MapBot: A Multi-Modal Agent for Geospatial Analysis

MapBot is an interactive system to manipulate, analyze, and visualize geospatial data. It combines frontier computer vision models with a large language model running in a Read–Eval–Print Loop (REPL). Users can upload or select aerial or satellite imagery, annotate objects, and query the data using natural language and a point-and-click interface. The LLM agent loop enables the orchestration of Segment Anything and DinoV2, Python code generation and execution, and the display of results in a web interface. This approach lowers the barrier to geospatial analysis for non-experts, enabling rapid annotation and querying of complex data through dialogue that includes map-based interaction.

Authors

• Martin Weiss Mila, Quebec AI Institute & Polytechnique Montreal
• Nasim Rahaman Mila, Quebec AI Institute & Max Planck Institute for Intelligent Systems
• Chris Pal

Keywords

• Geospatial Analysis
• Multi-Modal Models
• Agentic Systems