AeroRSI: Recursive Self-Improvement for Aerial Vision-Language Navigation

Shiwei Wu1Yuwen Pan1Zhi Li1 Yaowen Li2Yan Zhou3Xuemiao Xu4 Yu Liu2,†
1Shenzhen International Graduate School, Tsinghua University 2Department of Electronic Engineering, Tsinghua University 3Foshan University 4South China University of Technology †Corresponding author
Demo reelVideo coming soon

One round of repair at one failure, three generations of the policy on one test route, and the whole test set. Translucent drones are separate recorded flights composited along one camera path. Music: “The Spectre” by Alan Walker (instrumental).

TL;DR

A drone navigation policy repairs its own failures: before each failure, it flies on from every candidate action, keeps the actions it can complete, learns them, and repeats on the failures that remain.

46.4 → 67.5%test success rate, three rounds
38.6 → 19.1%collision rate, halved
+10.6points over AeroDPO (prior SOTA baseline), with no extra labels or inference cost

Abstract

Vision-language-action drone policies still fail often in closed-loop flight, and the repairs they need are beyond what human pilots can label or rule-based planners can reliably supply. AeroRSI lets the policy judge its own repairs by flying on from them: it keeps only the candidate actions it completes, so even rule-based proposals that hurt when imitated directly become reliable supervision. Since each repair can reveal the next failure, AeroRSI repeats on the failures that remain. On TravelUAV, three rounds lift a 2B-parameter policy from 46.4% to 67.5% success, 10.6 points above AeroDPO (prior SOTA baseline), with no extra human labels or inference cost.

Motivation

Each repair uncovers the next failure

Round k: the policy fails at A and judges its candidate repairs by flying on; after learning them, the next policy passes A and fails later at B; test success rate rises over three rounds.
(a) The policy fails at A and keeps only the candidate repairs it completes when flying on. (b) Having learned them, it passes A and fails at B, which A had hidden. (c) Test success rate by round; dashed: the previous state of the art from the same policy.
Method

One round of AeroRSI

Framework: adaptive failure anchoring, continuation-verified repair and retentive repair learning, repeated with the new policy.
1 · Anchor

Adaptive failure anchoring

The policy flies every training route. Anchors are drawn 1 to 16 actions before each failure, so they track the failures that remain.

2 · Verify

Continuation-verified repair

From each anchor, the policy flies on after every candidate: the failed action, its own samples and rule-based proposals. A candidate that completes twice becomes a repair.

3 · Learn

Retentive repair learning

The next policy learns the repairs, replays earlier ones, and keeps its own actions at normal states.

Front view 16 actions before a crashtf − 16
Front view 8 actions before a crashtf − 8
Front view 4 actions before a crashtf − 4
Front view 2 actions before a crashtf − 2
Front view 1 action before a crashtf − 1
Front view at the crashtf crash
Probes before one crash. Early probes may already be committed to the crash and late ones too late; several offsets let the verifier find where a repair still works.
Demos

A policy that improves from its own flights

1 · One round

It judges its own repairs by flying on

Before a crash, GEN 0 flies on after four of its own samples and one rule-based climb. The samples crash; the climb completes and becomes the repair. Trained on such repairs, GEN 1 climbs by itself.

Where only a proposal repaired an anchor, a new sample of the policy succeeds 62.1% of the time after one round, up from 11.0%.

One round at one anchorVideo coming soon
2 · Recurse

Each repair uncovers the next failure

The supervised policy crashes at A. After one round it passes A and crashes at B, which A had hidden; after two rounds it reaches the target.

Rounds 0 to 2 on one routeVideo coming soon
3 · Population

The whole test set improves

Each square is one of the 3,005 test routes: green reaches the target, red collides, amber fails otherwise. Success rises from 46.4% to 67.5%, collisions fall from 38.6% to 19.1%.

3,005 test routesVideo coming soon

Each drone is a recorded flight of a policy in TravelUAV, replayed from its poses; translucent drones are separate flights composited along one camera path. Test routes follow the evaluation protocol; candidates at an anchor fly under the paused collection cadence.

Results

TravelUAV test splits

Each round raises the success rate, with diminishing gains. Started from AeroDPO (prior SOTA baseline), the same loop adds 13.0 points.

SR (%)SeenUnseen obj.Unseen mapOverall
Supervised policy48.3856.4436.8546.39
AeroDPO (prior SOTA baseline)60.0166.4545.9356.87
AeroRSI (SFT init.)75.3282.1946.2467.49
AeroRSI (DPO init.)76.5279.6553.5569.85

Overall success rate by round

From the supervised policyFrom AeroDPO (prior SOTA baseline)
40506070 Test SR (%) 0123 Round 46.456.957.663.4 67.569.9

Both starting points gain in every round.

How test episodes end

Share of test routes ending in each failure type for rounds 0 to 3.

Collisions and wrong-way flights fall; timeouts grow.

BibTeX

@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}
}