Pith. sign in

REVIEW 4 major objections 6 minor 42 references

RecoverFly: A Failure-Aware Reinforcement Learning Post-Training Framework for Aerial Vision-Language Navigation

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read RecoverFly shows that re-training a UAV vision-language-action policy with token-level RL, failure replay, and a long-tail curriculum improves navigation success rates by 3.12 to 8.37 percentage points.

desk verdict Solid RL post-training recipe for UAV-VLA, but the headline gains over AerialVLA rest on quoted baselines rather than a same-pipeline rerun. read the letter →

arxiv 2608.09467 v1 pith:BBZ3PFBR submitted 2026-08-10 cs.CV cs.AI

Boxiong Wang , Hui Kang , Geng Sun , Jiahui Li , Chao Yu , Daxin Tian This is my paper · ORCID
classification cs.CVcs.AI
keywords aerialvision-languagenavigationreinforcementlearningpost-trainingtoken-levelPPOfailurereplaylong-tailscenecurriculumKLregularizationend-to-endVLApolicyTravelUAVbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that an end-to-end aerial vision-language-action (VLA) policy can be substantially improved after initial behavior cloning by a closed-loop reinforcement learning post-training stage that deliberately revisits its own failures. On the TravelUAV benchmark, the proposed framework, RecoverFly, combines token-level PPO with a dynamic failure replay pool, a two-stage long-tail scene curriculum, and reference-policy KL regularization. The result is the best reported success rate on the seen, unseen-map, and unseen-object splits, with absolute success-rate gains of 3.12 to 8.37 percentage points over the AerialVLA baseline under a rollout budget of roughly 30% of the training set. The paper argues that the key to these gains is turning sparse, delayed failure signals into reusable corrective experience while keeping the policy from drifting away from its learned capabilities.

What carries the argument

The central mechanism is the token-level PPO objective (Eq. 4), which applies the clipped importance-ratio surrogate separately to each valid action token while sharing a single action-level advantage across the tokens that decode into one 3-DoF command. This avoids a sequence-level joint ratio and per-token clipping decision, allowing the autoregressive action grammar to be optimized with online navigation rewards. Around this backbone, the dynamic failure replay pool (Eq. 5) stores unresolved task initializations and regenerates rollouts with the current policy, and the two-stage scene curriculum plus stage-wise reference-policy KL loss (Eq. 7) jointly steer the training distribution toward long-tailed scenes while constraining policy drift.

What would settle it

An ablation that swaps the token-level objective for a sequence-level objective while keeping all other components identical; if success rates do not drop, the token-level credit assignment is not the cause of the gains.

Watch

Extended reading notes

Core claim

The central discovery is that treating reinforcement learning as a failure-aware post-training stage, rather than a generic second-stage optimizer, converts closed-loop execution mistakes into a reusable learning signal for an autoregressive UAV control policy. RecoverFly stores unresolved task initializations in a dynamic failure pool and regenerates rollouts with the current policy, so that on-policy PPO keeps reliving the most informative failures without ever updating on stale trajectories. Token-level PPO assigns the same action-level advantage to every valid token in the autoregressive action sequence, enabling stable optimization of the grammar-constrained 3-DoF action space. A two-stage curriculum first trains under the original scene distribution and then rebalances toward rare scenes, while stage-wise KL regularization anchors the policy to the initial VLA policy and to the Stage I policy, limiting distribution shift. On TravelUAV, this combination yields the best performance across all three evaluation splits and improves success rate by 3.12 to 8.37 percentage points over the AerialVLA initialization.

Load-bearing premise

The framework assumes that the delayed success or failure of a flight is spread evenly across all the tokens that make up a single control command, so that rewarding each token with the same advantage is a fair credit assignment.

Editorial extensions

If this is right

  • End-to-end UAV-VLA policies can be improved without extra annotations, external detectors, or hand-designed controllers, purely from closed-loop interaction feedback.
  • The largest gains occur on hard, long-horizon trajectories, suggesting that corrective RL post-training addresses error accumulation rather than only refining short-range control.
  • Because the improvement transfers to unseen maps and unseen objects, the learned recovery behavior generalizes beyond the training environments.
  • The two-stage curriculum shows that deliberately rebalancing scene frequencies during post-training outperforms both uniform sampling and the original distribution, even with the same number of training stages.
  • A total rollout budget of about 30% of the training-set size is sufficient to achieve these gains, indicating sample-efficient use of online interaction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same failure-aware replay plus token-level RL recipe could apply to other grammar-constrained autoregressive action spaces, such as grounded manipulation or embodied instruction following, where action tokens have heterogeneous semantic roles.
  • The token-level advantage sharing is a strong credit-assignment assumption; a natural extension is to weight tokens by their predicted effect on the outcome rather than giving all tokens equal blame.
  • The reported gains are on a simulated benchmark; a testable extension is whether the learned recovery behavior transfers to real quadcopter flights under domain shift or with added sensor noise.
  • The curriculum currently acts on scene frequencies; a further step would be to make the curriculum adaptive to the policy's own failure distribution, coupling failure replay with scene difficulty.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper presents RecoverFly, a reinforcement-learning (RL) post-training framework for end-to-end UAV vision-language-action (VLA) policies on the TravelUAV benchmark. The method initializes from the AerialVLA checkpoint (an OpenVLA-7B backbone with a LoRA adapter) and combines four components: token-level PPO with a shared, action-level advantage for grammar-constrained autoregressive action tokens (Eq. 4), a dynamic failure replay pool that re-samples unresolved task initializations under the current policy (Eq. 5), a two-stage long-tail scene curriculum that moves from empirical scene frequencies to equal-quota sampling (Eq. 6), and stage-wise reference-policy KL regularization (Eq. 7). With a rollout budget of about 30% of the training set, the paper reports success-rate improvements over AerialVLA of 8.37, 5.39, and 3.12 percentage points on the Seen, Unseen Map, and Unseen Object splits, respectively, and best overall SR and SPL on all three splits, with three-seed means and standard deviations. Incremental ablations (Table 4) attribute the gains to the combination of failure replay, KL regularization, and the two-stage curriculum.

Significance. If the reported results hold, the paper makes a solid empirical contribution: it demonstrates that closed-loop RL post-training of a 7B-scale VLA policy is feasible at a modest rollout budget (about 30% of the training split) and yields practically meaningful gains (3 to 8 SR points) that are consistent across seen, unseen-map, and unseen-object splits. Strengths of the submission include the reporting of three-seed means and standard deviations for the main results, a clean ablation design that isolates each component, evaluation on held-out splits with no circularity between the fitted policy and the reported numbers, and transparent reporting of the compute budget (21 hours on 8 A100 GPUs). The consistently larger gains on hard trajectories than on easy ones are a credible sign that corrective feedback targets the error-accumulation regime. The main reservations concern the quoted (not re-run) AerialVLA baseline that anchors the headline claim, the absence of the appendices that contain the experimental configuration, and the single-seed ablation protocol.

major comments (4)
  1. [Experiments, 'Baselines' paragraph; Tables 1–3] The AerialVLA rows in Tables 1–3 are quoted from the AerialVLA publication rather than produced by re-running the AerialVLA checkpoint in the authors' evaluation harness. The 'Baselines' paragraph states: 'For all baselines, we report results directly from the corresponding publications on the same evaluation splits,' which is in tension with the paper's characterization of the AerialVLA comparison as 'a direct evaluation of RL post-training.' Because RecoverFly is initialized from AerialVLA, the authors necessarily have the checkpoint and can re-run it under their exact evaluation conditions (AirSim/TravelUAV version, metric implementation, success thresholds, termination rules, seeds). This is load-bearing for the headline claim: the gains of 8.37, 5.39, and 3.12 SR points are differences against these quoted values, and on the Unseen Object split the 3.12-point margin is modest relative to both typical cross-harness variability and the reported seed variance of RecoverFly itself (std 0.81), while the baseline's own variance is unreported. Table 4 compounds the issue because every ablation delta is computed against the same quoted AerialVLA values. I would accept quoted published numbers for third-party baselines, but for the initialization policy the comparison should be measured in the same harness; please re-run AerialVLA (and, if feasible, at least the strongest other baselines) and report the measured values.
  2. [Method (Eqs. 2, 8); Experiments, 'Implementation Details'; Appendices A.1, A.2, B, C] Several load-bearing configuration details are deferred to appendices that are absent from the submitted text: the reward constants (r_prog min/max and κ_p in Eq. 2), the replay ratio η and the maximum replay attempts N_max in Eq. 5, the KL and value coefficients β and c_v in Eq. 8, all PPO/GAE hyperparameters (γ, λ, ε, learning rate), and the details of the rare-scene partition and Stage II sampling strategy (Appendix A.1), the remaining implementation configuration (Appendix A.2), the replay behavior analysis (Appendix B), and the token-level versus sequence-level PPO comparison (Appendix C). The paper itself states that 'the details about the rest configurations can be found in Appendix A.2,' and that Appendix A.1 contains the Stage II sampling details, but no such material is available in the version under review. Without it, Eqs. 2 and 8 cannot be instantiated by an independent reader, and the paper's claim that token-level PPO provides stable optimization cannot be checked. Please include the appendices (or move the essential hyperparameters and sampling details into the main text).
  3. [Method, 'Token-Level Policy Optimization Backbone' (Eq. 4)] Eq. 4 assigns the same action-level advantage Â_t to every valid token of a decoded action, presupposing that credit for a delayed navigation outcome is uniform across tokens with different semantic roles: the three control tokens ⟨Δx, Δz, Δψ⟩ and the optional LAND token. Success or failure depends much more directly on the LAND decision than on, say, a small altitude adjustment, so if per-token credit is not uniform, the per-token clipped ratio in Eq. 4 can bias the update toward tokens that were merely unlikely rather than causally responsible. I regard this as a correctness-risk concern rather than an observed failure; however, the paper's stated comparison of token-level versus sequence-level PPO is deferred to the absent Appendix C, so the choice of backbone is currently unverified. Please report that comparison, add an analysis of per-token credit allocation (for example, ablating shared versus token-specific advantages), or justify the uniform-credit assumption from the action grammar.
  4. [Experiments, 'Ablation Study' (Tables 4–5)] All ablations in Tables 4 and 5 use seed 1 only, while Tables 1–3 report three-seed standard deviations. Under those variances, several ablation deltas are comparable to the noise level: ID1 versus AerialVLA on Seen (+0.21; std 0.15), ID3 versus AerialVLA on Unseen Map (−0.11; std 1.01), and the Table 5 rows carry no variance at all. In particular, the −6.37-point Unseen Map change when adding failure replay alone (ID2) is a single draw; the interpretation that replay 'redistributes performance across splits rather than providing a uniform gain' needs multi-seed support. Please report at least three-seed means and standard deviations for the ablation conditions, or justify explicitly why single-seed results are sufficient here.
minor comments (6)
  1. [Throughout (Eqs. 2–8)] Subscript and superscript formatting is inconsistent and frequently missing: for example, Eq. (2) renders 'κp(Dt−1−Dt)', Eq. (3) uses 'dt' and 'δt' without subscripts, and 'bAt' appears in place of Â_t in the token-level objective. Please correct the typesetting throughout.
  2. [Table 4] The AerialVLA baseline row shows four dashes ('− − − −') although the table has only three framework-component columns, and the rows would benefit from explicit labels ('AerialVLA', 'ID1', etc.). Please clarify the header and row structure.
  3. [Experiments, 'Ablation Study'] There are several typos: 'AsillustratedinTable4' (missing spaces), 'Furthermore, We isolate' (mid-sentence capitalization), and 'The state σ_i ... records the active state' should read 'records whether the entry is active, solved, or dropped.'
  4. [Method, discussion of Eq. (4)] In the sentence 'A sequence-level alternative forms a joint ratio ... while this applies a single importance weight,' the referent of 'this' is ambiguous; name the sequence-level objective explicitly.
  5. [Experiments, 'Implementation Details'] The rollout budget of 'approximately 30% of the training-set size' should specify whether replay rollouts are counted in that budget and how the budget is split between Stage I and Stage II; as written, the budget accounting is ambiguous.
  6. [Conclusions] The sentence 'RecoverFly outperforms all comparison methods across all three splits' is stronger than the tables support on every metric (e.g., LongFly achieves higher OSR on the Seen Full set and lower NE and higher OSR on the Unseen Object Hard subset); consider qualifying the claim to SR and SPL.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; empirical SR claims are measured on held-out splits, with minor self-citation to RLinf and quoted AerialVLA baselines creating verification gaps.

full rationale

RecoverFly is an empirical systems paper: the central claim is a measured success-rate improvement on the TravelUAV Seen, Unseen Map, and Unseen Object test splits, not a quantity derived from its own definitions. The token-level PPO objective in Eq. 4 is a training objective, and Tables 1-3 report evaluations on held-out splits, so no prediction reduces to a fitted value by construction. The paper's use of RLinf (Yu et al. 2025), co-authored by Chao Yu, as the implementation substrate and token-level PPO starting point is a self-citation, but it is not load-bearing: the reported gains and ablations stand on the authors' own rollouts and evaluation, and the citation provides infrastructure rather than the source of the empirical conclusion. The statement that 'For all baselines, we report results directly from the corresponding publications on the same evaluation splits' means the AerialVLA comparison numbers are quoted rather than re-run in the authors' pipeline; this is a reproducibility/verification gap, but it is not circularity because the baseline is an external benchmark, not a fitted parameter or a definitional identity. No uniqueness theorem, ansatz-hidden-in-citation, or renaming pattern is present. Deferred appendices (A.1, A.2, B, C) are a completeness concern, not evidence of circularity. The weakest-assumption concern about token-level credit assignment in Eq. 4 is a methodological caveat about whether per-token advantage sharing is appropriate, not a circular step. Overall, the derivation chain is self-contained: method components are defined independently of the reported metrics, and the empirical claims are falsifiable against held-out data.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central result rests on standard RL assumptions, simulator fidelity, and a set of hand-chosen reward, replay, curriculum, and KL hyperparameters whose exact values are not reported in the main text. No new physical entities are introduced.

free parameters (5)
  • Event reward constants R_e and progress reward scale kappa_p = Not reported (Appendix A.2)
    Reward in Eq. 2 shapes the learning signal for success, collision, stuck, away, early-stop, and timeout; no sensitivity analysis is provided, and values are deferred to the appendix.
  • Failure replay ratio eta and max replay attempts N_max = Not reported
    Control how often failed tasks are replayed and when they are dropped from the pool; central to the failure-replay mechanism.
  • KL and value loss coefficients beta and c_v = Not reported
    Balance the PPO objective, value loss, and reference-policy constraint in Eq. 8.
  • Stage II rare-scene partition and quota schedule = Not reported (Appendix A.1)
    Definition of long-tail scenes and the balanced sampling scheme determines the curriculum effect.
  • PPO and GAE hyperparameters (gamma, lambda, epsilon, learning rate) = Not reported
    Standard RL hyperparameters, but not listed in the main text; they affect stability and credit assignment.
assumptions (4)
  • standard math PPO with GAE provides stable policy improvement for token-level autoregressive policies when importance ratios are clipped per token.
    Invoked in Eq. 4 and GAE Eq. 3 without proof; standard RL theory.
  • domain assumption The TravelUAV/AirSim simulator provides faithful event detection and distance progress for closed-loop RL training.
    Rewards in Eq. 2 depend on collision, stuck, early-stop, and target distance computed by the simulator.
  • domain assumption The pretrained AerialVLA/OpenVLA-7B policy with LoRA provides a strong prior that RL post-training improves without catastrophic forgetting.
    The whole method initializes from AerialVLA and anchors via KL to the stage policies.
  • domain assumption Baselines reported from previous publications use the same evaluation splits and protocol as RecoverFly.
    Tables 1-3 compare RecoverFly against numbers taken directly from cited papers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RecoverFly: A Failure-Aware Reinforcement Learning Post-Training Framework for Aerial Vision-Language Navigation." pith.science (2026). https://pith.science/paper/BBZ3PFBR

@misc{pith2026260809467,
  author       = {Pith},
  title        = {Pith review of: RecoverFly: A Failure-Aware Reinforcement Learning Post-Training Framework for Aerial Vision-Language Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBZ3PFBR}},
  note         = {Machine review of arXiv:2608.09467}
}
read the original abstract

Unmanned aerial vehicle vision-language navigation (UAV-VLN) requires agents to translate visual observations and language instructions into reliable flight actions in complex environments. Although recent end-to-end UAV vision-language-action (UAV-VLA) policies reduce reliance on separately designed perception, planning, and control modules, their behavior-cloning objectives provide limited corrective supervision for interactive closed-loop execution. Reinforcement learning (RL) offers a promising solution, while its effectiveness is constrained by inefficient use of samples, long-tailed scene distributions, and policy distribution shift during optimization. To this end, we propose RecoverFly, a failure-aware RL post-training framework for end-to-end UAV-VLA policies. Specifically, RecoverFly adapts token-level RL for stable optimization of grammar-constrained autoregressive UAV actions, revisits unresolved failure cases to strengthen corrective learning and sample utilization, and combines a two-stage long-tail scene curriculum with reference-policy regularization to improve scene adaptation while preserving acquired capabilities. Experiments on the TravelUAV benchmark demonstrate that RecoverFly achieves the best performance on the seen, unseen-map, and unseen-object splits. Moreover, compared to the AerialVLA initialization, RecoverFly improves success rate by 3.12 to 8.37 percentage points under a total rollout budget of about 30\% of the training-set size, validating its effectiveness, robustness, and generalization capabilities.

Figures

Figures reproduced from arXiv: 2608.09467 by the authors.

Figure 1
Figure 1. Comparison of training paradigms for UAV￾VLN. Behavior cloning lacks corrective learning, and stan￾dard RL may discard informative failures during online sam￾pling. RecoverFly revisits unresolved tasks and learns recov￾ery behaviors from interaction feedback. ing end-to-end interface from onboard observations and in￾structions to executable controls, reducing reliance on sepa￾rately engineered perception, planning, … view at source ↗
Figure 2
Figure 2. Overview of RecoverFly. RecoverFly combines token-level PPO with dynamic failure replay to learn corrective behaviors from closed-loop interaction while preserving on-policy rollouts. Furthermore, a two-stage long-tail scene curriculum and stage-wise reference-policy KL regularization improve rare-scene adaptation and constrain policy distribution shift. Rocktäschel 2021). In contrast, RecoverFly introduces a dy￾nam… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 29 canonical work pages

  1. [1]

    Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments , year =

    Peter Anderson and Qi Wu and Damien Teney and Jake Bruce and Mark Johnson and Niko Sünderhauf and Ian Reid and Stephen Gould and Anton van den Hengel , booktitle =. Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments , year =

    work page

  2. [2]

    Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments , year =

    Jacob Krantz and Erik Wijmans and Arjun Majumdar and Dhruv Batra and Stefan Lee , booktitle =. Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments , year =

    work page

  3. [3]

    AerialVLN:

    Shubo Liu and Hongsheng Zhang and Yuankai Qi and Peng Wang and Yanning Zhang and Qi Wu , booktitle =. AerialVLN:. 2023 , pages =

    work page 2023

  4. [4]

    Aerial Vision-and-Dialog Navigation , year =

    Yue Fan and Winson Chen and Tongzhou Jiang and Chun Zhou and Yi Zhang and Xin Wang , booktitle =. Aerial Vision-and-Dialog Navigation , year =

    work page

  5. [5]

    Towards realistic

    Wang, Xiangyu and Yang, Donglin and Kwan, Hohin and Chen, Jinyu and Li, Hongsheng and Liao, Yue and Liu, Si and others , booktitle =. Towards realistic. 2025 , pages =

    work page 2025

  6. [6]

    2025 , archiveprefix =

    Lin, Peican and Sun, Gan and Liu, Chenxi and Li, Fazeng and Ren, Weihong and Cong, Yang , title =. 2025 , archiveprefix =. 2511.06182 , journal =

    arXiv 2025

  7. [7]

    2025 , archiveprefix =

    Jiang, Wen and Wang, Li and Huang, Kangyao and Fan, Wei and Liu, Jinyuan and Liu, Shaoyu and Duan, Hongwei and Xu, Bin and Ji, Xiangyang , title =. 2025 , archiveprefix =. 2512.22010 , journal =

    arXiv 2025

  8. [8]

    2026 , archiveprefix =

    Xu, Peng and Deng, Zhengnan and Deng, Jiayan and Gu, Zonghua and Wan, Shaohua , title =. 2026 , archiveprefix =. 2603.14363 , journal =

    arXiv 2026

Show all 42 references
  1. [9]

    OpenVLA:

    Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and Xiao, Ted and Balakrishna, Ashwin and Nair, Suraj and Rafailov, Rafael and Foster, Ethan and Lam, Grace and Sanketi, Pannag and others , booktitle =. OpenVLA:. 2025 , pages =

    2025

  2. [10]

    Tran and Radu Soricut and Anikait Singh and Jaspiar Singh and Pierre Sermanet and Pannag R

    Brianna Zitkovich and Tianhe Yu and Sichun Xu and Peng Xu and Ted Xiao and Fei Xia and Jialin Wu and Paul Wohlhart and Stefan Welker and Ayzaan Wahid and Quan Vuong and Vincent Vanhoucke and Huong T. Tran and Radu Soricut and Anikait Singh and Jaspiar Singh and Pierre Sermanet...

    2023

  3. [11]

    2017 , archiveprefix =

    John Schulman and Filip Wolski and Prafulla Dhariwal and Alec Radford and Oleg Klimov , title =. 2017 , archiveprefix =. 1707.06347 , journal =

    2017 arXiv

  4. [12]

    Long Ouyang and Jeffrey Wu and Xu Jiang and Diogo Almeida and Carroll L. Wainwright and Pamela Mishkin and Chong Zhang and Sandhini Agarwal and Katarina Slama and Alex Ray and John Schulman and Jacob Hilton and Fraser Kelton and Luke Miller and Maddie Simens and Amanda Askell ...

  5. [13]

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , year =

    St. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , year =

  6. [14]

    Shah, Shital and Dey, Debadeepta and Lovett, Chris and Kapoor, Ashish , booktitle =. Airsim:. 2017 , pages =

    2017

  7. [15]

    2025 , archiveprefix =

    Jiazhao Zhang and Anqi Li and Yunpeng Qi and Minghan Li and Jiahang Liu and Shaoan Wang and Haoran Liu and Gengze Zhou and Yuze Wu and Xingxing Li and Yuxin Fan and Wenjun Li and Zhibo Chen and Fei Gao and Qi Wu and Zhizheng Zhang and He Wang , title =. 2025 , archiveprefix =....

    2025

  8. [16]

    Reinforced Cross-Modal Matching and Self-Supervised Imitation Learning for Vision-Language Navigation , year =

    Wang, Xin and Huang, Qiuyuan and Celikyilmaz, Asli and Gao, Jianfeng and Shen, Dinghan and Wang, Yuan-Fang and Wang, William Yang and Zhang, Lei , booktitle =. Reinforced Cross-Modal Matching and Self-Supervised Imitation Learning for Vision-Language Navigation , year =

  9. [17]

    2025 , pages =

    Xiao, Jianqiang and Sun, Yuexuan and Shao, Yixin and Gan, Boxi and Liu, Rongqiang and Wu, Yanjin and Guan, Weili and Deng, Xiang , booktitle =. 2025 , pages =

    2025

  10. [18]

    OpenFly:

    Gao, Yunpeng and Li, Chenhui and You, Zhongrui and Liu, Junli and Li, Zhen and Chen, Pengan and Chen, Qizhi and Tang, Zhonghan and Wang, Liansheng and Yang, Penghui and others , booktitle =. OpenFly:

  11. [19]

    Citynavagent:

    Zhang, Weichen and Gao, Chen and Yu, Shiquan and Peng, Ruiying and Zhao, Baining and Zhang, Qian and Cui, Jinqiang and Chen, Xinlei and Li, Yong , booktitle =. Citynavagent:. 2025 , pages =

    2025

  12. [20]

    Li, Tianshun and Huai, Tianyi and Li, Zhen and Gao, Yichun and Li, Haoang and Zheng, Xinhu , booktitle =. SkyVLN:. 2025 , pages =

    2025

  13. [21]

    TypeFly:

    Chen, Guojun and Yu, Xiaojing and Ling, Neiwen and Zhong, Lin , journal =. TypeFly:. 2025 , pages =

    2025

  14. [22]

    See, Point, Fly:

    Hu, Chih Yao and Lin, Yang-Sen and Lee, Yuna and Su, Chih-Hai and Lee, Jie-Ying and Tsai, Shr-Ruei and Lin, Chin-Yang and Chen, Kuan-Wen and Ke, Tsung-Wei and Liu, Yu-Lun , booktitle =. See, Point, Fly:. 2025 , pages =

    2025

  15. [23]

    UAV-flow colosseo:

    Wang, Xiangyu and Yang, Donglin and Liao, Yue and Zheng, Wenhao and Dai, Bin and Li, Hongsheng and Liu, Si and others , booktitle =. UAV-flow colosseo:. 2026 , volume =

    2026

  16. [24]

    2025 , archiveprefix =

    Serpiva, Valerii and Lykov, Artem and Myshlyaev, Artyom and Khan, Muhammad Haris and Abdulkarim, Ali Alridha and Sautenkov, Oleg and Tsetserukou, Dzmitry , title =. 2025 , archiveprefix =. 2503.02572 , journal =

    2025 arXiv

  17. [25]

    2025 , archiveprefix =

    Lykov, Artem and Serpiva, Valerii and Khan, Muhammad Haris and Sautenkov, Oleg and Myshlyaev, Artyom and Tadevosyan, Grik and Yaqoot, Yasheerah and Tsetserukou, Dzmitry , title =. 2025 , archiveprefix =. 2503.01378 , journal =

    2025 arXiv

  18. [26]

    UAV-VLA:

    Sautenkov, Oleg and Yaqoot, Yasheerah and Lykov, Artem and Mustafa, Muhammad Ahsan and Tadevosyan, Grik and Akhmetkazy, Aibek and Cabrera, Miguel Altamirano and Martynov, Mikhail and Karaf, Sausar and Tsetserukou, Dzmitry , booktitle =. UAV-VLA:

  19. [27]

    Flightgpt:

    Cai, Hengxing and Dong, Jinhan and Tan, Jingjun and Deng, Jingcheng and Li, Sihang and Gao, Zhifeng and Wang, Haidong and Su, Zicheng and Sumalee, Agachai and Zhong, Renxin , booktitle =. Flightgpt:. 2025 , pages =

    2025

  20. [28]

    2025 , archiveprefix =

    Qi, Zhangyang and Zhang, Zhixiong and Yu, Yizhou and Wang, Jiaqi and Zhao, Hengshuang , title =. 2025 , archiveprefix =. 2506.17221 , journal =

    2025 arXiv

  21. [29]

    2025 , archiveprefix =

    Yu, Chao and Wang, Yuanqing and Guo, Zhen and Lin, Hao and Xu, Si and Zang, Hongzhi and Zhang, Quanlu and Wu, Yongji and Zhu, Chunyang and Hu, Junhao and others , title =. 2025 , archiveprefix =. 2509.15965 , journal =

    2025

  22. [30]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

  23. [31]

    Jordan and Pieter Abbeel , booktitle =

    John Schulman and Philipp Moritz and Sergey Levine and Michael I. Jordan and Pieter Abbeel , booktitle =. High-Dimensional Continuous Control Using Generalized Advantage Estimation , year =

  24. [32]

    Hindsight Experience Replay , year =

    Marcin Andrychowicz and Dwight Crow and Alex Ray and Jonas Schneider and Rachel Fong and Peter Welinder and Bob McGrew and Josh Tobin and Pieter Abbeel and Wojciech Zaremba , booktitle =. Hindsight Experience Replay , year =

  25. [33]

    Curriculum learning , year =

    Bengio, Yoshua and Louradour, J\'. Curriculum learning , year =

  26. [34]

    2025 , pages =

    Jungdae Lee and Taiki Miyanishi and Shuhei Kurita and Koya Sakamoto and Daichi Azuma and Yutaka Matsuo and Nakamasa Inoue , booktitle =. 2025 , pages =

    2025

  27. [35]

    History-Enhanced Two-Stage Transformer for Aerial Vision-and-Language Navigation , year =

    Xichen Ding and Jianzhe Gao and Cong Pan and Wenguan Wang and Jie Qin , booktitle =. History-Enhanced Two-Stage Transformer for Aerial Vision-and-Language Navigation , year =

  28. [36]

    LookasideVLN: Direction-Aware Aerial Vision-and-Language Navigation , year =

    Ning, Yuwei and Zhao, Ganlong and Qin, Yipeng and Liu, Si and Liu, Yang and Lin, Liang and Li, Guanbin , booktitle =. LookasideVLN: Direction-Aware Aerial Vision-and-Language Navigation , year =

  29. [37]

    2026 , pages =

    Fan, Chengjie and Pan, Cong and Liu, Zijian and Liu, Ningzhong and Qin, Jie , booktitle =. 2026 , pages =

    2026

  30. [38]

    Knepper and Yoav Artzi , booktitle =

    Valts Blukis and Yannick Terme and Eyvind Niklasson and Ross A. Knepper and Yoav Artzi , booktitle =. Learning to Map Natural Language Instructions to Physical Quadcopter Control using Simulated Flight , year =

  31. [39]

    Jordan and Philipp Moritz , booktitle =

    John Schulman and Sergey Levine and Pieter Abbeel and Michael I. Jordan and Philipp Moritz , booktitle =. Trust Region Policy Optimization , year =

  32. [40]

    Prioritized Level Replay , year =

    Minqi Jiang and Edward Grefenstette and Tim Rockt. Prioritized Level Replay , year =

  33. [41]

    2025 , archiveprefix =

    Lingfeng Zhang and Yuchen Zhang and Hongsheng Li and Haoxiang Fu and Yingbo Tang and Hangjun Ye and Long Chen and Xiaojun Liang and Xiaoshuai Hao and Wenbo Ding , title =. 2025 , archiveprefix =. 2511.13269 , journal =

    2025

  34. [42]

    2025 , archiveprefix =

    Mohamed Amine Ferrag and Abderrahmane Lakas and Merouane Debbah , title =. 2025 , archiveprefix =. 2511.11252 , journal =

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.