betaFeedback

WorksComputer Science

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

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.

Project code: https://github.com/starsholic/AeroRSI

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.

Introduction

Unmanned aerial vehicles (UAVs) that follow natural-language instructions could inspect infrastructure, search for people, and deliver goods without a human pilot. Toward this goal, recent vision-language-action (VLA) models for aerial vision-language navigation (VLN) [2, 3, 4, 5, 6, 7] directly map visual observations and instructions to flight actions [1, 8, 9, 10, 11], mirroring their success in robotic manipulation [12, 13, 14, 15, 16]. Despite this progress, existing aerial agents [1, 8, 9, 17, 18] still fail frequently in closed-loop flight, often due to collisions with obstacles. Since a single wrong action can terminate an entire flight, such failures severely limit their reliability.

These failures stem from a well-known mismatch in imitation learning [19, 20, 21], as a policy trained on demonstrations reaches states that the demonstrations never cover and accumulates errors along the flight. Correcting them requires repairs, namely actions that avoid a failure at the states the policy itself reaches, and existing approaches differ mainly in who judges such an action. Interactive imitation learning [21, 22, 23, 24] queries an expert for corrective actions, but human annotation is extremely expensive at scale, and rule-based planners [25, 26] ignore the language instruction. Automatic correction methods [1, 27, 28] avoid human labeling by deriving corrections from collision rollbacks, reference paths, or auxiliary teacher policies, yet they do not consider whether the policy can actually execute these corrections. Reinforcement learning [18, 29, 30, 31] instead relies on episode-level rewards, which indicate whether a flight fails but not which action should have been taken. As a result, annotation-free, action-level supervision that the policy itself can execute remains missing.

Recursive self-improvement offers a promising route toward such supervision. It has recently achieved remarkable success in language models [32, 33, 34, 35, 36, 37], where a model learns from its own verified outputs, and has begun to extend to robotics and navigation [28, 38, 39, 40, 41, 42, 43]. However, closed-loop flight poses three challenges that answer-level self-improvement leaves open. First, an answer can be verified on its own, whereas the value of a flight action depends on the subsequent flight of the policy, which calls for step-level values as in process supervision [44, 45]. Second, a language model learns from a fixed set of problems, whereas a flight terminates at its first failure, so later failures remain hidden until earlier ones are repaired. Third, self-improvement in language models mainly reinforces answers that the model can already produce [46, 47], whereas a failing flight policy often cannot sample the action it needs, even though simple physical priors, such as climbing or detouring around obstacles, can readily propose one, as rationalization does in STaR [32].

These challenges lead to our key insight that a failing policy should judge its own repairs. By continuing its flight after each candidate action, the current policy estimates the value of the candidate under its own behavior, which is what approximate policy iteration requires [48, 49, 50]. This self-verification turns weak sources into reliable supervision. Actions sampled from the policy often contain a repair, and rule-based proposals that degrade the policy under direct imitation become its most effective supervision once verified, whereas checking only the next few actions often favors candidates that fail later (Figure 1(a)). It also reveals how failures evolve. Verified repairs remain valid for subsequent policies, while new failures emerge elsewhere on the same routes (Figure 1(b)), so supervision must be renewed in every round. Moreover, actions learned from rule-based proposals in one round appear among its sampled actions in the next, indicating that the loop expands what the policy can do rather than reinforcing existing behavior.

Recursive self-improvement with self-verified repairs

Figure 1. Recursive self-improvement with self-verified repairs. (a) In round kk, πk\pi_k 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\pi_k, and only those it completes are kept. (b) After learning these repairs, πk+1\pi_{k+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.

Based on these insights, we present AERORSI, a recursive self-improvement framework that repeats three steps in every round. First, adaptive failure anchoring (AFA) collects the states leading to each failure of the current policy at multiple time scales, so that supervision follows the failures as they are revealed. Starting from these states, continuation-verified repair (CVR) builds candidates from actions sampled from the policy and from rule-based proposals, and accepts a candidate only if the original action fails again from the same state or falls clearly behind the best candidate, and the policy completes the candidate again in an independent re-flight. Finally, retentive repair learning (RRL) distills these verified repairs into the policy while keeping its behavior on states that need no repair.

Extensive experiments on the TravelUAV benchmark [3] demonstrate the effectiveness of AERORSI. Starting from a supervised policy, AERORSI improves the success rate from 46.4% to 67.5% in three rounds and reduces collisions by half (Figure 1(c)), outperforming the previous state-of-the-art method [1], which is trained from the same policy, by 10.6 points. When built upon the stronger policy of this method, AERORSI further improves the success rate to 69.9%, establishing a new state of the art without additional human annotation or inference cost.

In summary, our contributions are as follows.

  • We identify three challenges of self-improvement in closed-loop flight and show that letting a failing policy judge its own repairs turns its sampled actions and even rule-based proposals into reliable supervision.

  • We reveal that verified repairs remain valid while new failures emerge elsewhere on the routes, and that actions learned from rule-based proposals enter its sampled actions, which together make recursion essential.

  • We propose AERORSI, which anchors the failures of the current policy, verifies candidate repairs by its own continuation, and distills them while keeping its behavior elsewhere, achieving state-of-the-art results on TravelUAV with no extra annotation or inference cost.

Related Work

We review three lines of work most related to ours, namely aerial vision-language navigation, learning from corrections, and recursive self-improvement.

Aerial Vision-Language Navigation. Vision-language navigation was first studied for ground agents [51, 52, 53, 54, 55, 56, 57] and later extended to UAVs by AerialVLN [2], TravelUAV [3], OpenFly [4], and subsequent benchmarks [5, 6, 7]. Recent methods improve cross-embodiment pretraining [8], long-horizon context modeling [17], robustness to error accumulation [58], future prediction [59, 60, 61, 62, 63], and spatial reasoning [11, 64, 65]. AerialVLA [9] and AeroDPO [1] show that lightweight vision-language-action policies can match much larger models, and post-training further improves them through preference optimization on synthesized corrections [1] or online reinforcement learning [18, 29, 30, 31]. These efforts mostly improve the model architecture or learn from synthesized corrections and episode-level rewards, whereas AERORSI learns from repairs that the policy verifies itself, collected anew from its failures in every round, and is complementary to these architectures.

Learning from Corrections. Interactive imitation learning asks an expert to label the states that the learner itself encounters, as in DAgger [21] and its variants [22, 23, 24], while AggreVaTe [66] ranks actions by the cost-to-go of the expert. Navigation policies have also been learned from expert corrections [67] from crashes and disengagements [68, 69, 70], and in simulation [71, 72, 73]. To avoid human labeling, recent methods synthesize corrections by collision rollbacks [1], reference paths [27], or auxiliary residual policies [28], and learn from them with supervised, preference, or reinforcement objectives [74, 75, 76, 77], although preference updates may shift probability to unintended actions [78, 79, 80]. Since these corrections are judged by an expert, a heuristic, a reference path, or another policy, none of them accounts for whether the policy itself can execute them. CVR instead judges each candidate by how the current policy, not an expert, continues the flight after it, and RRL imitates the verified best candidate rather than contrasting it with the failed action.

Recursive Self-Improvement. Training a model on its own verified outputs underlies expert iteration [81, 82, 83], self-training [84], STaR [32], and ReST [33, 34], and repeated rounds of it have substantially improved the reasoning of language models [35, 36, 37, 85], although recent analyses find that it mainly reinforces outputs that the model can already produce [46, 47]. Weak supervisors can still teach stronger models [86], and autocurricula create new challenges as learning proceeds [87]. In robotics and navigation, agents likewise learn from data that they generate themselves [88, 89, 90], obtain corrections from residual policies [28], build on coding or multimodal agents [40, 41, 43], direct data collection with learned models [42], or retrain on their own trajectories and reasoning [38, 91, 92], with general formulations in [93, 94]. Closest to ours, CorrectNav [27] repeatedly retrains a policy on corrections of its errors toward reference paths, and local recovery supervision [39] checks whether a frozen student can continue from a teacher’s recoveries. Rollout-based policy iteration [49, 50] and learning to search [95, 96, 97] also judge actions by a continuation, but roll out every action of a small set or follow a reference policy. AERORSI brings this judgment to the failures of a VLA policy with nearly a million actions, verifying candidates from its samples and physical priors by Monte-Carlo continuations from restored states [98, 99, 100], without reference paths or a teacher, and RRL retains its behavior elsewhere, as in continual learning [101, 102, 103]. Our ablations compare these choices with such combinations (§ 4.3): unverified rule-based corrections help little or even hurt, checking only the first actions leaves more collisions, and reusing one round of supervision falls short of anchoring anew in every round.

Method

In this section, we first formulate recursive self-improvement for aerial navigation as approximate policy iteration and then present the three modules of AERORSI that implement it in every round.

Problem Formulation

Task. At each step tt, the UAV receives an observation oto_t consisting of its front and downward RGB views and the instruction, together with a coarse direction to the target as in the fuzzy direction guidance of [1]. The policy π\pi outputs an action ata_t that discretizes the forward, vertical, and yaw motion, optionally with a landing action. An episode succeeds if the UAV lands near the described target, and it terminates at the first failure, namely a collision, getting stuck, moving away from the target, landing prematurely, or exceeding the step limit. During training, the simulator can save and restore the full state sts_t of the UAV and its environment and provides depth maps and the target position. AERORSI uses these privileges only to generate training data, and the deployed policy observes oto_t alone, with the architecture and inference cost of the initial policy π0\pi_0.

Recursive Self-Improvement. Starting from π0\pi_0, each round kk generates supervision Dk\mathcal{D}_k from the experience of the current policy πk\pi_k without human annotation and uses Dk\mathcal{D}_k to update it to πk+1\pi_{k+1}. From the perspective of approximate policy iteration [48], an ideal round would relabel every state ss at which πk\pi_k fails with the improved action

a∗(s)=arg max⁡a  Qπk(s,a),(1)a^*(s) = \underset{a}{\operatorname{arg\,max}}\; Q^{\pi_k}(s,a), \tag*{(1)}

where Qπk(s,a)Q^{\pi_k}(s,a) is the expected return of executing aa in ss and following πk\pi_k afterward. Because QπkQ^{\pi_k} is defined by the subsequent flight of πk\pi_k itself, it is the failing policy that should judge its own repairs, as in rollout-based policy iteration [49, 50] and learning to search [95, 96]. Since the judge is the policy itself, each round also improves the next judge, which makes the loop recursive. Two properties of closed-loop flight make this target hard to obtain. The outcome of a continuation is stochastic, so a single rollout is an unreliable estimate of QπkQ^{\pi_k}. Moreover, as an episode terminates at its first failure, the states that need repair are revealed only as πk\pi_k improves. AERORSI therefore estimates QπkQ^{\pi_k} from verified continuations of the current policy and relabels the newly revealed failures in every round.

Overview

Figure 2 illustrates one round of AERORSI. The current policy πk\pi_k first flies all training routes, and AFA selects anchor states leading to its failures. CVR then turns each anchor into a verified repair or discards it, and RRL distills the repairs into πk+1\pi_{k+1} while keeping the behavior of πk\pi_k on states that need no repair, and the next round starts from the failures that πk+1\pi_{k+1} reveals.

Figure 2. Overview of AERORSI. The top row shows the policy πk\pi_k, 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\pi_k; (2) CVR restores each anchor, lets πk\pi_k continue the executed action a−a^{-}, its sampled actions, and rule-based proposals, and accepts as the repair c∗c^* the first candidate that qualifies again in a re-flight; and (3) RRL distills the verified repairs into a student copy of πk\pi_k, keeping the outputs of πk\pi_k on states that need no repair, to obtain πk+1\pi_{k+1}.

Adaptive Failure Anchoring (AFA)

Hand-designed anchors, such as the last action before a collision, cover only the failure types and time scales that a designer anticipates, whereas the failures of the policy change as it improves. AFA therefore anchors every failure revealed by the current policy at multiple time scales, regardless of its type.

For each training episode of πk\pi_k that ends in a failure, AFA places probes at exponentially spaced offsets before the end of the episode. Near offsets capture immediate causes, such as an action into an obstacle, and far offsets capture earlier decisions, such as a turn into a dead end. Each probe records the saved state ss, the observation oo, and the action a−a^{-} that πk\pi_k executed. Since the probes far outnumber the anchors that can be verified, AFA draws a fixed budget of anchors uniformly from all probes of the round. The anchors therefore follow the failure distribution of πk\pi_k without hand-crafted rules, so that frequent failure types and time scales receive more anchors, and failures that the policy has overcome receive none. Each anchor is finally assigned an evaluation horizon that grows with its offset and extends beyond the original failure point, so that a qualifying candidate must carry the flight past this point.

Continuation-Verified Repair (CVR)

Candidate Group. For an anchor with state ss, observation oo, and original action a−a^{-}, CVR builds the candidate group

C(s)={a−}∪{a(i)∼πk(⋅∣o)}i=1M∪P(s),(2)\mathcal{C}(s)=\{a^{-}\}\cup\{a^{(i)}\sim\pi_k(\cdot\mid o)\}_{i=1}^{M}\cup\mathcal{P}(s), \tag*{(2)}

which contains the original action, MM actions sampled from the current policy, and a set P(s)\mathcal{P}(s) of rule-based proposals, namely a potential-field avoidance action [25] computed from the depth map and the target position, a variant of it

that keeps the original forward motion, and a climb that keeps the original heading. The sampled actions are ones that the policy can already produce but does not choose, and the rule-based proposals add avoidance actions that it rarely samples. No candidate is trusted as a label before it passes verification.

Continuation Score. To estimate Qπk(s,c)Q^{\pi_k}(s,c) for a candidate cc, CVR restores ss, executes cc, and lets the current policy continue greedily until the evaluation horizon of the anchor or the end of the episode. The candidate receives the score

Sk(s,c)=Δd−λfI[fail]+λsI[success],(3)S_k(s,c)=\Delta d-\lambda_f\mathbb{I}[\mathrm{fail}]+\lambda_s\mathbb{I}[\mathrm{success}], \tag*{(3)}

where Δd\Delta d is the progress toward the target, a failure denotes a collision, a premature landing, or an invalid output, and λf\lambda_f and λs\lambda_s weight failure and success. Unlike a short-term criterion that inspects only the immediate outcome of cc, SkS_k judges a candidate by the subsequent flight of the policy itself and is thus a finite-horizon estimate of QπkQ^{\pi_k}.

Verified Acceptance. Since a single continuation is stochastic, CVR accepts a repair only after two checks against chance outcomes. A candidate qualifies if it reaches the end of the horizon without failure and with a minimum progress, or if it lands successfully, so that getting stuck or flying away counts as too little progress. The reproduction check first replays the original action from ss and keeps the anchor only if this action fails within the horizon or the best qualifying candidate outscores it by a margin. The confirmation check then re-flies the two best qualifying candidates from ss in order and accepts the first one, c∗c^*, that qualifies again. Each accepted anchor yields a repair (o,c⋆)(o, c^\star), and the next round verifies the new failures of πk+1\pi_{k+1}.

Retentive Repair Learning (RRL)

A policy trained only on repairs tends to drift on states that need no repair, because it transfers the action statistics of the repairs, such as frequent turns, to the rest of the flight. RRL therefore updates the policy with three components that together limit this drift.

Repair Supervision. RRL maximizes the likelihood of the verified repair c⋆c^\star of each anchor and, unlike preference optimization [75], adds no penalty on the original action. The original action is often only moderately worse than the repair and remains correct in many other states, so pushing it down explicitly risks degrading the flight elsewhere.

Behavior Retention. RRL adds a set N\mathcal{N} of states that need no repair, taken from successful training flights and covering early flight, the approach to the target, and the landing. Their targets a^k(o)\hat{a}_k(o) are the greedy outputs of the current policy πk\pi_k rather than the demonstrations or an earlier policy, so that retention anchors the policy to its own behavior where no repair is needed without pulling it back to an outdated one.

Repair Replay. From the second round on, RRL replays the repairs of the previous round with a share ρ\rho of the repair weight as a safeguard against forgetting them.

Together, the update of round kk minimizes

Lk=(1−λn)EDk[CE⁡(c⋆)]+λnEN[CE⁡(a^k)],(4)\mathcal{L}_k = (1-\lambda_n)\mathbb{E}_{D_k}\left[\operatorname{CE}(c^\star)\right] + \lambda_n\mathbb{E}_{\mathcal{N}}\left[\operatorname{CE}(\hat{a}_k)\right], \tag*{(4)}

where CE is the cross-entropy of the action tokens given oo, DkD_k draws a verified repair of round kk with probability 1−ρ1-\rho and a repair of the previous round with probability ρ\rho, and λn\lambda_n is the share of retention. The policy is initialized from πk\pi_k, and the trained policy becomes πk+1\pi_{k+1}.

Experiments

In this section, we first describe the experimental setup and compare AERORSI with existing methods, then ablate its design, and finally present case studies of recorded flights.

Experimental Setup

Benchmark and Metrics. We evaluate AERORSI on TravelUAV [3], a benchmark for aerial VLN built on AirSim and Unreal Engine [105]. Its training set contains 7,611 routes in 11 scenes, and its test set contains 1,418 routes in seen environments (Seen), 629 routes with unseen objects (Unseen Object), and 958 routes in unseen environments (Unseen Map). AERORSI collects experience only on the training routes. An episode succeeds if the UAV lands within 20 m of the target within 200 actions. Following [1], we report the navigation error (NE, in meters), success rate (SR), oracle success rate (OSR), and success weighted by path length (SPL), together with the collision rate (CR). All design choices were made on training routes held out in preliminary experiments, and every round of AERORSI is evaluated on all test routes. Since closed-loop flight is stochastic, evaluating the same policy twice changes the outcome of about a quarter of the routes and its overall SR by up to about two points, so we compare policies route by route and treat smaller differences as noise. The simulator pauses at observation boundaries during data collection so that states can be saved and restored exactly, whereas all evaluations follow the native protocol without pauses.

Baselines. We compare with two groups of methods. TravelUAV [3], NavFM [8], LongFly [17], and NeuroKalman [58] use the L1 assistant guidance of TravelUAV, whereas AerialVLA [9], RefineFly [18], and AeroDPO [1] use fuzzy direction guidance, as AERORSI does. AeroDPO is the closest baseline. It first trains a 2B-parameter policy on the expert demonstrations of AerialVLA by supervised fine-tuning, which we denote SFT, and then post-trains it by preference optimization on corrections derived from collision rollbacks. For a controlled comparison, AERORSI (SFT init.) starts from the same SFT policy, and we evaluate the released SFT and AeroDPO checkpoints under the same protocol as AERORSI, comparing them route by route.

Implementation Details. Table 1 lists the compared policies. To test whether AERORSI also improves a policy that has already been post-trained, we further start it from AeroDPO, denoted AERORSI (DPO init.). In both cases, π0\pi_0 is the initial policy, and the states that need no repair contribute only their observations, labeled with the outputs of the current policy. The supplementary material gives all settings and the cost of each round.

PolicyInitialized fromTrained on
Supervised policy (SFT) [1]Qwen3-VL-2B [104]Expert demonstrations
AeroDPO [1]SFTCollision corrections
AERORSI (SFT init.)SFTVerified repairs
AERORSI (DPO init.)AeroDPOVerified 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.

Main Results

Comparison with Existing Methods. Table 2 compares AERORSI after three rounds with existing methods. AERORSI (SFT init.) and AeroDPO post-train the same SFT policy without additional human labels, so their difference reflects the post-training alone. AERORSI raises the overall SR 10.6 points above AeroDPO, and the gain holds route by route, as AERORSI succeeds on 649 test routes on which AeroDPO fails while failing on only 330 on which AeroDPO succeeds. The gain concentrates where the loop collects experience. On Seen and Unseen Object,

MethodGuideSeen NE↓Seen SR↑Seen OSR↑Seen SPL↑Unseen Object NE↓Unseen Object SR↑Unseen Object OSR↑Unseen Object SPL↑Unseen Map NE↓Unseen Map SR↑Unseen Map OSR↑Unseen Map SPL↑Overall SR↑
TravelUAV [3]L1106.2816.1044.2614.30118.1122.4246.9020.51138.804.1820.773.8413.62
NavFoM [8]L193.0529.1749.2425.03108.0429.8347.9927.20125.106.3018.955.6822.02
LongFly [17]L160.0236.3965.8731.0766.7443.8764.5638.39108.3211.2730.279.3229.95
NeuroKalman [58]L171.5625.8658.7322.4371.0132.4860.8228.50100.328.3434.157.1221.66
AerialVLA [9]FD65.8847.9657.6938.5461.4556.6064.8646.6167.4237.5852.9228.2246.46
RefineFly [18]FD54.9656.3365.4045.9853.4259.7268.3151.3458.8842.9760.0931.8352.78
AeroDPO [1]FD51.8460.0167.7747.3047.6866.4572.0255.3068.7945.9355.6435.4556.87
AERORSI (SFT init.)FD31.9775.3280.4757.3829.2682.1984.2666.9965.5846.2454.7031.0667.49
AERORSI (DPO init.)FD31.0076.5283.5760.1526.2579.6584.7464.8858.8653.5564.5136.2569.85

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.

AERORSI outperforms all prior methods on every metric, raising the SR of AeroDPO by 15.3 and 15.7 points and reducing its NE by nearly 40%, whereas on Unseen Map, whose environments the loop never visits, it raises the SR of SFT by 9.4 points but only matches AeroDPO in SR, with a lower SPL. SPL rises less than SR, as many repairs climb over or detour around obstacles and lengthen the path. AERORSI (DPO init.) improves every metric on all three splits and raises the overall SR of AeroDPO by another 13.0 points to 69.9%, so repairs verified by the policy remain effective after preference optimization and complement it. On Unseen Map, the loop adds a similar gain to either initial policy, 9.4 and 7.6 points, so the final level in unseen environments is set mainly by the initial policy.

Improvement across Rounds. Table 3 reports every round. For AERORSI (SFT init.), the test SR rises from 46.4% to 57.6%, 66.0%, and 67.5%, and the SR on the training routes follows the same trend, so the repairs generalize beyond the anchored states. The gain per round decreases from 11.2 to 8.4 and 1.5 points, the last within the noise, as each round removes the failures that its verifier can repair. The failures that remain also change in kind (Figure 3). Collisions and flights away from the target decrease in every round, whereas timeouts grow from 1.2% to 6.8% of the routes,

RoundSeenUOUMSR↑SPL↑CR↓Train SR↑
AERORSI (SFT init.)
048.3856.4436.8546.3937.3738.6445.24
162.6267.8943.4257.6046.9629.4861.41
273.9179.0145.8266.0251.9621.2671.41
375.3282.1946.2467.4951.0019.1070.85
AERORSI (DPO init.)
060.0166.4545.9356.8745.2028.9556.21
169.9673.9346.7663.3949.8422.4666.03
272.5078.5448.8566.2250.0719.0770.84
376.5279.6553.5569.8553.5217.9072.78

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.

so the bottleneck shifts from avoiding obstacles to reaching the target in time. AERORSI (DPO init.) follows the same trend, and the gap between the two initializations shrinks from 10.5 to 2.4 points, almost all of it on Unseen Map, so in the environments that the loop experiences, the final result depends little on which policy it starts from.

Ablation Study

Table 4 and Figure 4 examine the insights behind AERORSI first and the designs of its modules afterwards.

(a) Verification in the first round(b) Recursion
SourceVerifiedSR↑SPL↑CR↓Variant
Potential field–39.7531.9225.01Round 1
Rule-based proposals–49.2337.5824.08Stale labels
Rule-based proposals✓61.5948.8824.922× anchors, once
Sampled actions–50.6241.0134.93w/o replay
Sampled actions✓51.6542.9536.33Round 2 (ours)
Both2 actions55.7647.0033.72
Both (ours)✓57.6046.9629.48
(c) Repair learning
VariantSR↑SPL↑CR↓
w/o retention57.4647.2832.01
+ DPO loss58.0447.2630.67
States from π0\pi_058.9748.3930.00
Advantage weights53.9343.9833.87
Round 1 (ours)57.6046.9629.48

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\pi_0 failed, each continued by a later policy (π0\pi_0 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\pi_3 over π0\pi_0 on 999 test routes with and without an altitude cap. Error bars are 95% confidence intervals, paired in (f).

Verification by the Policy. Verification by the policy is what turns candidate actions into useful supervision (Table 4(a)). Labeling every anchor with the potential-field action, as DAgger [21] would with a rule-based expert, lowers the SR of the initial policy by 6.6 points, and a random rule-based proposal raises it by only 2.8 points, as the policy trades collisions for getting stuck or flying away from the target (Figure 4(a)). Once the policy must complete the same proposals, the SR rises by 12.4 points to 61.6%. For its own sampled actions, which stay close to its behavior, verification adds only 1.0 point, so the benefit of verification grows with how far a source departs from the policy. This benefit comes from flying the candidates on rather than from the checks of Sec. 3.4, without which the SR is a similar 58.4%. Verification must also follow the flight beyond the first actions, as scoring the candidates after two actions instead leaves 4.2 points more collisions, and the candidate nearest to the target after two actions later fails or stalls at nearly a quarter of the anchors with a repair (Figure 5(a)). Repairs mostly hold beyond the horizon, as π0\pi_0, flying on from those of the first round until it lands, collides, or runs out of actions, reaches the target in 57% of episodes it had failed.

Case study of AERORSI

Figure 5. Case study of AERORSI. From a state before π0\pi_0 crashes, π0\pi_0 flies on from each candidate action (top), and π1\pi_1 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\pi_1 learns this repair. (c) The same action crashes when π0\pi_0 flies on but not when π1\pi_1 does, so every round must judge its repairs anew.

Sources of Candidates. The two sources of candidates complement each other, and their roles shift as the policy learns (Figure 4(b,Figure 4c)). The failing policy often holds a repair that it does not choose, as sampled actions supply 44% of the repairs accepted in the first round, and 78% of them still qualify when π0\pi_0 flies them again, against 38% for new samples that nearly repeat the failed action, so they are not lucky continuations. At the 21% of the anchors whose failure reproduces that only a rule-based proposal repaired, however, a sampled action of π0\pi_0 qualifies at only 11%, whereas after one round of learning from both sources, a sampled action of π1\pi_1 qualifies at 62%, as often as at the other anchors, and at 58% of the 42 such anchors in route groups held out from training, against 14% for π0\pi_0. The repairs learned from the proposals thus enter the policy’s sampled actions, and the leading source changes across rounds (Figure 4(c)). The proposals alone lead in the first round, as nearly all of their repairs change a single component of the action and thus teach a consistent correction, and both sources together lead in the second. No single source suffices in every round, so CVR verifies both and keeps the best candidate.

Necessity of Recursion. A new round is needed because the failures move, not because the repairs expire (Table 4(b)). Retraining π1\pi_1 on the repairs of the first round, without anchoring or verifying anew, reaches 60.3%, and collecting twice as many anchors from π0\pi_0 and training once, with as many updates as two rounds, reaches 63.0%, both below the 66.0% of the second round and with more collisions, whereas the second round without replay still reaches 65.7%. Yet the repairs remain valid, as those verified in the first round still qualify at 87% to 89% of their anchors when later policies continue them, compared with 81% when π0\pi_0 continues them again (Figure 4(d)). What changes is where the policy fails. The action at which π0\pi_0 failed qualifies at 57% to 67% of these anchors under later policies, against 27% under π0\pi_0, and on the training routes that two consecutive policies both fail, their failure points lie a median of 40 m apart after the first round and 28 m after the second, compared with 13 m when the same policy flies twice (Figure 4(e)).

Retentive Repair Learning. Behavior retention keeps the repairs from changing normal flight (Table 4(c)). Training on the repairs alone raises the CR from 29.5% to 32.0%, as the policy turns in 19.6% instead of 15.8% of its actions, drifting toward the 30% of the repairs. Taking the retained states from flights of π0\pi_0 instead of AeroDPO gives 59.0%, so the retention does not rely on another method. Weighting every qualifying candidate by its group-normalized advantage, as in GRPO [77], is 3.7 points below imitating the best candidate alone, suggesting that one consistent repair per state teaches more than several competing ones. Adding the DPO loss against the original action gives a similar 58.0%, while the DPO loss alone collapses the policy, so the gain comes from the verified repairs rather than the objective.

Flight Altitude. Many verified repairs climb over obstacles, and the median peak altitude of a test flight rises from 10 m for π0\pi_0 to 36 m for π3\pi_3, yet with the altitude capped at 15 m above the start on 999 test routes, π3\pi_3 keeps about four fifths of its gain over π0\pi_0, 13.2 of 16.2 points (Figure 4(f)).

Case Study

Missed Repairs. Figure 5 shows three anchors of the first round, with the front camera in the insets. Panels (a) and (b) reveal two ways in which a policy misses a repair that exists. In (a), the policy samples the repair but does not choose it. After two actions, a sharp turn is about 6 m behind the actions flying toward the target, yet those actions crash into a palm tree and only the turn completes the flight. A repair can thus cost progress before it pays off, so a criterion on the first actions would discard exactly the behavior that the policy lacks. In (b), the policy does not even sample the repair, as its sampled actions nearly repeat the action that crashes. Only one of the three rule-based proposals completes the flight, so rules can supply a repair, but only the flight of the policy can identify it. In both cases, π1\pi_1 takes the repair by itself after one round, as do most of its new sampled actions, so each round shifts what the next one can sample.

One Action under Two Policies. Figure 5(c) shows that the value of an action is not fixed. The action at which π0\pi_0 crashes completes the flight when π1\pi_1 flies on from it, as it does at most anchors accepted in the first round (Figure 4(d)). Whether an action repairs a failure thus depends on the policy that continues the flight, so a label judged under an earlier policy or an expert may not hold for the current one. This is why AERORSI lets the current policy judge its own repairs and verifies them anew in every round.

Conclusion

In this paper, we have presented AERORSI, a recursive self-improvement framework for aerial VLN in which a failing policy judges its own repairs by flying on from them. Verified in this way, the sampled actions of the policy and simple rule-based proposals become reliable supervision, and since repairs remain valid while failures move, every round anchors and verifies anew. On TravelUAV, three rounds raise the SR of a 2B-parameter policy from 46.4% to 67.5% and reduce collisions by half without additional human labels or changes at deployment.

This supplementary material is organized as follows.

  • Sec. A gives the implementation details of AERORSI: its settings, the rule-based proposals, the verification of an anchor (Algorithm S1), the ablation variants, and the evaluation and cost.

  • Sec. B gives the failure types of both initializations and the paired significance tests behind the comparisons of the main paper.

  • Sec. C gives further analyses behind Figure 4 of the main paper, in the order of its ablation study.

Implementation Details

Settings

Table S1 lists the settings of AERORSI, which are shared by both initializations and all three rounds. A fifth of the route groups of the training split is held out from training: the policies still fly these routes and AFA anchors their failures, but their repairs only serve to test whether repairs transfer to route groups that the policy has not learned from (Secs. C.2 and C.3). The retention states of RRL come from successful flights of AeroDPO on the training routes and are labeled with the greedy outputs of the current policy; taking them from the flights of π0\pi_0 instead gives a similar SR (Table 4(c) of the main paper).

ComponentSetting
PolicyQwen3-VL-2B; 384×768384 \times768 mosaic of the front and downward views; forward, vertical, and yaw components with 99 bins each, plus landing
TrainingLoRA of rank 64 on all linear layers of the language model, and the visual merger; 2 epochs per round; learning rate 2×10−52 \times10^{-5}; batch size 16
AFAOffsets of 1, 2, 4, 8, and 16 actions; 2,400 anchors per round; horizon of max⁡(8,δ+4)\max(8,\delta+4) actions for offset δ\delta, at most 20
CVRM=4M=4 samples at temperature 1.0 and three rule-based proposals; λf=100\lambda_f=100, λs=20\lambda_s=20; minimum progress of 4 m; margin m=4m=4 m; two confirmations at most
RRLλn=0.2\lambda_n=0.2 on 1,990 retention states; replay share ρ=30%\rho=30\% from round 2

Table S1. Settings of AERORSI. They are shared by both initializations and all rounds.

Rule-based Proposals

CVR adds three rule-based proposals to the sampled actions, each computed from the anchor state alone. The potential field attracts the UAV to the target and repels it from the obstacles in the depth map observed at the anchor. Its

Algorithm S1 Verification of one anchor by CVR

Figure S1.

translation variant keeps the forward command of the original action, takes the vertical command of the potential field, and limits the yaw to 0.224 rad, below the 0.25 rad at which an action turns in place instead of translating. The climb keeps the forward and yaw commands of the original action and rises 2 m more, by at most 5 m in one action. The proposals are deliberately simple and become useful only after verification (Table 4(a) of the main paper).

Verification

Algorithm S1 details how CVR verifies one anchor. A flight fails if it collides, lands prematurely, or produces an invalid output, and a candidate qualifies if its flight reaches the end of the horizon without failure and with at least 4 m of progress, or if it lands successfully. The input check discards an anchor at which the restored observation no longer leads the policy to its original action. The reproduction check keeps an anchor only if the original action is indeed worse than a candidate, and the confirmation check accepts a candidate only if it qualifies in two independent flights.

Variants of the Ablation Study

Three variants in Table 4 of the main paper are set up as follows. Without the two checks, CVR accepts the best qualifying candidate of every anchor after one flight, giving 1,374 repairs instead of 837 with the same 354 updates. Without replay, round 2 trains π1\pi_1 on its new repairs and the retention states with the same 444 updates. The DPO variant pairs the 837 repairs with the original action as the rejected one, with π0\pi_0 as the reference, β=0.1\beta=0.1, and the learning rate and 354 updates of round 1, without retention states; with the DPO loss alone, the outputs degenerate and the SR falls to 0.1%, whereas adding the likelihood of the repair with weight 1 gives 58.04%.

Evaluation and Cost

All policies are evaluated on the 3,005 test routes of TravelUAV with the native actions and termination of the benchmark, and an episode interrupted by a simulator error is flown again from its start. A round of AERORSI costs about 100 GPU hours (Table S2): flying the 7,611 training routes takes 17 to 20 hours and verifying the anchors about 6 hours on four GPUs, and training takes 11 to 23 minutes on one GPU.

Table S2. Simulator cost of each round of AERORSI (SFT init.). Round kk flies the training routes with πk−1\pi_{k-1}.

Round123
Flights of the training routes
Flights7,6117,6117,611
Actions (thousands)345376435
Verification
Flights of candidates15,23713,97512,451
Confirmation flights1,4841,3421,165
Actions (thousands)141125116

Table S2. Simulator cost of each round of AERORSI (SFT init.). Round kk flies the training routes with πk−1\pi_{k-1}.

Additional Results

Failure Types across Rounds

Table S3 extends Figure 3 of the main paper to both initializations. From either initial policy, collisions and flights away from the target decrease over the rounds, so the gain comes mainly from flights that no longer crash or lose the target.

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).

RoundCollisionWrong wayStuckEarly landingTimeout
AERORSI (SFT init.)
038.67.84.21.81.2
129.53.84.22.22.8
221.32.62.82.35.0
319.12.12.71.96.8
AERORSI (DPO init.)
029.05.54.21.33.2
122.53.53.22.64.9
219.13.43.62.65.1
317.93.12.92.93.3

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).

Paired Significance Tests

Table S4 tests the comparisons of the main paper route by route. For two runs on the same test routes, with bb routes on which only the first succeeds and cc routes on which only

Table S4. Paired tests on the test routes. Difference in SR (points) and zz over the routes common to both runs, for AERORSI (SFT init.) unless stated.

ComparisonΔ\DeltaSRzz
Rounds and baselines
Round 1 vs. round 0+11.2110.4
Round 2 vs. round 1+8.428.1
Round 3 vs. round 2+1.461.6
Round 3 vs. round 0+21.1018.5
Round 3 vs. AeroDPO+10.6210.2
DPO init., round 3 vs. AeroDPO+12.9812.7
DPO init. vs. SFT init., round 3+2.362.5
Ablations
Potential field vs. round 0−6.64−6.2
Rule-based proposals, unverified vs. verified−12.37−11.5
Without the two checks vs. round 1+0.770.8
Rule-based proposals vs. both, round 2−2.13−2.2
Sampled actions vs. both, round 2−2.23−2.3
Stale labels vs. round 2−5.73−5.5
2× anchors, once vs. round 2−3.06−3.2
w/o replay vs. round 2−0.33−0.4
Advantage weights vs. round 1−3.70−3.7
+ DPO loss vs. w/o retention+0.570.6

Table S4. Paired tests on the test routes. Difference in SR (points) and zz over the routes common to both runs, for AERORSI (SFT init.) unless stated.

the second does, we report the difference in SR and z=(b−c)/b+cz = (b-c)/\sqrt{b+c} of the McNemar test, where ∣z∣>1.96|z| > 1.96 means p<0.05p < 0.05. For the two comparisons that the main paper makes by collisions, the CR rises by 4.23 points with the two-action score (z=4.2z=4.2) and by 2.53 points without retention (z=2.6z=2.6).

Additional Analysis

Statistics of the Loop

Table S5 summarizes the collection of each round. As the policy improves, fewer anchors yield a repair, whereas the best of four sampled actions still qualifies at about two thirds of the anchors whose failure reproduces, and a growing share of the repairs comes from the sampled actions. From the second round on, about half of these sampled repairs change a single component of the failed action, against 31% in the first, closer to the rule-based proposals, nearly all of whose repairs do. Repairs are found both just before and long before a failure: in the first round, a repair is accepted at 64% of the anchors one action before a collision and at 56% of those 16 actions before.

Table S5. Statistics of the loop of AERORSI (SFT init.) unless stated. Anchors: those evaluated of the 2,400 drawn.

Round123
Anchors2,3902,3902,399
Share of all anchors (%)
Repair accepted43.838.334.3
Repair accepted, DPO init.46.138.833.9
Share of the anchors whose failure reproduces (%)
Best of four samples qualifies66.268.368.3
Only a rule-based proposal repairs21.016.416.5
Share of the accepted repairs (%)
From sampled actions43.951.954.4
of which change one component30.750.547.9

Table S5.

Sources of Candidates

Table S6 gives how often the candidates of each type qualify at the anchors whose failure reproduces. The rule-based proposals qualify at about half of these anchors, so they are useful candidates but unreliable labels without verification, and the sampled actions qualify more often as the policy improves. The repairs taken from sampled actions are not lucky continuations: at the anchors accepted in the first round, 77.6% of them qualify again when π0\pi_0 flies them once more (Table S8), whereas new samples of π0\pi_0 that nearly repeat the failed action, within two bins in every component, qualify at 37.8%.

Table S6. Candidates that qualify (%) at the anchors whose failure reproduces. Identical samples of an anchor count once.

RoundSampled actionsPotential fieldTranslationClimb
139.260.759.347.6
244.355.459.646.2
351.063.659.051.2

Table S6.

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\pi_0 failed, qualifies.

Continued byπ0\pi_0 againπ1\pi_1π2\pi_2π3\pi_3
Verified repair81.586.987.089.1
held out (211)82.588.291.992.9
from sampled actions (460)77.683.385.488.0
from proposals (588)84.589.888.290.0
Failed action26.557.063.366.6

Table S8.

Table S7 extends Figure 4(b) of the main paper to all groups of the 1,450 anchors of the first round whose failure reproduces. The groups are defined by the samples drawn during collection and the rates by four new samples of each policy, so the two do not share the same draws. The anchors that only a rule-based proposal repaired gain the most, also in the route groups held out from training, and at these anchors the qualifying samples of later policies come to resemble the proposals: the share that climbs by at least 10 bins grows from 12% for π0\pi_0 to 42% for π1\pi_1 and 57% for π3\pi_3.

Table S7. Pass@k of new samples (%) at the 1,450 reproducible anchors of the first round, grouped by what repaired them during collection.

Anchors repaired byπ0\pi_0 @1π0\pi_0 @4π1\pi_1 @1π1\pi_1 @4π3\pi_3 @1π3\pi_3 @4
All (100%)33.461.763.283.673.886.6
A sampled action (66%)45.679.971.790.980.492.2
Only a proposal (21%)11.030.462.184.376.787.9
Neither (13%)6.518.920.944.334.655.1
Held-out route groups
All41.874.968.791.879.092.7
Only a proposal (42)13.733.357.785.776.285.7

Table S7.

Repairs and Failures across Rounds

Table S8 flies the 1,048 repairs accepted in the first round again from their anchors, each continued by a later policy, as in Figure 4(d) of the main paper. The repairs remain valid under every later policy, from either source and in the route groups held out from training, whereas the action at which π0\pi_0 failed passes increasingly often. What changes instead is where the policy fails (Table S9): on the training routes that two consecutive policies both fail, the two failure points lie much farther apart than when one policy flies the same routes twice. Most repairs also hold to the end of the episode. From each of the 1,048 anchors accepted in the first round, π0\pi_0 executes the verified repair and flies on until it lands, collides, or reaches the limit of 200 actions, with the simulator paused at observation boundaries as in collection. It reaches the target in 57.2% of these episodes, 58.2% for repairs from sampled actions and 56.5% for those from rule-based proposals, and in 53.6% at the 604 anchors whose failure reproduces. The flights that fail again do so a median of 14 actions after the original failure point.

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.

FlightsRoutesWithin 10 mOver 50 mMedian
π0\pi_0, then π1\pi_12,12527.944.840.2 m
π1\pi_1, then π2\pi_21,33333.037.628.0 m
One policy twice5650.030.412.7 m

Table S9.

Flight Altitude

The median peak altitude of a test flight of AERORSI (SFT init.) rises from 10.0 m in round 0 to 15.6, 29.1, and 36.2 m in rounds 1 to 3. To test how much of the gain this explains, we evaluate π0\pi_0, π1\pi_1, and π3\pi_3 on 999 test routes drawn from the three splits with the altitude capped at 15 m above the start of each route, cutting any upward command to the room left below the cap. With the cap, their SR is 36.44%, 45.25%, and 49.65%, against 50.15%, 58.36%, and 66.37% without it on the same routes, so π3\pi_3 keeps 13.21 of its 16.22-point gain over π0\pi_0 (Figure 4(f) of the main paper).

References

  1. [1]Peng Xu, Chengcheng Wang, and Shaohua Wan. AeroDPO: Unleashing lightweight UAV navigation with high-fidelity perception and automated preference optimization. arXiv preprint arXiv:2608.07557, 2026. 1, 2, 3, 5, 6arxiv.org/abs/2608.07557
  2. [2]Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yanning Zhang, and Qi Wu. AerialVLN: Vision-and-language navigation for UAVs. In ICCV, 2023. 1, 2arxiv.org/abs/2308.06735
  3. [3]Xiangyu Wang, Donglin Yang, Ziqing Wang, Hinkson Kwan, Jinyu Chen, Wenjun Wu, Hongsheng Li, Yue Liao, and Si Liu. Towards realistic UAV vision-language navigation: Platform, benchmark, and methodology. In ICLR, 2025. 2, 5, 6arxiv.org/abs/2410.07087
  4. [4]Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. OpenFly: A comprehensive platform for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025. 2arxiv.org/abs/2502.18041
  5. [5]Jungdae Lee, Taiki Miyanishi, Shuhei Kurita, Koya Sakamoto, Daichi Azuma, Yutaka Matsuo, and Nakamasa Inoue. CityNav: A large-scale dataset for real-world aerial navigation. In ICCV, 2025. 2arxiv.org/abs/2406.14240
  6. [6]Xiangyu Wang, Donglin Yang, Yue Liao, Wenhao Zheng, Wenjun Wu, Bin Dai, Hongsheng Li, and Si Liu. UAV-Flow Colosseo: A real-world benchmark for flying-on-a-word UAV imitation learning. In NeurIPS, 2025.arxiv.org/abs/2505.15725
  7. [7]Jiajun Jiang, Chunliang Hua, Zichun Chen, Yanxing Wu, Zeyuan Yang, Jie Song, and Xiao Hu. SatNav: A scalable benchmark for long-horizon UAV vision-language navigation from satellite imagery. In NeurIPS, 2026. 1, 2arxiv.org/abs/2609.31507
  8. [8]Jiazhao Zhang, Anqi Li, Yunpeng Qi, Minghan Li, Jiaheng Liu, Shaoan Wang, Haoran Liu, Gengze Zhou, Yuze Wu, Xingxing Li, Yuxin Fan, Wenjun Li, Zhibo Chen, Fei Gao, Qi Wu, Zhizheng Zhang, and He Wang. Embodied navigation foundation model. In ICLR, 2026. 1, 2, 5, 6arxiv.org/abs/2509.12129
  9. [9]Peng Xu, Zhengnan Deng, Jiayan Deng, Zonghua Gu, and Shaohua Wan. AerialVLA: A vision-language-action model for UAV navigation via minimalist end-to-end control. arXiv preprint arXiv:2603.14363, 2026. 1, 2, 5, 6arxiv.org/abs/2603.14363
  10. [10]Xiaolou Sun, Wufei Si, Wenhui Ni, Yuntian Li, Dongming Wu, Fei Xie, Runwei Guan, He-Yang Xu, Henghui Ding, Yuan Wu, et al. AutoFly: Vision-language-action model for UAV autonomous navigation in the wild. In ICLR, 2026.arxiv.org/abs/2602.09657
  11. [11]Hengxing Cai, Jinhan Dong, Jingjun Tan, Jingcheng Deng, Sihang Li, Zhifeing Gao, Haidong Wang, Zicheng Su, Agachai Sumalee, and Renxin Zhong. FlightGPT: Towards generalizable and interpretable UAV vision-and-language navigation with vision-language models. In EMNLP, 2025. 1, 2arxiv.org/abs/2505.12835
  12. [12]Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. RT-1: Robotics transformer for real-world control at scale. In RSS, 2023. 1arxiv.org/abs/2212.06817
  13. [13]Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023.arxiv.org/abs/2307.15818
  14. [14]Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. In RSS, 2024.arxiv.org/abs/2405.12213
  15. [15]Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. OpenVLA: An open-source vision-language-action model. In CoRL, 2024.arxiv.org/abs/2406.09246
  16. [16]Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. π0: A vision-language-action flow model for general robot control. In RSS, 2025. 1DOI
  17. [17]Wen Jiang, Li Wang, Kangyao Huang, Wei Fan, Jinyuan Liu, Shaoyu Liu, Hongwei Duan, Bin Xu, and Xiangyang Ji. LongFly: Long-horizon UAV vision-and-language navigation with spatiotemporal context integration. arXiv preprint arXiv:2512.22010, 2025. 1, 2, 5, 6arxiv.org/abs/2512.22010
  18. [18]Boxiong Wang, Hui Kang, Geng Sun, Jiahui Li, Chao Yu, and Daxin Tian. RefineFly: Failure-aware post-training for aerial vision-language navigation. arXiv preprint arXiv:2608.09467, 2026. 1, 2, 5, 6arxiv.org/abs/2608.09467
  19. [19]Dean A. Pomerleau. ALVINN: An autonomous land vehicle in a neural network. In NeurIPS, 1988. 1
  20. [20]Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In AISTATS, 2010.
  21. [21]Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In AISTATS, 2011. 1, 2, 6arxiv.org/abs/1011.0686
  22. [22]Michael Kelly, Chelsea Sidrane, Katherine Driggs-Campbell, and Mykel J. Kochenderfer. HG-Dagger: Interactive imitation learning with human experts. In ICRA, 2019. 2arxiv.org/abs/1810.02890
  23. [23]Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. DART: Noise injection for robust imitation learning. In CoRL, 2017.arxiv.org/abs/1703.09327
  24. [24]Ryan Hoque, Ashwin Balakrishna, Ellen Novoseller, Albert Wilcox, Daniel S. Brown, and Ken Goldberg. ThriftyDAgger: Budget-aware novelty and risk gating for interactive imitation learning. In CoRL, 2021. 1, 2arxiv.org/abs/2109.08273
  25. [25]Oussama Khatib. Real-time obstacle avoidance for manipulators and mobile robots. The International Journal of Robotics Research, 1986. 1, 4
  26. [26]Xin Zhou, Zhepei Wang, Hongkai Ye, Chao Xu, and Fei Gao. EGO-Planner: An ESDF-free gradient-based local planner for quadrotors. IEEE Robotics and Automation Letters, 2021. 1arxiv.org/abs/2008.08835
  27. [27]Zhuoyuan Yu, Yuxing Long, Zihan Yang, Chengyan Zeng, Hongwei Fan, Jiaо Zhang, and Hao Dong. CorrectNav: Self-correction flywheel empowers vision-language-action navigation model. In AAAI, 2026. 1, 3arxiv.org/abs/2508.10416
  28. [28]Wenli Xiao, Haotian Lin, Andy Peng, Haoru Xue, Tairan He, Yuqi Xie, Fengyuan Hu, Jimmy Wu, Zhengyi Luo, Linxi Fan, et al. Self-improving vision-language-action models with data generation via residual RL. arXiv preprint arXiv:2511.00091, 2025. 1, 2, 3arxiv.org/abs/2511.00091
  29. [29]Ruibin Ju, Ruifeng Cai, Qin Xu, Zhuojie Chen, and Lin Wang. X-Fly: Learning to reach and stop in full-chunk aerial vision-language navigation. arXiv preprint arXiv:2610.21450, 2026. 2arxiv.org/abs/2610.21450
  30. [30]Baining Zhao, Jiacheng Xu, Weicheng Feng, Xin Zhang, Zhaolu Wang, Haoyang Wang, Shilong Ji, Ziyou Wang, Jianjie Fang, Zhiheng Zheng, et al. WorldVLN: Autoregressive world action model for aerial vision-language navigation. arXiv preprint arXiv:2605.15964, 2026.arxiv.org/abs/2605.15964
  31. [31]Zeyuan Ma, Jiaxin Chen, and Di Huang. From semantic grounding to decision optimization: A unified framework for long-horizon UAV vision-language navigation. In ACM MM, 2026. 2DOI
  32. [32]Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. STaR: Bootstrapping reasoning with reasoning. In NeurIPS, 2022. 2, 3arxiv.org/abs/2203.14465
  33. [33]Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, Arnaud Doucet, Orhan Firat, and Nando de Freitas. Reinforced self-training (ReST) for language modeling. arXiv preprint arXiv:2308.08998, 2023. 3arxiv.org/abs/2308.08998
  34. [34]Avi Singh, John D. Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J. Liu, James Harrison, Jaehoon Lee, Kelvin Xu, et al. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research, 2024. 3
  35. [35]Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. In ICML, 2024. 3arxiv.org/abs/2401.10020
  36. [36]Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, Yi Su, John D Co-Reyes, Avi Singh, Kate Baumli, Shariq Iqbal, Colton Bishop, Rebecca Roelofs, et al. Training language models to self-correct via reinforcement learning. In ICLR, 2025.arxiv.org/abs/2409.12917
  37. [37]DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 3DOI
  38. [38]Songze Li, Haoming Gu, Ziyang Gong, Zhitong Wang, Gengze Zhou, Zun Wang, Yuhan Zhu, Tianxiang Jiang, Juanzi Li, Qi Wu, et al. Data-centric recursive self-improvement for navigation via policy-synchronized language-action dual bootstrapping. arXiv preprint arXiv:2610.15782, 2026. 2, 3arxiv.org/abs/2610.15782
  39. [39]Yuzhi Zhang, Xinyu Liu, and Yu Zhang. Recursive self-improvement of visuomotor policies through local recovery supervision. arXiv preprint arXiv:2610.05151, 2026. 3arxiv.org/abs/2610.05151
  40. [40]Peidong Wang, Ying Chang, Zhiming Ma, Xiaoxiao Yang, Shi Feng, Daqi Liu, Yuzhe Yang, Xu Cao, Chuang Wang, Jian Su, et al. AdaRSI: On-demand recursive self-improvement for embodied agents. arXiv preprint arXiv:2610.20474, 2026. 3arxiv.org/abs/2610.20474
  41. [41]Haoran Lang, Haotao Lu, Shiyu Sang, Haoyang Luo, Guo Chen, Qun Li, Jingyi Yu, Ye Shi, and Jingya Wang. EmbodiRSI: Recursive self-improvement for data-efficient robot adaptation. arXiv preprint arXiv:2609.38905, 2026. 3arxiv.org/abs/2609.38905
  42. [42]Linxuan He, Yuying Tian, Lingxiang Fan, Jiaqi Pi, Yingqiao Lu, Shang Su, Mengkai Shi, and Shuo Feng. Policy-level recursive self-improvement for embodied AI with a criticality world model. arXiv preprint arXiv:2607.28251, 2026. 3arxiv.org/abs/2607.28251
  43. [43]Chenzhi Liu, Yue Zhang, Jiehong Lin, Jianan Wang, Bo Wang, Zhongrui Wang, and Xiaojun Qi. Mobi-Agent: Dual-loop recursive policy self-improvement for long-horizon mobile manipulation. In CoRL, 2026. 2, 3
  44. [44]Peiyi Wang, Lei Li, Zhihong Shao, R. X. Xu, Damai Dai, Yifei Li, Deli Chen, Y. Wu, and Zhifang Sui. Math-Shepherd: Verify and reinforce LLMs step-by-step without human annotations. In ACL, 2024. 2
  45. [45]Amirhossein Kazemnejad, Milad Aghajohari, Eva Portelance, Alessandro Sordoni, Siva Reddy, Aaron Courville, and Nicolas Le Roux. VinePPO: Refining credit assignment in RL training of LLMs. In ICML, 2025. 2
  46. [46]Audrey Huang, Adam Block, Dylan J. Foster, Dhruv Rohatgi, Cyril Zhang, Max Simchowitz, Jordan T. Ash, and Akshay Krishnamurthy. Self-improvement in language models: The sharpening mechanism. In ICLR, 2025. 2, 3arxiv.org/abs/2412.01951
  47. [47]Yuda Song, Hanlin Zhang, Carson Eisenach, Sham Kakade, Dean Foster, and Udaya Ghai. Mind the gap: Examining the self-improvement capabilities of large language models. In ICLR, 2025. 2, 3arxiv.org/abs/2412.02674
  48. [48]Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In ICML, 2002. 2, 3
  49. [49]Alan Fern, Robert Givan, and Sungwook Yoon. Approximate policy iteration with a policy language bias: Solving relational Markov decision processes. Journal of Artificial Intelligence Research, 2006. 3DOI
  50. [50]Bruno Scherrer, Victor Gabillon, Mohammad Ghavamzadeh, and Matthieu Geist. Approximate modified policy iteration. arXiv preprint arXiv:1205.3054, 2012. 2, 3arxiv.org/abs/1205.3054
  51. [51]Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, 2018. 2arxiv.org/abs/1711.07280
  52. [52]Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation. In NeurIPS, 2018.arxiv.org/abs/1806.02724
  53. [53]Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, 2022.arxiv.org/abs/2202.11742
  54. [54]Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scal- ing data generation in vision-and-language navigation. In ICCV, 2023.
  55. [55]Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, 2020.arxiv.org/abs/2004.02857
  56. [56]Dong An, Hanqing Wang, Wenguan Wang, Zun Wang, Yan Huang, Keji He, and Liang Wang. ETPNav: Evolving topological planning for vision-language navigation in continuous environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.DOI
  57. [57]Dhruv Shah, Blazej Osinski, Brian Ichter, and Sergey Levine. LM-Nav: Robotic navigation with large pre-trained models of language, vision, and action. In CoRL, 2022. 2arxiv.org/abs/2207.04429
  58. [58]Yin Tang, Jiawei Ma, Jinrui Zhang, Alex Jinpeng Wang, and Deyu Zhang. Mitigating error accumulation in continuous navigation via memory-augmented Kalman filtering. In ICML, 2026. 2, 5, 6arxiv.org/abs/2602.11183
  59. [59]Kunhui Wang, Xintong Zhang, Junyu Gao, and Changsheng Xu. ForeFly: A dual-horizon world action model for aerial vision-language navigation. arXiv preprint arXiv:2609.33581, 2026. 2arxiv.org/abs/2609.33581
  60. [60]Yan Deng and Fei Xu. DreamFly: Causal memory and receding-horizon diffusion planning for aerial vision-language navigation. arXiv preprint arXiv:2608.12308, 2026.arxiv.org/abs/2608.12308
  61. [61]Shengtao Zheng, Kai Li, Weichen Zhang, Yu Meng, Chen Gao, Xinlei Chen, Yong Li, and Xiao-Ping Zhang. World-Fly: A world-model-based vision-language-action model for UAV navigation. arXiv preprint arXiv:2606.06147, 2026.arxiv.org/abs/2606.06147
  62. [62]Xuchen Liu, Jiawei Huang, Shihao Xia, Bingxi Liu, Jinqiang Cui, and Jiajun Yang. ImagineUAV: Aerial vision-language navigation via world-action modeling and kinodynamic planning. arXiv preprint arXiv:2606.01205, 2026.arxiv.org/abs/2606.01205
  63. [63]Yutong Liu, Xiaojei Li, Mingzhu Xu, and Jianlong Wu. Air-Foresight: Current-to-future spatial map imagination with cross-space planning consistency for UAV-VLN. In ACM MM, 2026. 2
  64. [64]Yunpeng Gao, Zhigang Wang, Pengfei Han, Linglin Jing, Dong Wang, and Bin Zhao. Exploring spatial representation to enhance LLM reasoning in aerial vision-language navigation. arXiv preprint arXiv:2410.08500, 2024. 2arxiv.org/abs/2410.08500
  65. [65]Ruipu Wu, Yige Zhang, Jinyu Chen, Linjiang Huang, Shifeng Zhang, Xu Zhou, Liang Wang, and Si Liu. Aero-Duo: Aerial duo for UAV-based vision and language navigation. In ACM MM, 2025. 2DOI
  66. [66]Stéphane Ross and J. Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret learning. arXiv preprint arXiv:1406.5979, 2014. 3arxiv.org/abs/1406.5979
  67. [67]Stephane Ross, Narek Melik-Barkhudarov, Kumar Shaurya Shankar, Andreas Wendel, Debadeepta Dey, J. Andrew Bagnell, and Martial Hebert. Learning monocular reactive UAV control in cluttered natural environments. In ICRA, 2013. 3arxiv.org/abs/1211.1690
  68. [68]Dhiraj Gandhi, Lerrel Pinto, and Abhinav Gupta. Learning to fly by crashing. In IROS, 2017. 3arxiv.org/abs/1704.05588
  69. [69]Gregory Kahn, Pieter Abbeel, and Sergey Levine. BADGR: An autonomous self-supervised learning-based navigation system. IEEE Robotics and Automation Letters, 2021.arxiv.org/abs/2002.05700
  70. [70]Gregory Kahn, Pieter Abbeel, and Sergey Levine. LaND: Learning to navigate from disengagements. IEEE Robotics and Automation Letters, 2021. 3arxiv.org/abs/2010.04689
  71. [71]Fereshteh Sadeghi and Sergey Levine. CAD2RL: Real single-image flight without a single real image. In RSS, 2017. 3arxiv.org/abs/1611.04201
  72. [72]Antonio Loquercio, Elia Kaufmann, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Learning high-speed flight in the wild. Science Robotics, 2021.
  73. [73]Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 2023. 3DOI
  74. [74]Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3arxiv.org/abs/2203.02155
  75. [75]Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 5arxiv.org/abs/2305.18290
  76. [76]Jiwoo Hong, Noah Lee, and James Thorne. ORPO: Monolithic preference optimization without reference model. In EMNLP, 2024.arxiv.org/abs/2403.07691
  77. [77]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 3, 8arxiv.org/abs/2402.03300
  78. [78]Noam Razin, Sadhika Malladi, Adithya Bhaskar, Danqi Chen, Sanjeev Arora, and Boris Hanin. Unintentional un-alignment: Likelihood displacement in direct preference optimization. In ICLR, 2025. 3DOI
  79. [79]Fahim Tajwar, Anikait Singh, Archit Sharma, Rafael Rafailov, Jeff Schneider, Tengyang Xie, Stefano Ermon, Chelsea Finn, and Aviral Kumar. Preference fine-tuning of LLMs should leverage suboptimal, on-policy data. In ICML, 2024.arxiv.org/abs/2404.14367
  80. [80]Richard Yuanzhe Pang, Weizhe Yuan, Kyunghyun Cho, He He, Sainbayar Sukhbaatar, and Jason Weston. Iterative reasoning preference optimization. In NeurIPS, 2024. 3arxiv.org/abs/2404.19733
  81. [81]Thomas Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In NeurIPS, 2017. 3arxiv.org/abs/1705.08439
  82. [82]David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of Go without human knowledge. Nature, 2017.DOI
  83. [83]David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharsan Kumaran, Thore Graepel, et al. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 2018. 3DOI
  84. [84]Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves ImageNet classification. In CVPR, 2020.arxiv.org/abs/1911.04252
  85. [85]Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Tai Loy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In ICLR, 2024.arxiv.org/abs/2305.20050
  86. [86]Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. Weak-to-Strong Generalization: Eliciting strong capabilities with weak supervision. In ICML, 2024.arxiv.org/abs/2312.09390
  87. [87]Bowen Baker, Ingmar Kanitscheider, Todor Markov, Yi Wu, Glenn Powell, Bob McGrew, and Igor Mordatch. Emergent tool use from multi-agent autocurricula. In ICLR, 2020.arxiv.org/abs/1909.07528
  88. [88]Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X. Lee, Maria Bauza, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. RoboCat: A self-improving generalist agent for robotic manipulation. arXiv preprint arXiv:2306.11706, 2023.arxiv.org/abs/2306.11706
  89. [89]Zhiyuan Zhou, Pranav Atreya, Abraham Lee, Homer Walke, Oier Mees, and Sergey Levine. Autonomous improvement of instruction following skills via foundation models. In CoRL, 2024.arxiv.org/abs/2407.20635
  90. [90]Seyed Kamyar Seyed Ghasemipour, Ayzaan Wahid, Jonathan Tompson, Pannag Sanketi, and Igor Mordatch. Self-improving embodied foundation models. In NeurIPS, 2025.arxiv.org/abs/2509.15155
  91. [91]Songze Li, Zun Wang, Gengze Zhou, Jialu Li, Xiangyu Zeng, Ziyang Gong, Limin Wang, Yu Qiao, Qi Wu, Mohit Bansal, et al. Learning goal-oriented vision-and-language navigation with self-improving demonstrations at scale. arXiv preprint arXiv:2509.24910, 2025.arxiv.org/abs/2509.24910
  92. [92]Bingqian Lin, Yunshuang Nie, Khun Loun Zai, Ziming Wei, Mingfei Han, Rongtao Xu, Minzhe Niu, Jianhua Han, Hanwang Zhang, Liang Lin, et al. EvolveNav: Empowering LLM-based vision-language navigation via self-improving embodied reasoning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026.DOI
  93. [93]Hai Zhou, Fan Yang, Zeyi Chen, Lejun Zhang, Chenyang Fu, Jianping Li, Jiayi Zhu, Xianyang Liu, Yingxuan Yang, LiLi Meng, et al. Who improves the robot? toward physical recursive self-improvement. arXiv preprint arXiv:2610.18842, 2026.arxiv.org/abs/2610.18842
  94. [94]Hongyao Tang, Yi Ma, Pengyi Li, and Yifu Yuan. Generalized Agent Iteration: One formal framework for iterative policy improvement and recursive self-improvement. arXiv preprint arXiv:2609.13406, 2026.arxiv.org/abs/2609.13406
  95. [95]Hal Daumé III, John Langford, and Daniel Marcu. Search-based structured prediction. Machine Learning, 2009.arxiv.org/abs/0907.0786
  96. [96]Kai-Wei Chang, Akshay Krishnamurthy, Alekh Agarwal, Hal Daumé III, and John Langford. Learning to search better than your teacher. In ICML, 2015.arxiv.org/abs/1502.02206
  97. [97]Wen Sun, Arun Venkatraman, Geoffrey J. Gordon, Byron Boots, and J. Andrew Bagnell. Deeply AggreVaTeD: Differentiable imitation learning for sequential prediction. In ICML, 2017.arxiv.org/abs/1703.01030
  98. [98]Gerald Tesauro and Gregory R. Galperin. On-line policy improvement using Monte-Carlo search. In NeurIPS, 1996.
  99. [99]Dimitri P. Bertsekas, John N. Tsitsiklis, and Cynara Wu. Rollout algorithms for combinatorial optimization. Journal of Heuristics, 1997.
  100. [100]Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune. First return, then explore. Nature, 2021.
  101. [101]Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.
  102. [102]James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 2017.DOI
  103. [103]David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy P. Lillicrap, and Greg Wayne. Experience replay for continual learning. In NeurIPS, 2019.
  104. [104]Shuai Bai, Yuan Chen, Ruize Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianhao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL technical report. arXiv preprint arXiv:2511.1631, 2025.arxiv.org/abs/2511.21631
  105. [105]Shital Shah, Debadeepta Dey, Chris Stachniss, and Ashish Kapoor. AirSim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics, 2018.arxiv.org/abs/1705.05065

Paper details

vlnc:W_9r3m9s5s
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents