E²BA: Environment Exploration and Backtracking Agent for Visual Language Object Navigation

摘要

Robot navigation in an unknown environment is a challenging task, due to the lack of spatial awareness and semantic understanding of the environment. Previous works predominantly relied on prior scene knowledge and semantic information, lacking generalization and transferability. This paper proposes an environment exploration and backtracking agent (E2BA) for visual language object navigation, which leverages the rich semantic prior knowledge and commonsense reasoning of large language models (LLMs) to explore the environment and find the object. By fusing LLM scores and spatial geometric costs using particle filters, we select a redefined optimal frontier as sub-goal for environment exploration. To avoid redundant exploration and paths, we design a backtracking discriminator to evaluate the state of the agent and determine the timing of backtracking triggering through a double-level cascade mechanism. Additionally, we design a random instruction fuzzy semantic guessing task to verify the application diversity of this method. Comprehensive experiments on the Habitat-Matterport 3D dataset show that our method achieves a success rate of 0.704, which is higher than the existing baseline method. This study explores the potential application of LLMs in environment exploration without the need for additional training and semantic supplementation.