AeroRSI: Recursive Self-Improvement for Aerial Vision-Language Navigation
Authors: Shiwei Wu, Yuwen Pan, Zhi Li, Yaowen Li, Yan Zhou, Xuemiao Xu, Yu Liu
Organizations: SIGS, Tsinghua University · Department of Electronic Engineering, Tsinghua University · Foshan University · South China University of Technology
Vision-language-action (VLA) models enable unmanned aerial vehicles (UAVs) to follow natural-language instructions, yet still fail frequently in closed-loop flight. Correcting these failures requires repairs, i.e., actions that avoid them at the states the policy itself reaches, which human pilots cannot label at scale and rule-based planners cannot supply reliably. In this paper, we present AeroRSI, a recursive self-improvement framework in which an aerial navigation policy iteratively learns from its own failures. Our key insight is that a failing policy should judge its own repairs by flying on from them. Judged in this way, its sampled actions and even rule-based proposals that hurt it when imitated directly become reliable supervision. Moreover, an episode terminates at its first failure, concealing subsequent failures until the earlier ones are repaired. AeroRSI thus renews its supervision in every round, anchoring the failures of the current policy, verifying repairs by its continuation, and distilling them while keeping its behavior elsewhere. On TravelUAV, three rounds raise the success rate of a supervised 2B-parameter policy from 46.4% to 67.5% and reduce collisions by half, surpassing the previous state of the art by 10.6 points from the same initialization without extra human labels or inference cost. Code and models are available at https://github.com/starsholic/AeroRSI.
Figures & tables
Figure 2: Overview of AeroRSI . The top row shows the policy πk , of which only the visual merger and the LoRA are trained. In each round, (1) AFA samples anchors from probes before every failure of πk ; (2) CVR restores each anchor, lets πk continue the executed action a− , its sampled actions, and rule-based proposals, and accepts as the repair c⋆ the first candidate that qualifies again in a re-flight; and (3) RRL distills the verified repairs into a student copy of πk , keeping the outputs of πk on states that need no repair, to obtain πk+1 .
Policy
Initialized from
Trained on
Supervised policy (SFT) [ 1 ]
Qwen3-VL-2B [ 104 ]
Expert demonstrations
AeroDPO [ 1 ]
SFT
Collision corrections
AeroRSI (SFT init.)
SFT
Verified repairs
AeroRSI (DPO init.)
AeroDPO
Verified repairs
Table 1: Compared policies. They share one 2B-parameter model with fuzzy direction guidance; the supervised policy (SFT) imitates the expert demonstrations of AerialVLA [ 9 ] , and the suffix of AeroRSI names its initial policy.
Method
Guide
Seen
Unseen Object
Unseen Map
Overall
NE ↓
SR ↑
OSR ↑
SPL ↑
NE ↓
SR ↑
OSR ↑
SPL ↑
NE ↓
SR ↑
OSR ↑
SPL ↑
SR ↑
TravelUAV [ 3 ]
L1
106.28
16.10
44.26
14.30
118.11
22.42
46.90
20.51
138.80
4.18
20.77
3.84
13.62
NavFoM [ 8 ]
L1
93.05
29.17
49.24
25.03
108.04
29.83
47.99
27.20
125.10
6.30
18.95
5.68
22.02
LongFly [ 17 ]
L1
60.02
36.39
65.87
31.07
66.74
43.87
64.56
38.39
108.32
11.27
30.27
9.32
29.95
NeuroKalman [ 58 ]
L1
71.56
25.86
58.73
22.43
71.01
32.48
60.82
28.50
100.32
8.34
34.15
7.12
21.66
AerialVLA [ 9 ]
FD
65.88
47.96
57.69
38.54
61.45
56.60
64.86
46.61
67.42
37.58
52.92
28.22
46.46
Table 2: Comparison with existing methods on the TravelUAV test splits. L1: L1 assistant guidance of [ 3 ] ; FD: fuzzy direction guidance of [ 1 ] . AeroRSI (SFT init.) starts from the same supervised policy as AeroDPO, and AeroRSI (DPO init.) from AeroDPO (Table 1 ); both after three rounds. Overall: SR over all test routes, weighted by split sizes where only per-split results exist. Best results are in bold.
Round
Seen
UO
UM
SR ↑
SPL ↑
CR ↓
Train SR ↑
AeroRSI (SFT init.)
0
48.38
56.44
36.85
46.39
37.37
38.64
45.24
1
62.62
67.89
43.42
57.60
46.96
29.48
61.41
2
73.91
79.01
45.82
66.02
51.96
21.26
71.41
3
75.32
82.19
46.24
67.49
51.00
19.10
70.85
AeroRSI (DPO init.)
Table 3: Results across rounds. Seen, UO, and UM give the SR on each split; SR, SPL, and CR are computed over all test routes. Train SR: SR on the training routes. Round 0 is the initial policy. The best result of each setting is in bold.
Figure 3: Failure types across rounds. Share of test routes that end in each type of failure for the policy of every round of AeroRSI (SFT init.), with the total failure rate after each bar.
Table 4: Ablation study on all test routes. Variants in (a) and (c) are compared with the first round and those in (b) with the second. In (a), unverified variants label every anchor without flying the label, verified variants keep only the labels that the policy completes, and the two-action variant scores the candidates after two actions instead. Paired significance tests are in the supplementary material.
Figure 4: Analysis of the design of AeroRSI . (a) SR and CR after the first round when the labels are verified by the policy (filled) or not (hollow), as in Table 4 (a). (b) Pass@ k of newly sampled actions of each policy at the reproducible anchors of the first round, at all of them (solid) and at those that only a rule-based proposal repaired (dashed). (c) Test SR when a round learns from the rule-based proposals, the sampled actions, or both, each single-source round starting from the previous round with both sources. (d) At the anchors accepted in the first round, the verified repair and the action at which π0 failed, each continued by a later policy ( π0 continues the repair again). (e) Distance between the failure points of two flights on the training routes on which both fail. (f) SR gain of π3 over π0 on 999 test routes with and without an altitude cap. Error bars are 95% confidence intervals, paired in (f).
Figure 5: Case study of AeroRSI . From a state before π0 crashes, π0 flies on from each candidate action (top), and π1 flies after one round (bottom). (a) Flying toward the target looks best at first but crashes, whereas a turn that falls behind completes the flight, so a repair can be judged only by flying on. (b) The sampled actions nearly repeat the crashing action, so only a rule-based proposal repairs it, and π1 learns this repair. (c) The same action crashes when π0 flies on but not when π1 does, so every round must judge its repairs anew.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Setting
Policy
Qwen3-VL-2B; 384×768 mosaic of the front and downward views; forward, vertical, and yaw components with 99 bins each, plus landing
Training
LoRA of rank 64 on all linear layers of the language model, and the visual merger; 2 epochs per round; learning rate 2×10−5 ; batch size 16
AFA
Offsets of 1, 2, 4, 8, and 16 actions; 2,400 anchors per round; horizon of max(8,δ+4) actions for offset δ , at most 20
CVR
M=4 samples at temperature 1.0 and three rule-based proposals; λf=100 , λs=20 ; minimum progress of 4 m; margin m=4 m; two confirmations at most
RRL
λn=0.2 on 1,990 retention states; replay share ρ=30% from round 2
Appendix
Table S1 : Settings of AeroRSI . They are shared by both initializations and all rounds.
Round
1
2
3
Flights of the training routes
Flights
7,611
7,611
7,611
Actions (thousands)
345
376
435
Verification
Flights of candidates
15,237
13,975
12,451
Confirmation flights
1,484
1,342
1,165
Appendix
Table S2 : Simulator cost of each round of AeroRSI (SFT init.). Round k flies the training routes with πk−1 .
Round
Collision
Wrong way
Stuck
Early landing
Timeout
AeroRSI (SFT init.)
0
38.6
7.8
4.2
1.8
1.2
1
29.5
3.8
4.2
2.2
2.8
2
21.3
2.6
2.8
2.3
5.0
3
19.1
2.1
2.7
1.9
6.8
AeroRSI (DPO init.)
Appendix
Table S3 : Failure types across rounds. Share (%) of all test routes that end in each type of failure; the other routes succeed (Table 3 of the main paper).
Comparison
Δ SR
z
Rounds and baselines
Round 1 vs. round 0
+11.21
10.4
Round 2 vs. round 1
+8.42
8.1
Round 3 vs. round 2
+1.46
1.6
Round 3 vs. round 0
+21.10
18.5
Round 3 vs. AeroDPO
+10.62
10.2
Appendix
Table S4 : Paired tests on the test routes. Difference in SR (points) and z over the routes common to both runs, for AeroRSI (SFT init.) unless stated.
Round
1
2
3
Anchors
2,390
2,390
2,399
Share of all anchors (%)
Repair accepted
43.8
38.3
34.3
Repair accepted, DPO init.
46.1
38.8
33.9
Share of the anchors whose failure reproduces (%)
Best of four samples qualifies
66.2
68.3
68.3
Appendix
Table S5 : Statistics of the loop of AeroRSI (SFT init.) unless stated. Anchors: those evaluated of the 2,400 drawn.
Round
Sampled actions
Potential field
Translation
Climb
1
39.2
60.7
59.3
47.6
2
44.3
55.4
59.6
46.2
3
51.0
63.6
59.0
51.2
Appendix
Table S6 : Candidates that qualify (%) at the anchors whose failure reproduces. Identical samples of an anchor count once.
π0
π1
π3
Anchors repaired by
@1
@4
@1
@4
@1
@4
All (100%)
33.4
61.7
63.2
83.6
73.8
86.6
A sampled action (66%)
45.6
79.9
71.7
90.9
80.4
92.2
Only a proposal (21%)
11.0
30.4
62.1
84.3
76.7
87.9
Neither (13%)
6.5
18.9
20.9
44.3
34.6
55.1
Held-out route groups
Appendix
Table S7 : Pass@ k of new samples (%) at the 1,450 reproducible anchors of the first round, grouped by what repaired them during collection.
Continued by
π0 again
π1
π2
π3
Verified repair
81.5
86.9
87.0
89.1
held out (211)
82.5
88.2
91.9
92.9
from sampled actions (460)
77.6
83.3
85.4
88.0
from proposals (588)
84.5
89.8
88.2
90.0
Failed action
26.5
57.0
63.3
66.6
Appendix
Table S8 : Verified repairs under later policies. Share (%) of the 1,048 anchors accepted in the first round at which the repair, or the action at which π0 failed, qualifies.
Flights
Routes
Within 10 m
Over 50 m
Median
π0 , then π1
2,125
27.9
44.8
40.2 m
π1 , then π2
1,333
33.0
37.6
28.0 m
One policy twice
56
50.0
30.4
12.7 m
Appendix
Table S9 : Distance between the failure points of two flights on the routes that both fail: share (%) within 10 m and over 50 m, and the median. One policy flown twice is an earlier policy on 384 held-out routes.
Foshan Graduate School of Innovation, Northeastern University, Foshan 528311, China · College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · Faculty of Robot Science and Engineering, Northeastern University, Shenyang 110819, China +3