L3MVN: Leveraging Large Language Models for Visual Target Navigation

TL;DR

This paper addresses visual target navigation in unseen indoor environments by injecting common-sense priors through large language models. It introduces L3MVN, a modular framework that builds semantic and frontier maps and uses LLMs in zero-shot and feed-forward modes to identify informative frontiers, followed by a local policy guided by the Fast Marching Method. The approach achieves superior performance on Gibson and HM3D compared with map-based baselines, with ablations confirming the value of LLM guidance and cost-utility exploration; it also demonstrates feasibility in real-world robot experiments. The work highlights the potential of leveraging pretrained language models to reduce training costs and improve generalization in embodied navigation tasks, pointing to future work on tighter robot-LLM interaction and sim-to-real transfer.

Abstract

Visual target navigation in unknown environments is a crucial problem in robotics. Despite extensive investigation of classical and learning-based approaches in the past, robots lack common-sense knowledge about household objects and layouts. Prior state-of-the-art approaches to this task rely on learning the priors during the training and typically require significant expensive resources and time for learning. To address this, we propose a new framework for visual target navigation that leverages Large Language Models (LLM) to impart common sense for object searching. Specifically, we introduce two paradigms: (i) zero-shot and (ii) feed-forward approaches that use language to find the relevant frontier from the semantic map as a long-term goal and explore the environment efficiently. Our analysis demonstrates the notable zero-shot generalization and transfer capabilities from the use of language. Experiments on Gibson and Habitat-Matterport 3D (HM3D) demonstrate that the proposed framework significantly outperforms existing map-based methods in terms of success rate and generalization. Ablation analysis also indicates that the common-sense knowledge from the language model leads to more efficient semantic exploration. Finally, we provide a real robot experiment to verify the applicability of our framework in real-world scenarios. The supplementary video and code can be accessed via the following link: https://sites.google.com/view/l3mvn.

Figures (6)

• Figure 1: Visual target navigation example. The robot explores the environment and uses language models to find more relevant frontier (shown in green, with the highest score) based on the observation and the target.
• Figure 2: The architecture of the target navigation framework. The framework takes RGB-D images as input to generate a semantic map and frontiers, and selects a long-term goal based on the maps and object category using the inference of the language model. Once the long-term goal is reached, a local policy guides the final action for the robot.
• Figure 3: Example of zero-shot approach.
• Figure 4: Example of the fine-tuning-based feed-forward approach.
• Figure 5: The visual target navigation experiment process in the Habitat platform for finding a bed. The gray channel represents the barrier, the blue spot and the circle denote the long-term goal selected by our policy, the red thick line represents the trajectory of the robot, the red thin line denotes the frontiers, and other colors represent the semantic objects.