AeroDPO: 2B UAV model matches 7B baselines, slashes collisions

AeroDPO, from researchers Peng Xu, Chengcheng Wang, and Shaohua Wan, challenges the assumption that UAV navigation requires massive language models. The team's cross-scale evaluations show perception quality matters more than reasoning capacity. A lightweight 2B model with high-fidelity visual inputs completely matches the overall success rates of 7B baselines, enabling real-time edge deployment where billions of parameters would cause prohibitive latency.

However, pure behavior cloning exposed a critical flaw: lack of negative feedback leads to high collision rates in out-of-distribution scenarios. AeroDPO fixes this with a zero-cost automated Direct Preference Optimization pipeline. When the simulated drone collides, the system rewinds the physical state, extracts the causal error, and synthesizes collision-avoidance maneuvers via privileged interventions. An offline vision-language inspector filters visual ambiguities. This automated data flywheel pushes success to 49.16% on unmapped scenes while suppressing collisions, establishing a new state-of-the-art.