Venue: ICLR 2026 · Authors: Xiaolou Sun, Wufei Si, Wenhui Ni, Yuntian Li, Dongming Wu, Fei Xie, Runwei Guan, He-Yang Xu, Henghui Ding, Yuan Wu, Yutao Yue, Yongming Huang, Hui Xiong · arXiv: 2602.09657 · Category: Embodied navigation / VLN (aerial) · Trend tag: Vision-Language-Action models for autonomous UAV flight.

flowchart LR
  RGB[Onboard RGB camera] --> PDE[Pseudo-depth encoder<br/>depth-aware features from RGB]
  RGB --> VE[Visual encoder]
  Inst[Coarse directional guidance] --> VLA
  PDE --> VLA
  VE --> VLA
  subgraph VLA["AutoFly VLA model"]
    Stage1[Stage 1: align vision + depth + language] --> Stage2[Stage 2: action grounding]
  end
  VLA --> Act[UAV action<br/>continuous planning + obstacle avoidance]

VLN research for UAVs typically assumes detailed, pre-specified step-by-step instructions. But real outdoor exploration happens in unknown environments where such instructions are unavailable: the drone has only coarse directional guidance and must navigate autonomously through continuous planning and obstacle avoidance. This is the gap AutoFly targets.

AutoFly is an end-to-end Vision-Language-Action (VLA) model for autonomous UAV navigation:

• Pseudo-depth encoder: derives depth-aware spatial features directly from RGB imagery, improving spatial reasoning without a real depth sensor.
• Progressive two-stage training: first aligns visual representations, depth information, and language understanding; then grounds these into navigation actions.
• A new autonomous-navigation dataset emphasizing obstacle avoidance and continuous planning (rather than instruction-following), incorporating real-world data.

AutoFly reports a +3.9% success rate over state-of-the-art VLA baselines, with consistent performance across both simulated and real-world environments. The authors state model, data, and code are released. (Exact simulator and per-task numbers omitted here pending the camera-ready tables.)

AutoFly reframes aerial VLN away from instruction-following toward autonomous flight in the wild under sparse guidance, and shows that RGB-only pseudo-depth can substitute for true depth sensing in a UAV VLA. It extends the VLA paradigm — dominant in manipulation and ground navigation — into the aerial embodiment.

• arXiv: https://arxiv.org/abs/2602.09657
• OpenReview: https://openreview.net/forum?id=88RKxlFUNY

• ICLR 2026 Survey
• OmniVLA (navigation)
• NavFoM (navigation foundation model, includes drone embodiment)

← Back to ICLR-2026