LMPath: Language-Mediated UAV Exploration for Object Search
Researchers have developed LMPath, a pipeline that uses generative language models and vision models to create semantic exploration priors for UAV search missions. Traditional autonomous UAV search relies on geometric coverage patterns that ignore semantic context, wasting time in large-scale environments. LMPath takes a basic geofence and an object of interest prompt, then uses language models to predict likely regions for the object and a foundation vision model to segment sub-regions from satellite imagery. This prior enables UAV path generation with objectives like minimizing expected search time, maximizing discovery probability within a distance limit, or narrowing search to high-likelihood sub-regions. The approach leverages semantics to improve efficiency over blind coverage patterns.
Key facts
- LMPath uses generative language models to determine regions likely to contain an object of interest.
- A foundation vision model segments satellite imagery into sub-regions for exploration priors.
- Traditional UAV search uses geometric coverage patterns without semantic context.
- LMPath can generate paths to minimize expected time to locate the object.
- It can also maximize probability of finding the object given a limited travel distance.
- The pipeline narrows search space to sub-regions most likely to contain the target.
- The approach is designed for large-scale environments to reduce time waste.
- LMPath is presented in arXiv paper 2605.13782.
Entities
Institutions
- arXiv