Indoor Semantic Map Prediction for Object Search Based on Relationship Reasoning and Active Training

An accurate and complete semantic map is crucial for effective object search. However, in real indoor environments, factors such as object occlusion, the robot's limited field of view, and the highly diverse spatial structures and semantic layouts can significantly degrade the performance of semantic mapping and object search. To address these challenges, we design a semantic map prediction method enhanced with semantic-spatial relationship reasoning, which models the semantic and spatial correlations between semantic entities. To further improve the performance in the new environments, we introduce an informationĝ€'gainĝ€'based active training strategy that guides the robot to actively collect informative samples and continuously refine the semantic map prediction model. Finally, we propose an uncertaintyĝ€'driven target object search method. Our approach is evaluated in simulated environments based on the Matterport3D dataset and demonstrates significant improvements in both semantic mapping and object search.