Navigation in Underground Parking Lot by Semantic Occupancy Grid Map Prediction

Abstract

Autonomous navigation in complex environments, such as underground parking lots, poses significant challenges due to the absence of prior maps and reliance on real-time perception. This paper proposes a comprehensive framework for mapless exploration and navigation, integrating a Semantic Occupancy Grid Network (SOCCNet) and a Navigator module. SOCCNet predicts local semantic occupancy grids from sensor data, achieving an average pixel accuracy of 0.8234 on test maps. The Navigator constructs and updates a global semantic occupancy grid using a Bayesian approach, incorporating distancebased weighting to mitigate uncertainties in distant predictions. Exploration targets, termed Topology Nodes, are sampled and scored based on proximity and semantic likelihood, guiding the robot via an A* planner. Evaluated in the Isaac Sim environment across multiple trials, the framework successfully explored all spaces in 11 out of 12 trials (91.7% success rate), despite occasional revisits to known areas. Our results demonstrate robust adaptability and efficiency, offering a practical solution for autonomous navigation in unmapped, dynamic settings.