ETD

• AUV
• AUVs
• behavior
• behaviour
• language
• manager
• managing
• mission
• missions
• model
• models
• small
• tree
• trees

Autonomous Underwater Vehicle (AUV) missions are constrained by limited onboard computation, harsh communication conditions, and strict energy and time budgets, which makes cloud-dependent Large Language Models (LLMs) unsuitable for mission management despite their strong reasoning capabilities. This thesis proposes a mission management approach based on a compact, fine-tuned Small Language Model (SLM) that enables operators to issue natural language commands, which are transformed into sequences of high-level tasks executed through behavior trees interfaced with the vehicle’s guidance system.

The SLM, instantiated as FLAN-T5-base and adapted via Low-Rank Adaptation (LoRA), is prompted with a structured combination of available operations, mission context derived from a dedicated mission memory, and the operator’s request, and is trained to output purely symbolic task tags that map directly to behavior-tree subtrees.

The method incorporates an explicit confidence estimation mechanism, adapted from the BSDetector framework, that combines Observed Consistency and Self-reflection Certainty into a scalar confidence score used to automatically trigger clarification dialogues when the model is uncertain, or to replan after execution failures by exploiting mission logs and failure context. The approach is first validated on a proof-of-concept virtual system inspired by the RAMI Marine Robots competition, where the SLM selects operations such as waypoint acquisition, gate traversal, buoy mapping, and pipeline inspection based on evolving mission knowledge. It is then integrated into the mission manager of a simulated AUV whose control stack is natively based on behavior trees, and evaluated using token- and sequence-level accuracy metrics and confidence statistics on dedicated training and testing datasets. Results show that the SLM-based manager achieves high sequence accuracy, correctly triggers the desired tasks, and assigns higher confidence to correct than to incorrect plans, while offering performance comparable to a rule-based mission manager but with substantially improved accessibility and usability for human operators under realistic resource constraints.