DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation

Analysis

Why This Paper Matters

Aerial vision-language navigation (VLN) is a challenging task that requires an agent to integrate visual information over time, plan actions, and decide when to stop, all under partial observability. While recent VLA models have shown promise in perception-to-action paradigms, they often struggle with limited historical context, short planning horizons, and unreliable termination in aerial settings. DreamFly directly tackles these issues, making it a significant step toward more capable aerial navigation systems.

The paper's focus on causal memory and receding-horizon planning is particularly relevant as it addresses the temporal and sequential nature of navigation. By preventing future information leakage and using future actions as auxiliary targets, DreamFly improves both training efficiency and closed-loop performance. This work is likely to inspire further research in diffusion-based planning for embodied AI, especially in domains where long-horizon decision-making is critical.

Technical Contributions

• Causally Aligned Historical Memory: DreamFly augments current visual representations with only past observations, ensuring temporal reasoning without future leakage. This is crucial for real-world deployment where future information is unavailable.
• Receding-Horizon Diffusion Planning: The policy predicts a K-step action chunk but executes only the first action, then replans. This plan-K, execute-one strategy leverages future actions as auxiliary targets while maintaining closed-loop feedback, balancing foresight and adaptability.
• LiteStop for Explicit Termination: LiteStop estimates stop probability directly from action logits at the initial all-mask state, decoupling termination from action generation. This simplifies the stopping decision and improves reliability.
• Integration with Dream-VLA: The framework builds on Dream-VLA, demonstrating how diffusion-based VLA models can be adapted for aerial navigation with targeted enhancements.

Results

DreamFly was evaluated on the OpenFly benchmark, showing consistent improvements in both seen and unseen environments. Specifically, it achieved 32.04% SR and 28.22% SPL on test-seen, and 29.46% SR and 23.54% SPL on test-unseen, outperforming all compared methods on both metrics. Additionally, it attained the lowest navigation error, indicating more precise goal reaching. These results highlight the effectiveness of jointly modeling historical context, future action structure, and explicit termination.

Significance

The success of DreamFly demonstrates that addressing temporal reasoning and planning horizon can significantly boost aerial VLN performance. This work provides a blueprint for integrating causal memory and diffusion planning in embodied AI, which could extend to other navigation tasks, such as ground robots or autonomous driving. By decoupling termination from action generation, it also offers a more modular approach to decision-making. As aerial robots become more prevalent in search-and-rescue, delivery, and inspection, DreamFly's contributions could enable more reliable and efficient autonomous navigation in complex environments.