跳转至

文章背景与核心概要

无人机视觉语言导航（UAV-VLN）要求机器在复杂的3D环境中实现快速且具反应性的控制。传统的端到端方案虽然展现出巨大潜力，但往往依赖拥有数十亿参数的大型语言模型，导致极高的延迟，难以在边缘设备上实现实时部署。为此，本文作者对这种“重参数”范式提出了挑战，并通过跨尺度评估揭示了一个关键洞察：感知质量在根本上优于语言推理能力。

为了在不依赖不可扩展的人工标注的前提下解决轻量级策略在纯行为克隆（BC）下缺乏显式负向反馈、易在分布外（OOD）场景中碰撞的痛点，作者引入了 AeroDPO——一个由确定性物理仿真状态回滚驱动的零成本自动化直接偏好优化流水线。该方法赋能一个轻量级的 2B 参数模型实现了 SOTA 级别的自主飞行性能，彻底打破了边缘部署的延迟瓶颈。

AeroDPO: Unleashing Lightweight UAV Navigation with High-Fidelity Perception and Automated Preference Optimization

arXiv: 2608.07557 [cs.RO]
Authors: Peng Xu, Chengcheng Wang, Shaohua Wan
Submission Date: August 2, 2026 (Last revised August 25, 2026)
Code Repository: GitHub - XuPeng23/AeroDPO

arXiv: 2608.07557 [cs.RO]
Authors: Peng Xu, Chengcheng Wang, Shaohua Wan
Submission Date: August 2, 2026 (Last revised August 25, 2026)
Code Repository: GitHub - XuPeng23/AeroDPO

Executive Summary

Vision-Language Navigation for Unmanned Aerial Vehicles (UAV-VLN) demands rapid, reactive control in complex 3D environments. While minimalist end-to-end paradigms show strong potential, they traditionally rely on massive language models with billions of parameters, resulting in prohibitive latency for real-time edge deployment.

Vision-Language Navigation for Unmanned Aerial Vehicles (UAV-VLN) demands rapid, reactive control in complex 3D environments. While minimalist end-to-end paradigms show strong potential, they traditionally rely on massive language models with billions of parameters, resulting in prohibitive latency for real-time edge deployment.

In this work, the authors challenge this parameter-heavy paradigm. Comprehensive cross-scale evaluations reveal a critical insight: perception quality fundamentally outweighs language reasoning capacity. A lightweight 2B model equipped with high-fidelity visual inputs can fully match the success rates of massive 7B baselines. However, this minimalist policy exposes a core vulnerability inherent to pure Behavior Cloning (BC)—lacking explicit negative feedback, the agent fails to internalize spatial constraints and suffers from high collision rates in out-of-distribution (OOD) scenarios.

In this work, the authors challenge this parameter-heavy paradigm. Comprehensive cross-scale evaluations reveal a critical insight: perception quality fundamentally outweighs language reasoning capacity. A lightweight 2B model equipped with high-fidelity visual inputs can fully match the success rates of massive 7B baselines. However, this minimalist policy exposes a core vulnerability inherent to pure Behavior Cloning (BC)—lacking explicit negative feedback, the agent fails to internalize spatial constraints and suffers from high collision rates in out-of-distribution (OOD) scenarios.

To resolve this without relying on unscalable human annotations, the authors introduce AeroDPO, a zero-cost automated Direct Preference Optimization pipeline powered by deterministic physical simulation state rollback.

To resolve this without relying on unscalable human annotations, the authors introduce AeroDPO, a zero-cost automated Direct Preference Optimization pipeline powered by deterministic physical simulation state rollback.

Key Methodology: The AeroDPO Pipeline

AeroDPO addresses the robustness flaws of pure Behavior Cloning through an automated data flywheel:

Key Methodology: The AeroDPO Pipeline

AeroDPO addresses the robustness flaws of pure Behavior Cloning through an automated data flywheel:

  1. State Rollback & Causal Error Extraction: Upon detecting a collision, the system autonomously rewinds the simulation environment to isolate and extract causal reasoning errors as rejected actions.
  2. Privileged Interventions: It applies decoupled privileged interventions to synthesize collision-avoidance preferred maneuvers.
  3. Visual Filtering: An offline vision-language inspector is leveraged to filter out visual ambiguities.
  1. State Rollback & Causal Error Extraction: Upon detecting a collision, the system autonomously rewinds the simulation environment to isolate and extract causal reasoning errors as rejected actions.
  2. Privileged Interventions: It applies decoupled privileged interventions to synthesize collision-avoidance preferred maneuvers.
  3. Visual Filtering: An offline vision-language inspector is leveraged to filter out visual ambiguities.

Performance Highlights

  • Success Rate: Boosts performance to 49.16% on unmapped scenarios.
  • Safety: Drastically suppresses collision rates compared to baseline models.
  • Efficiency: Empowers a lightweight 2B parameter model to achieve SOTA performance for autonomous aerial agents without edge-deployment latency bottlenecks.

Performance Highlights

  • Success Rate: Boosts performance to 49.16% on unmapped scenarios.
  • Safety: Drastically suppresses collision rates compared to baseline models.
  • Efficiency: Empowers a lightweight 2B parameter model to achieve SOTA performance for autonomous aerial agents without edge-deployment latency bottlenecks.

Metadata & Reference Information

Metadata & Reference Information