Lightweight 2B Model Matches 7B Baselines in UAV Navigation

A new paper on arXiv (2608.07557) challenges the reliance on massive language models for Vision-Language Navigation in Unmanned Aerial Vehicles (UAV-VLN). The authors demonstrate that a lightweight 2B parameter model, when equipped with high-fidelity visual inputs, achieves success rates comparable to 7B parameter baselines. This finding suggests that perception quality is more critical than language reasoning capacity for UAV navigation tasks. However, the minimalist policy reveals a robustness flaw inherent to pure Behavior Cloning (BC): without explicit negative feedback, the agent fails to internalize spatial constraints and exhibits high collision rates in out-of-distribution scenarios. The paper proposes a solution called AeroDPO, which likely involves automated preference optimization to address this issue. The research underscores the importance of high-fidelity perception and the need for robust training methods in edge deployment of UAV navigation systems.