EV-infotaxis: Extended Vision-language Model-based Infotaxis for Efficient Gas Source Term Estimation and Search

Abstract

As hazardous gas leakage incidents increase, rapid and accurate source term estimation (STE) has become essential. Infotaxis is an widely-used information-theoretic search method based on Bayesian estimation to infer the information of a source utilizing a gas dispersion model and a sensor model. However, when the gas dispersion model deviates from the actual dispersion distribution, its estimation and search performance degrade. To address this limitation, this study proposes extended vision-language model-based Infotaxis (EV-Infotaxis), which integrates an extended vision-language model (EVLM) into the search process. The proposed method employs Grounding DINO for object detection incorporating environmental context and utilizes CLIP to compute semantic similarity between images and texts entered by a user, improving the efficiency of VLM-based search. Additionally, the gas sensor information is effectively fused with the image information from the EVLM, enabling more efficient exploration and exploitation. Simulation results verify the effectiveness of the proposed framework, confirming that EV-Infotaxis achieves superior search efficiency compared with Infotaxis. © ICROS 2025.