Aerial vision-language navigation (VLN) agents typically rely on large vision-language models, and querying them at every step is expensive. One way to reduce that cost is to predict a chunk of actions at once and execute the whole chunk before replanning. The authors report that this full-chunk approach cuts inference time from 195 ms to 42 ms per waypoint for their policy.
However, chunked execution creates a distribution shift: the policy is queried from states it reached itself, not from expert demonstrations, and those demonstrations do not teach it how to recover from such states or when to stop. On the OpenFly test-seen benchmark, failure to stop accounted for 47.78% of outcomes under full-chunk behavior cloning.
X-Fly addresses this with a staged pipeline. Recovery DAgger labels learner-visited decision states with smooth action chunks that reconnect to future expert segments and adds STOP at the first in-region state. Then Flow-GRPO jointly refines the chunked flow controller and a Bernoulli STOP policy using closed-loop outcomes from the same task, supplemented by outcome-verified stopping supervision.
With eight controls per decision, X-Fly achieves 74.18% success rate on test-seen and 51.73% on test-unseen, compared with 16.45% and 18.43% for full-chunk BC and 32.04% and 29.46% for DreamFly. The results suggest that explicitly training both reaching and stopping behaviors is key to making full-chunk aerial VLN practical.