DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
DreamFly reports stronger aerial VLN results by combining causal memory, plan-ahead diffusion, and a separate stop signal.
The framework adds historical visual context without using future observations, then predicts a K-step action chunk while executing only the first action before replanning. Its LiteStop component estimates stopping directly from initial action logits, separating termination from action generation. On OpenFly, the paper says DreamFly beats compared methods on success rate and SPL in both seen and unseen test splits, with the lowest navigation error. ArXiv · AI/CL/LG's note
The framework adds historical visual context without using future observations, then predicts a K-step action chunk while executing only the first action before replanning. Its LiteStop component estimates stopping directly from initial action logits, separating termination from action generation. On OpenFly, the paper says DreamFly beats compared methods on success rate and SPL in both seen and unseen test splits, with the lowest navigation error. ArXiv · AI/CL/LG's note
score 5