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

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

FSD-VLN, a novel Fast-Slow Dual-System framework, addresses critical challenges in long-horizon Vision-Language Navigation (VLN) for unmanned aerial vehicles (UAVs). It explicitly decouples high-level semantic reasoning from low-latency action generation, resolving the structural inconsistency common in prior approaches. The framework features a slow system that extracts stable semantic priors from a pretrained vision-language model and a fast system, built on a Diffusion Transformer (DiT), which models action distributions for flight command generation. This design improves trajectory consistency by capturing cross-temporal action dependencies. Evaluated in large-scale simulated low-altitude urban environments, FSD-VLN achieves up to 2x higher success rates in unseen environments, improving SR from 5.1% to 13.6% and SPL from 3.5% to 10.7% compared to OpenFly. It also reduces single-action inference latency by over 50% (from 402 ms to 176 ms) and overall task execution time by 53% (from 307.6 s to 144.7 s).

Key takeaway

For Robotics Engineers developing long-horizon UAV navigation systems, you should consider adopting a decoupled fast-slow architecture like FSD-VLN. This approach effectively balances deep semantic reasoning with low-latency control, overcoming the stability and efficiency trade-offs of traditional methods. Implementing a Diffusion Transformer for action generation and time-weighted optimization can significantly improve navigation success rates and reduce execution times in complex, unseen environments.

Key insights

Decoupling high-level semantic reasoning from low-latency action generation enables stable, efficient long-horizon aerial navigation.

Principles

• Asynchronous dual-system architectures balance deep perception and real-time control.
• Explicitly modeling cross-temporal action dependencies enhances trajectory consistency.
• Time-weighted optimization stabilizes gradient propagation in long-sequence learning.

Method

FSD-VLN employs an asynchronous slow system (VLM for semantic priors) and a fast system (Diffusion Transformer for action distributions), combined with time-weighted MSE loss and global adaptive normalization.

In practice

• Consider dual-system designs for robotics needing both complex reasoning and fast control.
• Apply Diffusion Transformers for generative action modeling in sequential decision tasks.
• Utilize time-weighted loss functions to improve stability in long-horizon training.

Topics

• Vision-Language Navigation
• Unmanned Aerial Vehicles
• Fast-Slow Dual-System
• Diffusion Transformer
• Long-Horizon Navigation
• Low-Latency Control

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• FSD-VLN: Fast-Slow Dual-System Modeling for Aerial Long-Horizon Vision-Language Navigation — Decoupling semantic reasoning and flight control via a fast-slow dual-system significantly enhances aerial Vision-Language Navigation performance.
• FSD-VLN: Fast-Slow Dual-System Modeling for Aerial Long-Horizon Vision-Language Navigation — Decoupling semantic reasoning from low-latency flight control significantly enhances long-horizon aerial Vision-Language Navigation.
• Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation — An asynchronous VLM and diffusion model architecture enables fine-grained, long-horizon UAV navigation with real-time continuous control.
• LiteVLA-H: Dual-Rate Vision-Language-Action Inference for Onboard Aerial Guidance and Semantic Perception — Edge VLA latency is pre-fill dominant, necessitating dual-rate scheduling for reactive guidance and semantic awareness.
• CLEAR: Cognition and Latent Evaluation for Adaptive Routing in End-to-End Autonomous Driving — CLEAR achieves efficient, multi-modal autonomous driving by combining single-step generative planning with LLM-driven cognitive reasoning.
• Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models — Vision-Language-Action models can achieve strong one-step action generation by biasing diffusion training towards high-noise states.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.