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

Shiwei Wu, Yuwen Pan, Zhi Li, Yaowen Li, Yan Zhou, Xuemiao Xu, Yu LiuarXiv 2026 · arXiv

TL;DR

AeroRSI is a recursive self-improvement framework that allows aerial drones to learn from their own navigation failures without human intervention. By having a model "test-fly" potential corrections and only keeping those that lead to successful mission completion, the system iteratively reveals and fixes hidden failures, boosting success rates on the TravelUAV benchmark from 46.4% to 67.5%.

What Problem Does This Paper Address?

Vision-Language-Action (VLA) models allow drones to follow natural-language instructions (e.g., "Fly past the red building and land near the fountain"). However, these models frequently fail in "closed-loop" flight—meaning they make a small mistake, end up in a state they weren't trained for, and then crash or get stuck.

Correcting these errors is difficult for two reasons:

  1. Labeling Scarcity: Human pilots cannot easily provide labels for the infinite variety of "mistake states" a drone might reach.
  2. Hidden Failures: An episode ends at the first failure (like a collision). This hides subsequent failures further down the path. You can't fix a mistake at the end of a flight if the drone always crashes at the beginning.

Methods

The authors introduce AeroRSI, a framework that treats self-improvement as a recursive loop. Instead of relying on an external expert, the failing policy itself acts as the judge of its own "repairs."

The process consists of three main modules per round:

  1. Adaptive Failure Anchoring (AFA): The system identifies "anchor states" leading up to a failure. It places probes at different time scales—some near the crash to catch immediate causes, and some further back to catch poor strategic decisions.
  2. Continuation-Verified Repair (CVR): For every anchor state, the model generates candidate actions from its own samples or simple rule-based proposals (like "climb higher"). Crucially, it verifies these candidates by executing the action and letting the current policy attempt to finish the flight. If the drone reaches the goal, that candidate action is saved as a "verified repair."
  3. Retentive Repair Learning (RRL): The model is updated using the verified repairs. To prevent "catastrophic forgetting" or drifting away from good behaviors, it uses a retention loss that anchors the model to its own successful greedy outputs on states that didn't need repair.
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.

Datasets

The researchers evaluated AeroRSI on TravelUAV, a high-fidelity benchmark for aerial navigation built on AirSim and Unreal Engine. It includes:

  • 7,611 training routes across 11 diverse environments.
  • Test sets categorized by "Seen" environments, "Unseen Objects," and "Unseen Maps" to test generalization.

Key Insights and Ideas

  • The Policy as its Own Judge: The "aha moment" is that a failing policy is actually a great judge of repairs. If a candidate action allows the current (imperfect) policy to successfully complete the rest of the flight, that action is a highly reliable supervision signal.
  • Recursion is Essential: As shown in Figure 1, fixing an early failure (A) reveals a "hidden" failure (B) further along the path. By repeating the process across multiple rounds, the model "peels the onion" of its own incompetence.
  • Turning Weak Signals into Gold Labels: Simple rule-based proposals (like a potential-field avoidance move) might actually hurt performance if imitated blindly. However, once verified by a successful flight continuation, they become high-quality training data.
Figure 1. Recursive self-improvement with self-verified repairs. (a) In round k, πk fails at A and judges its own candidate repairs. From the state before A, each candidate from its sampled actions or rule-based proposals is continued by πk, and only those it completes are kept. (b) After learning these repairs, πk+1 avoids A and reaches B, a failure concealed by the earlier one at A, where the next round verifies new repairs. (c) Test success rate on TravelUAV; the dashed line is the previous state of the art [1] from the same supervised policy.

Key Results

  • Significant Performance Boost: Starting from a supervised 2B-parameter model, AeroRSI raised the success rate from 46.4% to 67.5% in just three rounds.
  • New State-of-the-Art: It outperformed the previous best method (AeroDPO) by 10.6 points using the exact same initialization and without any extra human labels.
  • Collision Reduction: The framework effectively taught the drone to fly higher and more cautiously, reducing the collision rate by approximately 50%.
  • Failure Shift: Over successive rounds, the primary cause of failure shifted from "Collisions" to "Timeouts," suggesting the model became much safer but occasionally too slow to reach the target.
Failure types across rounds: Collisions drop significantly while Timeouts slightly increase.

Limitations and Future Work

  • Simulation Cost: While AeroRSI eliminates human labeling, it is computationally expensive in simulation. Each round requires thousands of "test flights" to verify repairs.
  • Sim-to-Real Gap: The method relies on the ability to "save and restore" states in a simulator to test different actions from the exact same moment. This cannot be done directly in the real world, though the resulting policy can be deployed on real hardware.
  • Language Grounding: While navigation improves, the paper doesn't deeply explore if the model's actual understanding of complex linguistic nuances improves, or if it's primarily learning better spatial "common sense."

Summary updated

Abstract

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.
BibTeX
@misc{wu2026aerorsi,
  author = {Shiwei Wu and Yuwen Pan and Zhi Li and Yaowen Li and Yan Zhou and Xuemiao Xu and Yu Liu},
  title = {{AeroRSI: Recursive Self-Improvement for Aerial Vision-Language Navigation}},
  year = {2026},
  eprint = {2611.24077},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24077},
}
Figure