Behavior-triggered self-reflection for zero-shot open-domain UAV target search

Abstract

Open-domain Object Goal Navigation requires a UAV agent to locate arbitrary target objects in previously unseen environments without task-specific training. Existing LLM-based navigation methods often operate in a reactive manner and lack explicit mechanisms for long-term exploration awareness and online self-correction, leading to inefficient behaviors such as revisiting explored regions and getting trapped in local exploration loops. To address these limitations, we propose a closed-loop zero-shot navigation framework that integrates episodic spatial memory and behavior-triggered reflection into the LLM decision process. The memory module summarizes exploration history as a structured language context to improve spatial awareness, while the reflection module detects common navigation failures and provides corrective guidance without additional training or reward signals. Experiments on the UAV-ON benchmark demonstrate that the proposed method consistently improves navigation success rate and path efficiency over existing zero-shot baselines. Ablation studies further validate the effectiveness and complementary roles of episodic memory and behavior-triggered reflection, highlighting the benefit of integrating exploration-aware memory with online self-correction for long-horizon UAV target search.