8th Annual Conference on Robot LearningAbstract
While unmanned aerial vehicles (UAVs) are proven beneficial in search and rescue (SAR) missions, real-world scalability remains challenging due to limited high-level decision-making capabilities and rarely addressed human-in-the-loop command and communications. Recent LLM-modulo planning frameworks combine the broad knowledge of LLMs with domain-specific reasoning for more robust planning, yet adapting them for online SAR planning requires further changes to scale while preserving real-time performance. We introduce Say-REAPEx, an LLM-modulo online planning framework that discards irrelevant or low-feasibility actions using domain knowledge to shrink the planning space, and leverages online heuristic search to reduce uncertainty in future rewards. In validation tests based on realistic SAR missions, Say-REAPEx is 70% more efficient than existing frameworks while achieving comparable or better success rates.
Björn Döschl
Research Associate
Research associate and PhD candidate at the Institute of Flight Systems, focusing on UAV autonomy, swarm intelligence, and multi-agent systems.
