Say-REAPEx: An LLM-Modulo UAV Online Planning Framework for Search and Rescue

Björn Döschl, Jane Jean Kiam

Keywords: Search and Rescue Robots, AI-Enabled Robotics, Aerial Systems: Applications

TL;DR: Say-REAPEx, in SAR, plans high-level tasks with a set of low-level actions, by discarding irrelevant or lowly feasible actions based on domain-specific knowledge, while leveraging online heuristic search to reduce uncertainty of future rewards.

Abstract:

While unmanned aerial vehicles (UAVs) are
proven beneficial in search and rescue (SAR) missions, the
scalability of their deployment is in practice still challenging as
high-level decision-making capabilities for UAVs still lack, and
the natural human-in-the-loop command and communications
in a SAR mission are rarely tackled. Some promising largelanguage-model- (LLM-)modulo planning frameworks have
been developed for general robotics, combining the strengths
of LLMs given their vast training data, but complementing
them with domain-specific knowledge and reasoning capabilities
for more robust planning. However, adopting the existing
frameworks for online planning in a SAR mission requires
further adaptations to scale for larger problems, while assuring
the real-time planning capability. We introduce Say-REAPEx,
an LLM-modulo online planning framework that discards
irrelevant or lowly feasible actions based on domain-specific
knowledge in order to reduce the size of the planning problem,
while leveraging online heuristic search to reduce uncertainty
of future rewards. Results of validation tests based on realistic
SAR missions show that Say-REAPEx is 70 % more efficient
compared to existing frameworks, while maintaining better and
comparable success rate.

Submission Number: 7