and facilitates room-related robotic applications. The source code is available for download at https://github.com/sxy370921/RRCE 2D .