Aero-VPA: Few-Step Flow-Matching Control over Non-Reconstructive JEPA Latents for Agile Flight on Edge Hardware

📄 arXiv: 2611.04392v1 📥 PDF

作者: Yeonseok Lee

分类: cs.LO, cs.LG, cs.RO

发布日期: 2026-11-03

备注: 38 pages, 3 figures, 22 tables. Simulation study (PX4 SITL + Gazebo Harmonic). Version 2 of a preprint first posted on Zenodo (doi:10.5281/zenodo.23633375). Code: https://github.com/ylee-sling/Aero-VPA ; trained policy: https://huggingface.co/ylee-sling/aero-vpa ; benchmark dataset: https://huggingface.co/datasets/ylee-sling/aero-vpa-benchmark

🔗 代码/项目: GITHUB | HUGGINGFACE

💡 一句话要点

提出Aero-VPA以解决边缘硬件上的快速飞行控制问题

🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture) 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 边缘计算 无人机导航 视觉-语言-动作 实时决策 深度学习

📋 核心要点

  1. 现有的VLA模型在执行任务时需要大量的顺序推理，导致在边缘硬件上性能受限。
  2. Aero-VPA通过设计轻量级的视觉-预测-动作架构，减少了对内存的需求，并提高了决策效率。
  3. 实验结果表明，改进后的策略在180个任务中成功率显著提升，且减少了碰撞率，验证了方法的有效性。

📝 摘要（中文）

视觉-语言-动作（VLA）模型在任务间的泛化能力较强，但其自回归解码器需要根据动作时间范围进行顺序前向推理，这在嵌入式空中平台上受限于从内存流式传输数十亿参数。本文提出了Aero-VPA，一种轻量级的视觉-预测-动作架构，旨在实现边缘硬件上的3D导航。该架构通过紧凑的RGB-D视觉变换器和五头集成预测器，预测下一个潜在状态，并利用条件流匹配求解器生成速度和偏航率。本文首次报告了闭环评估结果，显示在程序生成的森林中进行的四旋翼任务中，经过改进的DAgger策略成功率从21.1%提升至52.2%。

🔬 方法详解

问题定义：本文旨在解决在边缘硬件上进行快速飞行控制时，现有VLA模型因需要大量顺序推理而导致的性能瓶颈。现有方法在嵌入式平台上难以处理数十亿参数的流式传输，影响了实时决策能力。

核心思路：Aero-VPA的核心思路是通过构建一个轻量级的视觉-预测-动作架构，利用RGB-D视觉变换器和集成预测器来高效预测潜在状态，从而减少计算负担并提高决策速度。

技术框架：该架构包括一个紧凑的RGB-D视觉变换器和一个五头集成预测器，后者在没有像素重建的情况下进行训练，预测下一个潜在状态。条件流匹配求解器则在两个欧拉步骤中生成速度和偏航率。

关键创新：最重要的创新在于采用了条件流匹配求解器，使得每个决策仅需五次网络评估，从而显著提高了决策效率，解决了传统方法的延迟问题。

关键设计：在设计中，采用了深度记忆、障碍约束和视野限制等关键参数设置，确保了在复杂环境中的安全性和有效性。

📊 实验亮点

实验结果显示，经过改进的DAgger策略在180个任务中的成功率从21.1%提升至52.2%，同时碰撞率从72.2%降低至22.2%。这些结果表明，Aero-VPA在复杂环境中的表现显著优于传统方法，验证了其有效性。

🎯 应用场景

Aero-VPA的研究成果在无人机导航、自动驾驶和机器人控制等领域具有广泛的应用潜力。其轻量级架构使得在资源受限的边缘设备上实现高效的实时决策成为可能，未来可用于各种复杂环境下的自主飞行任务。

📄 摘要（原文）

Vision-Language-Action (VLA) models generalize across tasks, but their autoregressive decoders need sequential forward passes in proportion to the action horizon, each bounded below on embedded aerial platforms by streaming billions of parameters from memory. We present Aero-VPA, a lightweight Vision-Predictor-Action architecture for 3D navigation on edge hardware. A compact RGB-D Vision Transformer and a five-head ensemble predictor, trained without pixel reconstruction, forecast the next latent state; a Conditional Flow Matching solver generates a velocity and yaw-rate chunk in two Euler steps, so each decision takes five sequential network evaluations; and the ensemble dispersion drives an uncertainty fallback. Version 1 specified the architecture, deployment and protocol without results. This version reports the first closed-loop evaluation: PX4 software-in-the-loop simulation of quadrotor missions in procedurally generated forests. Behaviour cloning failed on all 60 missions despite open-loop errors of 0.02-0.04 m/s: the policy copies its last command, a copycat form of causal confusion. DAgger with a privileged, safety-gated supervisor gave the first successes but plateaued at 11.1-21.1% across six checkpoints. A map-free runtime shield with a 15 s depth memory, barrier constraints, a cap on motion outside the 73-degree field of view and backoff along the flown path raised the first DAgger policy's success on 180 paired missions from 21.1% to 52.2% (95% CI [45.0, 59.4]; exact McNemar p = 1.2 x 10^-10) and cut its crashes from 72.2% to 22.2%, while timeouts rose from 2.8% to 23.3%. All six checkpoints gained 30.0-34.4 percentage points, at mean speeds of 1.16-1.36 m/s. The code (https://github.com/ylee-sling/Aero-VPA), the trained policy (72.6 MB, https://huggingface.co/ylee-sling/aero-vpa) and the full benchmark dataset are public.