FineCog-Nav: Fine-grained Cognitive Modules for Zero-shot UAV Vision-Language Navigation

Forum topic · 小凯 · 2026-04-21

Summary

FineCog-Nav is a zero-shot UAV vision-language navigation (VLN) framework inspired by human cognition, presented in an arXiv paper by Dian Shao and colleagues. The top-down framework decomposes navigation into fine-grained modules covering language processing, perception, attention, memory, imagination, reasoning, and decision-making. Each module is powered by a moderate-sized foundation model with role-specific prompts and structured input-output protocols, improving collaboration and interpretability compared to approaches that rely on large base models with generic prompts and loosely coordinated components. The authors also introduce AerialVLN-Fine, a benchmark of 300 curated trajectories from AerialVLN featuring sentence-level instruction-trajectory alignment and refined instructions with explicit visual endpoints and landmark references. Experiments show FineCog-Nav consistently outperforms zero-shot baselines in instruction following, long-horizon planning, and generalization to unseen environments, demonstrating the effectiveness of fine-grained cognitive modularity for zero-shot aerial navigation.

Paper Overview

Field: Computer Vision (CV) Authors: Dian Shao, Zhengzheng Xu, Peiyang Wang, Like Liu, Yule Wang, Jieqi Shi, Jing Huo arXiv: 2604.16298 Project page: https://smartdianlab.github.io/projects-FineCogNav

Abstract

UAV vision-language navigation (VLN) requires an agent to navigate complex 3D environments from an egocentric perspective while following ambiguous multi-step instructions over long horizons. Existing zero-shot methods remain limited, as they often rely on large base models, generic prompts, and loosely coordinated modules.

In this work, the authors propose FineCog-Nav, a top-down framework inspired by human cognition that organizes navigation into fine-grained modules for:

Language processing
Perception
Attention
Memory
Imagination
Reasoning
Decision-making

Each module is driven by a moderate-sized foundation model with role-specific prompts and structured input-output protocols, enabling effective collaboration and improved interpretability.

AerialVLN-Fine Benchmark

To support fine-grained evaluation, the authors construct AerialVLN-Fine, a benchmark of 300 curated trajectories from AerialVLN with:

Sentence-level instruction-trajectory alignment
Refined instructions with explicit visual endpoints and landmark references

Results

Experiments show that FineCog-Nav consistently outperforms zero-shot baselines in instruction following, long-horizon planning, and generalization to unseen environments. These results demonstrate the effectiveness of fine-grained cognitive modularity for zero-shot aerial navigation.

---

*Auto-collected on 2026-04-21.*