DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
DreamFly is a diffusion-based aerial vision-language navigation (VLN) framework built on Dream-VLA. It introduces causally aligned historical memory that augments the current visual representation using only observations preceding the current decision step. Navigation is formulated as receding-horizon diffusion planning, where the policy predicts a K-step action chunk but executes only the first action before replanning. LiteStop estimates stop probability directly from action logits at the initial all-mask state.
Aerial vision-language navigation requires an embodied agent to integrate visual evidence over time, plan future actions, and determine when it has reached a navigation goal under partial observability. DreamFly addresses challenges of limited historical context, short planning horizons, and unreliable implicit termination by introducing causal memory, receding-horizon diffusion planning, and LiteStop.
The causally aligned historical memory prevents future information leakage by using only past observations. The plan-K, execute-one strategy uses future actions as auxiliary planning targets while preserving closed-loop visual feedback. LiteStop decouples stop probability estimation from the diffusion process by using action logits at the initial all-mask state.
This research advances embodied AI for aerial robots, potentially improving autonomous drone navigation in partially observable environments. The framework builds on Dream-VLA, indicating a trend toward adapting vision-language-action models for specialized robotic domains.
Improved aerial VLN could enable more reliable autonomous drones for inspection, delivery, and search-and-rescue, reducing the need for human teleoperation and enabling operation in GPS-denied or visually complex environments.
Observable next signals include follow-up papers evaluating DreamFly on real-world aerial navigation benchmarks, integration with commercial drone platforms, and extensions to other embodied navigation tasks such as ground robots or underwater vehicles.