X-Fly: Learning to Reach and Stop in Full-Chunk Aerial Vision-Language Navigation
This research presents X-Fly, a staged training framework for full-chunk aerial vision-language navigation. By executing complete predicted action chunks before replanning, the method reduces model inference time from 195 ms to 42 ms per waypoint and improves recovery from non-expert states. The approach combines Recovery DAgger labeling with Flow-GRPO refinement to jointly optimize a chunked flow controller and a Bernoulli STOP policy. On the OpenFly test-seen dataset, X-Fly achieves a 74.18% success rate compared to 16.45% for full-chunk behavior cloning and 32.04% for DreamFly.