This post introduces DreamFly, a paper in computer vision by Yan Deng and Fei Xu, published on arXiv (2508.03419, 2026-08-13).
Research Area
Aerial vision-language navigation (VLN)
Overview
Aerial VLN requires an embodied agent to integrate visual evidence over time, plan future actions, and determine when a navigation goal has been reached under partial observability. While recent VLA models offer a promising perception-to-action paradigm, adapting them to aerial navigation remains challenging due to limited historical context, short planning horizons, and unreliable implicit termination.
Key Contributions
DreamFly is a diffusion-based aerial VLN framework built on Dream-VLA, with three main components:
- Causal aligned memory: enhances the current visual representation using only observations from decision steps *before* the current one, enabling temporal reasoning without leaking future information.
- Receding-horizon diffusion planning: navigation is formalized so the policy predicts K-step action chunks but executes only the first action before replanning. This "plan K steps, execute one" strategy uses future actions as auxiliary planning objectives while retaining closed-loop visual feedback.
- LiteStop: estimates stop probability directly from action logits initialized from a fully masked state, decoupling explicit termination from action generation.
Results
Experiments on the OpenFly benchmark show consistent improvements in both seen and unseen environments:
| Split | SR | SPL | |---|---|---| | Test seen | 32.04% | 28.22% | | Test unseen | 29.46% | 23.54% |
DreamFly outperforms all compared methods on both metrics and achieves the lowest navigation error. These results demonstrate the effectiveness of jointly modeling historical context, future action structure, and explicit termination for aerial VLN.
Original Abstract
> Aerial vision-language navigation (VLN) requires an embodied agent to integrate visual evidence over time...
Link: https://arxiv.org/abs/2508.03419