AeroRSI-2B

Policies of AeroRSI: Recursive Self-Improvement for Aerial Vision-Language Navigation.

Code · Paper · Daily Paper · Project page

AeroRSI lets an aerial navigation policy learn from its own failures, round after round. At each state shortly before a failure, the policy flies on from every candidate action (its own samples and rule-based proposals) to judge which one it can actually complete; it then learns the verified repairs while keeping its behavior elsewhere, and repeats on the failures that remain.

Checkpoints

Folder Starts from Seen SR Unseen Object SR Unseen Map SR Overall SR
AeroRSI-2B-SFTinit the supervised policy AeroVLA-2B-384 75.32 82.19 46.24 67.49
AeroRSI-2B-DPOinit AeroDPO 76.52 79.65 53.55 69.85

Both are after three rounds, evaluated on the 3,005 test routes of TravelUAV with the benchmark's native actions and termination. For reference, the supervised policy reaches 46.39 and AeroDPO 56.87 overall under the same evaluator.

Each folder is a LoRA adapter (rank 64 on the seven language projections, plus the visual merger) for the supervised policy AeroVLA-2B-384-Merged released in XuPeng23/AeroDPO.

Usage

git clone https://github.com/starsholic/AeroRSI.git && cd AeroRSI
bash scripts/setup_env.sh && conda activate aerorsi
hf download XuPeng23/AeroDPO --include "AeroVLA-2B-384-Merged/*" --local-dir checkpoints
hf download DavidWu1116/AeroRSI --include "AeroRSI-2B-SFTinit/*" --local-dir checkpoints
aerorsi prepare --dataset-root dataset_raw          # TravelUAV routes, see docs/INSTALL.md
aerorsi evaluate --adapter checkpoints/AeroRSI-2B-SFTinit --output outputs/eval --gpus 0 1 2 3
aerorsi metrics --runs outputs/eval

The policy takes a 384×768 mosaic of the front and down views and the prompt Fly <direction> and find the target. <object description>\nAction: and answers with three two-digit bins for forward, down and yaw, optionally followed by LAND. Load the adapter with aerorsi.policy.Policy, which keeps the visual merger in FP32 under BF16 autocast as in training.

Citation

@article{wu2026aerorsi,
  title   = {AeroRSI: Recursive Self-Improvement for Aerial Vision-Language Navigation},
  author  = {Wu, Shiwei and Pan, Yuwen and Li, Zhi and Li, Yaowen and Zhou, Yan and Xu, Xuemiao and Liu, Yu},
  journal = {arXiv preprint arXiv:2611.24077},
  year    = {2026}
}
Downloads last month
-
Video Preview
loading

Model tree for DavidWu1116/AeroRSI

Base model

XuPeng23/AeroDPO
Adapter
(1)
this model

Paper for DavidWu1116/AeroRSI