GMM-searcher: efficient object search in large-scale scenes using large language models

Autonomous object search in large-scale, dynamic environments presents significant challenges for robots. This paper introduces GMM-Searcher, a framework designed to efficiently guide robots in object search tasks while adapting to complex and changing environments. By leveraging the reasoning capabilities of large language models (LLMs), GMM-Searcher directs the robot toward areas where the object is most likely to be found. For large-scale environments, an adaptive-resolution topological graph (ARTG) is combined with Gaussian Mixture Models (GMM) to optimize memory usage while preserving high environmental fidelity. GMM also stores search experiences to enhance performance in repeated tasks and adapt to environmental changes, informing the LLM's decision-making process for more efficient subsequent searches. Real-world experiments show that GMM-Searcher significantly enhances search efficiency and equips robots with lifelong learning capabilities and adaptability.