Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation

A benchmark for hybrid UAV navigation and reasoning tasks.

FLIGHT closes the gap between long-range VLN and fine-grained UAV VLA tasks by combining multi-stage semantic missions with dense 6-DoF trajectory supervision.

Long-Horizon Flow

Extends atomic UAV-Flow instructions into compositional missions that combine traversal, translation, approach, orbiting, takeoff, landing, and task termination.

Fine-Grained VLN

Recasts visual-language navigation with dense landmark grounding and continuous action sequences instead of sparse discrete navigation decisions.

Pilot Reasoning

Adds textual supervision that summarizes the current flight state and anticipates the next subgoal, giving VLMs an explicit task-progress signal.

Task comparison between FLIGHT, traditional VLN, and UAV-Flow

Dense trajectories, video streams, and language annotations.

Demonstrations are collected in semantically rich Unrealzoo simulation scenes, aligned with FPV video, UAV state, 6-DoF pose, task semantics, and reasoning texts.

6,689

Fine-Grained VLN trajectories

Average trajectory length of 154.5 m with continuous actions sampled at 10 Hz.

4,099

Long-Horizon Flow trajectories

13,815 action instances with an average of 3.37 subtasks per trajectory.

6-DoF

Continuous action supervision

Records position and attitude, then converts adjacent states into relative motion commands.

Collect

Human pilots and simple-route planners produce expert UAV trajectories and aligned first-person flight videos.

Annotate

Motion metadata and segmented video clips guide VLM annotation of task semantics at fine temporal granularity.

Reason

Pilot Reasoning texts are generated from current and next-step task semantics, then verified for planning consistency.

FLIGHT data collection and annotation pipeline

Fast semantic planning, high-frequency continuous control.

FLIGHT VLA decouples a low-frequency Streaming Pilot VLM from a high-frequency diffusion action model, allowing long-context reasoning and responsive closed-loop control to coexist.

Streaming Pilot VLM

Consumes recent FPV video, UAV state, task instruction, and historical Pilot Reasoning memory to infer task progress and next-step plans.

Diffusion Action Model

Uses real-time visual features, relative UAV state, and the latest semantic feature to generate fine-grained future action chunks.

FLIGHT VLA asynchronous fast-slow framework

Closed-loop evaluation across long-horizon and fine-grained tasks.

FLIGHT VLA improves multi-stage task completion, subgoal adherence, terminal control, and UAV video reasoning over representative VLA and VLN baselines.

Long-Horizon Flow	SR	SSR	NDTW
OpenVLA-UAV	10.50	30.87	18.02
Memory VLA	18.50	38.69	16.58
FLIGHT VLA	59.00	78.97	45.78

Fine-Grained VLN	SR	OSR	NE	IASR
CMA+LAG	6.5	22.0	52.56	4.5
NaVid	13.0	25.5	46.69	5.0
FLIGHT VLA	13.5	37.0	45.52	11.0

Qualitative visualization of FLIGHT VLA navigation results

All metric values are reported as percentages except NE, which is measured in meters.

Citation

Please cite FLIGHT if this benchmark, dataset, or framework supports your research.

@misc{zheng2026flight,
  title         = {Think Like a Pilot: Fine-Grained Long-Horizon UAV Navigation},
  author        = {Zheng, Xiangyi and Wang, Xiangyu and Liao, Qinan and Tang, Zimu and Liao, Yue and Lyu, Dongyue and Wang, Guodong and Liu, Junjie and Liu, Si},
  year          = {2026},
  eprint        = {2606.06836},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.06836}
}