Autonomous mobile robots are essential for applications like search and rescue, but navigating unfamiliar environments requires sophisticated decision-making based on semantics which are typically obtained from human input. A key limitation in traditional navigation approaches includes a lack of semantic understanding core to human reasoning. This work aims to bridge this gap by leveraging the power of large language models (LLMs) and traditional navigation techniques to enhance robotic situational awareness.

While LLMs excel at language tasks, integrating them into robotic systems presents challenges. The "world knowledge" within LLMs must be grounded in the robot's physical sensor data for meaningful use. Additionally, safety in robotic systems demands explainable outputs. This dissertation presents two frameworks that address these limitations: code generation from natural language and a contextual reasoning framework that encodes environmental states as natural language. These frameworks build upon a state-of-the-art exploration system developed for the DARPA Subterranean Challenge, which enabled successful autonomous navigation in GPS-denied environments.

This work's contributions include a flexible exploration system for mobile robots that emphasizes autonomy with high-level human input, along with innovations in subterranean communications. Most importantly, this research introduces two LLM-based frameworks – code generation and contextual reasoning – to improve robots' understanding and navigation abilities within their environments.