SwarmSAT Agent: A Tool-Grounded Autonomous Multi-Agent LLM Framework for Rapid Satellite Constellation Retasking of Pop-Up Events

Adib Bazgir, Yuwen Zhang

Keywords: Multi-agent systems, Satellite constellation tasking, Large language models (LLMs), Auction-based coordination, Pop-up event response

Abstract: Coordinating a satellite constellation for time-critical pop-up events (e.g., wildfires) requires fast, resource-aware tasking under partial visibility, limited onboard energy, and operator-imposed rules. We present SwarmSAT, a physically grounded multi-agent constellation tasking framework that treats pop-up response as a lightweight sequential handoff auction. A Cluster Lead agent detects a pop-up, identifies eligible satellites via orbital visibility windows derived from real TLE propagation, and orchestrates turn-based bid submission from satellite agents. Each satellite computes a composite bid that trades off predicted peak imaging quality against current battery state, enabling resource-aware assignment and explicit redundancy suppression. We evaluate SwarmSAT in a discrete-time orbital simulation with battery recovery/consumption dynamics, routine background targets, and injected pop-up disasters, and compare against a greedy baseline scheduler. In a representative Australia bushfire scenario, SwarmSAT eliminates redundant pop-up imaging (0\% redundant vs.\ 98.6\% for greedy) and reduces time-to-first observation from 720\,s to 60\,s. Across four geographically distinct pop-up scenarios, SwarmSAT improves average pop-up utility by \(\sim\)36\% while reducing time-to-first observation by \(\sim\)99\%, and it maintains near-zero redundancy by selecting a single winner to execute the task. Beyond performance, SwarmSAT enables controllable operations where it satisfies 4/4 natural-language operator constraints (battery floor, quality floor, satellite exclusion, and no-go zone avoidance), whereas the greedy baseline violates 3/4. We further report a multi-pop-up stress test illustrating degradation under simultaneous disasters, motivating coordination mechanisms for scalable real-world event response.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 52