Large language models improve team robot rescue in changing disaster areas

Persistent LLM Orchestration of Heterogeneous Robot Teams via Partial Delegation to Multi-Agent Protocols

Multiagent SystemsRobotics

Summary

Robots working together in disaster zones need to respond to changes like spreading fire and moving people, but current plans often assume missions end or conditions stay the same. The authors found that using large language models (LLMs) to guide robots can struggle when missions don’t end and the environment keeps changing. They improved this by summarizing the robot team’s state, letting some robots work autonomously together, and reminding the system to update tasks when new events happen. Their approach saved more people in a simulated rescue with less computing effort than other methods.

What this means in practice

  • For robotics engineers: Coordinate heterogeneous robot teams in dynamic environments like disaster sites to maintain effective ongoing operations despite changing conditions.
  • For emergency response coordinators: Automate task planning for robot teams to improve rescue rates and situational awareness during large-scale emergencies with moving hazards and evacuees.

Authors

Maria G. Mendoza, Victoria M. Tuck, Zachary Ravichandran, Shankar Sastry, George J. Pappas

Abstract

Natural language specifications let non-expert users task robot teams, and large language models (LLMs) have recently proven effective at grounding such semantic specifications into plans for heterogeneous teams. Existing approaches assume missions that terminate and environments that, while unknown, do not change, but both assumptions fail in many emerging applications. For example, a disaster-response mission does not terminate, and its environment changes continuously as hazards spread and evacuees move. We identify three ways an LLM orchestrator breaks in this setting: its prompt grows without bound, its reasoning degrades as agents and targets accumulate, and it stops actively monitoring after covering the search space once. We build on the successful framework of grounded generative intelligence, and address each issue with a classical idea. A Markovian state summary, rebuilt from the semantic map at every invocation, replaces the growing interaction history. Partial delegation lets the orchestrator assign a subset of robots to an autonomous multi-agent protocol and reserve its reasoning for generating subtasks that require detailed parameters and careful ordering. Lastly, observation staleness, surfaced explicitly in the prompt, together with task-and event-triggered re-prompting, keeps the team vigilant. In a simulated urban search-and-rescue mission with spreading fire and moving evacuees, our framework rescues 98% of civilians versus 89% for the strongest baseline while using six times fewer tokens. Additional ablations isolate the contribution of each mechanism.