Frame-Level Control for Dual-Altitude UAV Collaborative Navigation: A Self-Evolving Agentic Approach

Dual-Altitude UAV collaborative navigation assigns two UAVs at distinct altitudes to work together and visually search outdoor regions for a specific target: a high-altitude UAV provides bird's-eye-view (BEV) observation for overall guidance, while a low-altitude UAV uses observations from a front-mounted camera to navigate around obstacles. Existing methods often deploy fixed models on UAVs. However, once deployed, these models struggle to adapt to different environments or learn from new experiences. In this work, we propose Frame-Level Agentic Navigation Evolution (FLANE), a self-evolving agentic approach that automatically learns from its experiences. To start with, we introduce a training-free navigation pipeline based on visual prompting and dual-view occupancy estimation. On top of this, we introduce a self-evolution paradigm for frame-level UAV navigation control based on self reflection. Specifically, we first establish frame-level motion control APIs that can adjust UAV path and motion styles. Second, for self reflection, since it is impossible to feed all frames into a large language model (LLM), we visualize 1) frame-level measurements as UAV statistics charts and 2) the UAV trajectory as a BEV flight map, and use them instead. Given statistics and flight map visualization, LLMs are then instructed to reflect, summarize, and extract compact behavioral rules by programming with frame-level motion control APIs. On the HaL-13k dataset, FLANE achieves up to a 39.43% success rate, outperforming previous state-of-the-arts by 22.86%.