X-Fly improves aerial drone navigation with smarter chunked actions
X-Fly: Learning to Reach and Stop in Full-Chunk Aerial Vision-Language Navigation
RoboticsArtificial IntelligenceComputer Vision and Pattern Recognition
Summary
Navigation for aerial drones using large vision-language models is slow because these models take time to process each waypoint. The authors found that grouping waypoints into chunks speeds things up but causes problems when the drone doesn't know when to stop or how to recover from unexpected states. They developed X-Fly, a two-step training method that teaches drones to safely follow chunked instructions and decide when to stop. This approach almost doubles success rates in familiar and new environments compared to earlier methods.
What this means in practice
- •For drone software engineers: Enable more efficient and accurate drone navigation by improving decision making on when to replan and stop during long flight paths.
- •For urban planning teams: Provide aerial navigation tools that better handle dynamic environments and unpredictable states for city-scale drone mapping projects.
Authors
Ruibin Ju, Ruifeng Cai, Min Xu, Lin Wang
Abstract
Aerial vision-language navigation (VLN) agents built on large vision-language models are expensive to query. Executing a complete predicted action chunk before replanning amortizes this cost, reducing model inference from 195 to 42 ms per executed waypoint for our policy. It also moves the states at which the policy is queried away from expert demonstrations, which supervise neither recovery from these states nor termination at the learner's own arrival. On OpenFly test-seen, no-stop failures account for 47.78% of full-chunk behavior cloning (BC) outcomes. We present X-Fly, a staged training framework for full-chunk aerial VLN. Recovery DAgger labels learner-visited decision states with smooth chunks that reconnect to future expert segments and with STOP at the first in-region state. Flow-GRPO then jointly refines the chunked flow controller and a Bernoulli STOP policy from same-task closed-loop outcomes, supplemented by outcome-verified stopping supervision. With eight controls per decision, X-Fly reaches 74.18% SR 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.