FSD-VLN Fast-Slow Dual-System Modeling for Aerial Long-Horizon Vision-Language Navigation

Jul 16, 2026

Upload by Aurelia

Vision-Language Navigation (VLN) supports UAVs by grounding natural-language instructions in real-time visual observations, enabling adaptive autonomy beyond GPS- or pre-programmed approaches. Long-horizon UAV-VLN requires both high-level semantic reasoning and low-latency flight control, yet prior methods face structural inconsistency between global multimodal understanding and temporally coherent action generation. FSD-VLN addresses this via an efficient fast-slow dual-system that decouples semantic priors and low-latency action modeling, improves cross-temporal dependency capture, and uses time-aware adaptive optimization for stable training.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

Abstract

Introduction

Background and problem

Limitations of existing methods

Proposed approach

arXiv :2607 .08359v 1 [ cs .RO] 9 Jul 2026 FSD-VLN: Fast-Slow Dual-System Modeling for Aerial Long-Horizon Vision-Language Navigation Xueke Zhu∗1, Qingyan Meng∗1, Liutao Yu 1 , Wei Zhang 1 , Zhengyu Ma 1 , Huihui Zhou 1 ,2 , Yonghong Tian 1 ,3 ∗ These authors contributed equally to this work. 1 Pengcheng Laboratory 2 Shenzhen Institutes of Advanced Technology 3 Peking University Abstract Vision-Language Navigation (VLN) enables unmanned aerial vehicles (UAVs) to navigate autonomously in unfamiliar environments by grounding natural language instructions in real-time vision observations. Compared to traditional GPS-based or pre-programmed navigation methods, VLN offers more intuitive human-machine interaction and greater adaptability. Effective UAV-VLN systems typically require the integration of high-level semantic reasoning and low-latency flight control. However, existing approaches often suffer from a structural inconsistency between global multimodal understanding and temporally coherent action generation, leading to unstable trajectories and significant decision latency during long-horizon navigation. To address this challenge, we propose FSD-VLN, an efficient fast-slow dualsystem framework that explicitly decouples high-level semantic reasoning from lowlatency action generation. The framework consists of two asynchronous pathways: a slow system that extracts stable semantic priors from a pretrained vision–language model, and a fast system built upon a Diffusion Transformer (DiT) that models action distributions for flight command generation. This decoupled design enables stable multi-step prediction and improves trajectory consistency by explicitly capturing cross-temporal action dependencies. Furthermore, a time-aware adaptive optimization strategy is designed to enhance long-horizon training stability and mitigate gradient oscillations during optimization. Extensive experiments in largescale simulated low-altitude environments demonstrate that FSD-VLN achieves up to 2 × higher success rate for navigation in unseen environments compared to prior arts, while reducing both single-action inference latency and overall task execution time by over 50% . These results highlight the importance of explicitly modeling the cooperation between semantic reasoning and temporally coherent control, providing a principled solution for long-horizon aerial VLN. Keywords: Vision-Language Navigation, Long-Horizon Modeling, Fast-Slow Dual-System, Low-Latency Decision Making FSD-VLN Instruction "Move upwards to a large white building , which is a tall and rectangular structure with balconies . Then , advance forward to it . As you slightly turn right and keep proceeding straight , you will reach it . Slightly turn left and head straight to it . Finally , slightly stop and drop off at it . Fast system BVLF: Buffer of vision-language features Figure 1: The navigation pipeline of the proposed FSD-VLN framework. During navigation, the slow system utilizes a pre-trained vision-language model to encode visual observations and language instructions into latent embeddings. The fast system adopts a variant of Diffusion Transformer (DiT), consisting of alternating cross-attention and selfattention modules, to process the UAV state embedding sequence and the visual-language embeddings produced by the slow system, respectively. The fast-slow dual system operates asynchronously to generate low-latency UAV actions. 1 Introduction Unmanned Aerial Vehicles (UAVs) are being extensively utilized in a broad spectrum of missions, from environmental monitoring to emergency response, requiring robust decision-making in different scenarios. Traditional UAV navigation often relies heavily on maps or manually designed control strategies[33, 40, 6, 32, 18, 44, 22], making it difficult to adapt to complex, dynamic, and semantically rich task requirements in open environments. Recently, the remarkable breakthroughs in large-scale foun

FAQs

What problem does FSD-VLN address in long-horizon aerial VLN?

It addresses the structural mismatch between global multimodal semantic understanding and temporally coherent action generation, which causes unstable trajectories and decision latency during long-horizon navigation.

How does FSD-VLN separate semantic reasoning from action generation?

It uses a slow system to extract stable semantic priors from a pretrained vision-language model and a fast system, based on a Diffusion Transformer variant, to model action distributions for low-latency flight command generation.

What helps FSD-VLN improve long-horizon training stability and reduce optimization issues?

A time-aware adaptive optimization strategy is used to enhance long-horizon training stability and mitigate gradient oscillations during optimization.

Jul 16, 2026

Upload by Aurelia