Bayesian Active Learning for Intent Disambiguation in Interactive Robot Planning
This study addresses the problem of intent inference failures in interactive robot planning caused by ambiguous and incomplete natural language instructions. We propose an active clarification framework that integrates large language models (LLMs) with Bayesian optimization. Specifically, multi-turn dialogue disambiguation is formulated as an active learning process grounded in Signal Temporal Logic (STL). The LLM generates formal specifications and natural language queries, while Bayesian optimization estimates uncertainty and selects questioning strategies that maximize information gain; a formal planner then synthesizes verifiable trajectories. Evaluations across simulated and real-world task domains demonstrate that our approach significantly improves task satisfaction and reduces clarification turns compared to LLM baselines, effectively narrowing the performance gap between smaller and larger language models.