This research by MIT focuses on using Large Language Models (LLMs) for resilience and adaptation in autonomous aerial robots. The method introduced, REAL, uses LLMs to enhance the mission planning and control framework of an autonomous robot. LLMs provide prior knowledge to increase resilience in challenging scenarios, interpret natural-language and log/diagnostic information for mission planning, and adapt control inputs with minimal user-provided knowledge about the robot’s dynamics. The study demonstrates REAL’s effectiveness in a real-world multirotor, reducing position tracking errors and making decisions to avoid potentially dangerous situations.
Publication date: 2 Nov 2023
Project Page: https://arxiv.org/abs/2311.01403v1
Paper: https://arxiv.org/pdf/2311.01403
