Pith. sign in

REVIEW 3 major objections 5 minor 90 references

Uncertainty-Aware World Model for Aerial Image-Goal Navigation

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that scoring a drone's candidate trajectories by only the part of the prediction-goal mismatch that cannot be explained by predicted future uncertainty improves aerial image-goal navigation, without the cost of sampling…

desk verdict A genuinely new trajectory-scoring mechanism for navigation world models, backed by thorough experiments and unusually candid supplements, though its central uncertainty interpretation rests on an unmeasured assumption. read the letter →

arxiv 2608.05597 v1 pith:SNWLNKZC submitted 2026-08-06 cs.CV

Deyi Zhu , Haoyu Fan , Yinan Zhu , Weichen Zhang , Shilin Ma , Xinlei Chen , Yansong Tang This is my paper · ORCID
classification cs.CV
keywords uncertainty-awareworldmodelimage-goalnavigationaerialout-of-distributiondetectionlatenttrajectoryscoringfuture-stateuncertaintyUAV
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 a drone navigation system can rank candidate trajectories more accurately if it stops asking 'how far is this goal from one predicted future?' and instead asks 'is this goal compatible with the whole set of plausible futures?' The authors build UA-NWM, a latent world model that predicts a deterministic future representation plus a learned uncertainty subspace around it. Any difference between the predicted future and the goal image is split into two parts: differences that lie along learned plausible-uncertainty directions, and a residual that cannot be explained. Only that residual is penalized when selecting a trajectory. The result, if correct, is better offline trajectory ranking, better closed-loop simulation navigation (76% success versus 70% for the deterministic backbone), and real-world drone flights, all at inference latency close to a fast deterministic baseline.

What carries the argument

The central mechanism is the Hierarchical Error Projection (HEP) module. HEP builds a coarse-to-fine pyramid of grid scales {1,2,7,14}; at each scale and cell it predicts a small set (rank R=2) of basis directions from the deterministic prediction and context latents via cross-attention and an MLP, then ridge-projects the current residual in that cell onto the basis span and subtracts the projected part from every patch in the cell. The residual left after the finest scale is the unexplained error used for scoring. This module is what converts trajectory scoring into a one-pass conditional out-of-distribution check: if the goal differs from the prediction mostly along learned uncertainty directions, the residual stays small and the candidate is accepted, whereas a goal whose mismatch is orthogonal to all learned plausible variations receives a large residual penalty.

What would settle it

Run a controlled experiment where the future ambiguity is known exactly: generate goals that differ from the prediction only along a predefined plausible variation (such as building color or occlusion pattern) and goals that differ by an impossible variation; if UA-NWM's residual score does not separate the two, the subspace is not capturing future uncertainty. Also compare the sequential greedy HEP residual with an exact orthogonal projection onto the union of all scale subspaces; if the ranking on AirGoal-10k changes materially, the hierarchical approximation itself, not the uncertainty decomposition, is the active ingredient.

Watch

Extended reading notes

Core claim

The paper's central claim is that future-state uncertainty should be modeled explicitly during trajectory scoring rather than averaged over or sampled. Concretely, UA-NWM treats a candidate trajectory as defining a conditional distribution of plausible future observations in a frozen DINO feature space. Using a module called HEP (Hierarchical Error Projection), it learns a low-rank subspace of plausible deviations around the deterministic prediction, conditioned only on the observation context and the candidate action sequence. The discrepancy between goal and predicted future is then projected onto this subspace; the explainable parallel component is discounted, and the perpendicular residual is used as the trajectory cost. UA-NWM therefore rejects a trajectory only when its goal discrepancy cannot be accounted for by plausible future variation, which the paper argues is why it outperforms deterministic point-prediction scorers and sampling-based stochastic scorers across their AirGoal-10k benchmark and in real-world deployment.

Load-bearing premise

The load-bearing premise is that the few learned uncertainty directions around the average prediction really do cover the ways the future can plausibly differ, so that whatever residual remains after projection is genuine evidence of a mismatched trajectory rather than model error or an artifact of the greedy projection.

Editorial extensions

If this is right

  • Trajectory selection no longer needs multiple stochastic future samples; the subspace is scored in one forward pass, keeping per-frame inference near the deterministic baseline (8.47 ms versus 8.06 ms offline).
  • Candidates whose goal mismatch is uncertainty-explainable are no longer falsely rejected, which the paper shows improves ranking and planning accuracy.
  • The same scorer can be plugged into sampling-based planners such as CEM, improving closed-loop success rate and SPL over the deterministic backbone and prior world models.
  • The scoring mechanism transfers to 2D ground navigation benchmarks, suggesting the idea is not limited to aerial views.
  • Real-world zero-shot deployment on a physical quadrotor succeeds without fine-tuning, using only RGB observations and the goal image.

Reading between the lines

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

  • The decomposition is effectively a learned nuisance-removal step: any goal-conditioned planner that suffers from predictable ambiguity (occlusion, drift) could adopt the same 'score only the unexplained residual' recipe.
  • Because the subspace is learned from context and action alone, residual energy will also absorb genuine model error when the world model is miscalibrated; a useful stress test is to measure how the score behaves under intentionally induced model errors versus genuinely incompatible goals.
  • The paper's own supplement notes that the sequential greedy projection is not the exact distance to the union of scale subspaces; comparing the greedy residual with an exact global projection on the same benchmark would reveal how much of the gain rides on the hierarchical approximation versus the uncertainty idea itself.
  • HEP also suggests a cheap stochastic generator: sampling inside the subspace produces diverse plausible futures, although the paper notes spatial coherence is limited and treats richer uncertainty representations as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes UA-NWM, a latent world model for aerial image-goal navigation that scores candidate trajectories by the unexplained residual between a predicted future representation and the goal representation after projecting the discrepancy onto a learned, context- and trajectory-dependent low-rank subspace. The key novelty is the Hierarchical Error Projection (HEP) module, which decomposes the prediction--goal discrepancy into an uncertainty-explainable component lying in a predicted subspace and an orthogonal unexplained residual, and scores trajectories using only the latter. The authors construct a new AirSim-based benchmark (AirGoal-10k), compare against policy-based and world-model baselines in offline trajectory ranking, standalone planning, and online closed-loop navigation, and report real-world UAV deployment. They claim consistent improvements in ATE/RPE and success rate while maintaining low inference latency.

Significance. If the central claim holds, UA-NWM is a useful contribution: it offers a fast, single-forward-pass alternative to stochastic future sampling for trajectory scoring, and the HEP decomposition is an interesting way to build uncertainty awareness into a deterministic latent world model. The new AirGoal-10k benchmark and the real-world deployment experiments are valuable assets for the community. The supplement is unusually candid about the assumptions behind the scoring rule, which helps a reviewer identify exactly what needs to be verified. However, the significance is tempered by three concerns: the theoretical justification of the e_perp-only score rests on an unmeasured high-anisotropy assumption; the experimental claims rely on a new benchmark with re-implemented baselines and no reported statistical uncertainty; and the learned subspace is not demonstrated to correspond to plausible future states rather than a generic low-rank residual model.

major comments (3)
  1. [Supplement G.1, Eq. (52)] The decision to score with only ||e_perp||^2 is derived from the Gaussian residual model, but the derivation shows that the calibrated score also contains ||e_parallel||^2/(tau^2+sigma^2). Dropping the parallel term is valid only in the high-anisotropy regime tau^2 >> sigma^2, yet the paper never measures tau^2 or sigma^2, nor does it verify that the learned subspace S contains plausible future states rather than generic low-rank residual directions. Section H's admission that random samples from S are spatially incoherent 'jigsaw puzzles' makes this gap load-bearing: if S is not a faithful model of the conditional future-state distribution, then the uncertainty interpretation of e_perp is not established. I ask for a quantitative analysis of the explained residual for correct versus incorrect candidate trajectories (e.g., ||e_parallel||/||e|| for matched and mismatched goals), or an ablation that uses the full Mahalanobis score, so that the e_perp-only rule can be shown to be the actual source of the reported gains rather than a manifestation of a learned residual metric.
  2. [Section 5.1.2 / Supplement D.1 / Tables 1-3] All head-to-head comparisons are conducted on AirGoal-10k, a benchmark introduced in this paper, and all baselines are re-implemented or modified for 3D (e.g., NWM with 25 diffusion steps instead of 250, One-Step WM trained from scratch because its official checkpoint is unavailable). The central state-of-the-art claim therefore rests on the authors' implementation choices and on a single evaluation set. In addition, no error bars, multiple seeds, or significance tests are reported; for the online experiment in Table 3, the difference between UA-NWM (76% SR) and RAE-NWM (70% SR) is only 6 successes out of 100 episodes. Please release code and trained checkpoints, report baseline tuning details, and provide multiple-seed results with confidence intervals, especially for the online closed-loop evaluation.
  3. [Eq. (15) / Supplement B.2] The HEP training loss L_HEP directly minimizes the normalized version of the inference-time score s_HEP, and the training signal is the discrepancy between the predicted future and the ground-truth future on correct transitions. This is a legitimate fitted scoring rule, but it means the paper's 'conditional OOD detection' formulation is not a parameter-free derivation: the projection is trained to minimize exactly the residual energy it later uses to score. I therefore ask the authors to (i) state this more plainly, and (ii) add an experiment that separates the benefit of the hierarchical low-rank structure from the benefit of a learned residual metric, for example by comparing HEP against a non-hierarchical learned projection of the same rank or against a direct learned distance on the residual. Such an experiment would also address the over-projection risk implied by the rank ablation in Table 5, where R=2 outperforms larger ranks.
minor comments (5)
  1. [Section 4.2 / Supplement G.2] The main paper refers to 'the subspace S' as if HEP performed a single orthogonal projection, but the supplement correctly notes that the coarse-to-fine procedure is a greedy sequential ridge projection rather than an exact global projection onto the sum of scale-wise subspaces. The main text should state this distinction explicitly, since it affects the geometric interpretation of e_perp.
  2. [Tables 1 and 3] The 'Time/Frame' and 'Time/Step' metrics are not defined consistently: offline per-frame latency is measured on a GPU with a fixed batch size, while online per-step time includes CEM planning and rollouts. Please clarify exactly what is measured in each column so that the latency comparison is reproducible.
  3. [Section 5.6] The real-world experiments are described qualitatively over five tasks with no quantitative metrics such as success rate, path efficiency, or final distance to goal. Please label these as pilot demonstrations and report quantitative measures if available, or state clearly that the results are qualitative.
  4. [Supplement F, Table 9] The 2D benchmark comparison includes NWM, RAE-NWM, and UA-NWM but not One-Step WM or MWM. Please clarify whether those baselines were evaluated on the 2D benchmarks and, if not, why they are omitted from the transfer experiment.
  5. [Throughout] The paper would benefit from a notation table or a reminder that 'norm(.)' denotes channel-wise L2 normalization in Eq. (4) but that the scoring rule in Eq. (5) is not normalized by the total discrepancy; the relationship between the normalized training loss and the unnormalized inference score should be stated in the main text, not only in Supplement B.2.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the learned residual-energy scorer is transparently trained and validated on held-out data and external benchmarks; the Gaussian derivation is explicitly an illustrative assumption, not a circular input.

full rationale

The paper's derivation chain is a learned scoring function, not a parameter-free first-principles prediction. HEP is trained with L_HEP (Eq. 15), which minimizes the normalized unexplained residual fraction on correct transitions, while inference uses the unnormalized residual energy (Eq. 5). The paper explicitly distinguishes these quantities: "This differs from the normalized training loss: at test time, retaining the absolute magnitude gives sharper discrimination between candidate trajectories" (Supplement B.2). Evaluation is performed on held-out test trajectories with identical candidate sets across baselines, and Supplement F further transfers the scorer to RECON and GO Stanford with NoMaD-generated candidates, so the scoring rule is not statistically forced by fitting the test data. The Supplement G.1 Gaussian derivation is explicitly framed as an illustrative model: "Our purpose is to make the assumptions behind the HEP interpretation explicit, rather than to claim that HEP estimates a fully calibrated likelihood of future observations." The high-anisotropy regime (tau^2 >> sigma^2) is acknowledged as an assumption, not a hidden circular input. The paper also candidly notes in Supplement H that random samples from the subspace are spatially incoherent "shuffled jigsaw puzzle" states, which weakens the uncertainty interpretation but is a correctness/interpretability limitation rather than circular reasoning. Self-citations appear only in related-work context and are not load-bearing; no uniqueness import, ansatz-by-citation, or renaming of a known result is present. The central empirical claim is therefore self-contained against held-out and external benchmarks.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The core scoring mechanism rests on several fitting choices (basis rank, scale pyramid, ridge penalty, rollout horizon, token counts) and on stated but unverified modeling assumptions from the supplement: the manifold hypothesis, the Gaussian residual model, the high-anisotropy regime, and the adequacy of the greedy projection. The AirGoal-10k benchmark is an invented evaluation entity, which concentrates evidence inside the paper's own pipeline.

free parameters (6)
  • Basis rank R = 2
    Selected by ablation over R = 1..5 on the validation or test set of AirGoal-10k (Table 5). The central scoring mechanism depends on this choice.
  • Scale pyramid G = {1,2,7,14}
    Chosen by the authors; ablation (Table 4) shows each scale contributes and removing any scale degrades performance. This is a hand-chosen architectural set affecting the scoring.
  • Ridge regularization lambda = 1e-3
    Chosen for the ridge-regularized projection in HEP (Eq. 30). The projection result depends on this value.
  • CEM parameters (samples, elites, iterations) = 32, 16, 3
    Shared with prior NWM evaluation protocol, but a hand-chosen planning budget that affects the online and offline planning results.
  • Context length C and horizon H = C=4, H=8
    Chosen by the authors; the rollout horizon is ablated in Supplement E.1 and affects performance.
  • Number of compressed and delta tokens K, M = K=32, M=32
    Selected by ablation (Table 8). These architectural capacities affect the representational quality of the backbone.
assumptions (5)
  • domain assumption DINOv3 features are a suitable space for measuring navigation goal compatibility.
    The whole method operates in frozen DINOv3 latent space, and all scoring happens there. The paper does not validate whether DINOv3 distances are calibrated to trajectory correctness beyond the reported benchmark numbers.
  • domain assumption The manifold hypothesis applies to plausible future DINO features: plausible futures concentrate near a low-dimensional manifold.
    Invoked in the Introduction (Section 1) and implicitly in the design of all HEP subspaces. The paper cites Bengio et al. for the manifold hypothesis but provides no evidence for its specific application to future-state DINO features.
  • ad hoc to paper The Gaussian residual model with V a + xi and the high-anisotropy regime tau^2 >> sigma^2 describe the future-state residual distribution.
    Supplement G.1 introduces this model to justify using only the orthogonal residual. The paper explicitly says this is an illustrative assumption and that in the general case a calibrated score would also penalize the parallel component. The regime tau^2 >> sigma^2 is not measured.
  • ad hoc to paper Greedy coarse-to-fine sequential projection behaves similarly to an exact projection for scoring purposes.
    Supplement G.2 admits that the greedy procedure is not an exact orthogonal projection onto the union of scale-wise subspaces unless the subspaces are mutually orthogonal, which is not enforced. The main text's Figure 3 explanation frames the residual as a distance to a subspace, which is only approximately true.
  • domain assumption Trajectory-scoring performance on AirGoal-10k transfers to real-world drone flight.
    The real-world experiments are qualitative and the sim-to-real gap is bridged without fine-tuning. The paper provides no quantitative real-world metrics to support this transfer.
invented entities (2)
  • Uncertainty subspace S per candidate trajectory
    purpose: A low-rank set of directions in DINO feature space approximating plausible future-state variation around the deterministic prediction, used to decompose the goal discrepancy.
    The subspace is a latent modeling construct defined by the learned HEP bases. It has no independent falsifiable handle outside the paper's own scoring objective.
  • AirGoal-10k benchmark
    purpose: A new simulated UAV image-goal navigation benchmark constructed from AirSim scenes and prior VLN start states, used to evaluate the method.
    The benchmark is introduced by the paper and is not a public standard benchmark at the time of writing. Validation on this benchmark cannot provide fully independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware World Model for Aerial Image-Goal Navigation." pith.science (2026). https://pith.science/paper/SNWLNKZC

@misc{pith2026260805597,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware World Model for Aerial Image-Goal Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SNWLNKZC}},
  note         = {Machine review of arXiv:2608.05597}
}
read the original abstract

Aerial image-goal navigation requires an unmanned aerial vehicle (UAV) to reach a target location specified by a goal image. Existing world-model-based methods rank candidate trajectories using predicted futures, but typically rely on only one or a few point predictions, which is inadequate for large-scale outdoor environments with substantial future-state uncertainty. To address this limitation, we propose the Uncertainty-Aware Navigation World Model (UA-NWM), an efficient latent world model for aerial image-goal navigation, which formulates trajectory scoring as conditional out-of-distribution detection. UA-NWM represents plausible futures with an uncertainty subspace and decomposes the prediction--goal discrepancy into uncertainty-explainable and unexplainable components. Only the unexplainable residual is used for scoring, enabling robust selection without multiple future samples. Extensive experiments demonstrate that UA-NWM consistently outperforms existing navigation world models while maintaining low inference latency. Real-world UAV experiments further validate its practical applicability. Project page: https://duryi.github.io/UA-NWM-Project-Page

Figures

Figures reproduced from arXiv: 2608.05597 by the authors.

Figure 1
Figure 1. Illustration of future-state uncertainty. sion foundation models (VFMs) [40, 54] to improve effi￾ciency, but most of them remain deterministic, and can only produce a single estimate, rather than representing the full future-state distribution. Recent methods [7, 25, 53] begin exploring stochastic prediction in VFM latent spaces. Despite these advances, a critical limitation remains— future-state uncertainty is not … view at source ↗
Figure 2
Figure 2. Comparison of UA-NWM and previous navigation world models in terms of future prediction and trajectory scoring. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. (a) Illustration of the uncertainty-aware trajectory scoring idea. Conventional deterministic methods score based on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: The overall training procedure of UA-NWM. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Overview of the AirGoal-10k benchmark. (a) Diverse [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of future predictions. The DINO latent features predicted by RAE-NWM and UA-NWM are decoded [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Illustration of real-world UAV experiments. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Detailed training and inference procedure of UA-NWM. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Comparison of (a) trajectory ranking and (b) standalone [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Configuration of custom-built UAV platform and system architecture for real-world experiments. [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Different strategies of sampling plausible states on pre [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Visualizations of hierarchical error projection process. [PITH_FULL_IMAGE:figures/full_fig_p026_12.png]
Figure 13
Figure 13. Figure 13: Additional visualizations of samples from the predicted uncertainty subspace. [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparisons on the trajectory ranking task. [PITH_FULL_IMAGE:figures/full_fig_p028_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparisons on the standalone planning task. [PITH_FULL_IMAGE:figures/full_fig_p029_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparisons on the standalone planning task. [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]
Figure 17
Figure 17. Figure 17: Qualitative results of online closed-loop navigation experiments. [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]
Figure 18
Figure 18. Figure 18: Additional qualitative results of real-world UAV experiments. [PITH_FULL_IMAGE:figures/full_fig_p032_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

90 extracted references · 51 canonical work pages

  1. [1]

    Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800,

    Niket Agarwal, Arslan Ali, Jon Allen, Martin Antolini, Ade- line Aubame, Alisson Azzolini, Junjie Bai, Maciej Bala, Yo- gesh Balaji, Josh Bapst, et al. Cosmos 3: Omnimodal world models for physical ai.arXiv preprint arXiv:2606.02800,

    arXiv

  2. [2]

    NavWAM: A Navigation World Action Model for Goal-Conditioned Visual Navigation

    Daichi Azuma, Taiki Miyanishi, Koya Sakamoto, Shuhei Kurita, Yaonan Zhu, Petr Khrapchenkov, Motoaki Kawan- abe, Yusuke Iwasawa, and Yutaka Matsuo. Navwam: A nav- igation world action model for goal-conditioned visual navi- gation.arXiv preprint arXiv:2606.13494, 2026. 3

    work page Pith review arXiv 2026

  3. [3]

    Uncertainty-informed active perception for open vocabulary object goal navigation

    Utkarsh Bajpai, Julius R ¨uckin, Cyrill Stachniss, and Marija Popovi´c. Uncertainty-informed active perception for open vocabulary object goal navigation. In2025 European Con- ference on Mobile Robots (ECMR), pages 1–7. IEEE, 2025. 13

    2025

  4. [4]

    Back to the fea- tures: Dino as a foundation for video world models.arXiv preprint arXiv:2507.19468, 2025

    Federico Baldassarre, Marc Szafraniec, Basile Terver, Vasil Khalidov, Francisco Massa, Yann LeCun, Patrick Labatut, Maximilian Seitzer, and Piotr Bojanowski. Back to the fea- tures: Dino as a foundation for video world models.arXiv preprint arXiv:2507.19468, 2025. 1

    arXiv 2025

  5. [5]

    Navigation world models

    Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15791–15801. IEEE, 2025. 1, 2, 3, 6, 7, 18, 19, 21, 22, 25

    2025

  6. [6]

    Rep- resentation learning: A review and new perspectives.IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013

    Yoshua Bengio, Aaron Courville, and Pascal Vincent. Rep- resentation learning: A review and new perspectives.IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828, 2013. 2

    2013

  7. [7]

    Vfmf: World modeling by forecast- ing vision foundation model features.arXiv preprint arXiv:2512.11225, 2025

    Gabrijel Boduljak, Yushi Lan, Christian Rupprecht, and Andrea Vedaldi. Vfmf: World modeling by forecast- ing vision foundation model features.arXiv preprint arXiv:2512.11225, 2025. 1

    arXiv 2025

  8. [8]

    Vi- sual navigation for mobile robots: A survey.Journal of In- telligent and Robotic Systems, 53(3):263–296, 2008

    Francisco Bonin-Font, Alberto Ortiz, and Gabriel Oliver. Vi- sual navigation for mobile robots: A survey.Journal of In- telligent and Robotic Systems, 53(3):263–296, 2008. 2

    2008

Show all 90 references
  1. [9]

    Ge- nie: Generative interactive environments

    Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Ge- nie: Generative interactive environments. InInternational Conference on Machine Learning, pages 4603–4623. PMLR,

  2. [10]

    Lawam: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026

    Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, et al. Lawam: Latent world action models for efficient dynamics-aware robot policies.arXiv preprint arXiv:2606.15768, 2026. 1

    2026

  3. [11]

    Geninav: Gener- ative model driven image-goal navigation via imagination- guided consistency flow matching

    Yuqi Chen, Junjie Gao, Yongzhou Pan, Siyuan Song, Zixuan Zhang, Jiaping Xiao, and Mir Feroskhan. Geninav: Gener- ative model driven image-goal navigation via imagination- guided consistency flow matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2026

  4. [12]

    Perception-aware path planning.arXiv preprint arXiv:1605.04151, 2016

    Gabriele Costante, Christian Forster, Jeffrey Delmerico, Paolo Valigi, and Davide Scaramuzza. Perception-aware path planning.arXiv preprint arXiv:1605.04151, 2016. 13

    2016 arXiv

  5. [13]

    Adanav: Adaptive reasoning with un- certainty for vision-language navigation.arXiv preprint arXiv:2509.24387, 2025

    Xin Ding, Jianyu Wei, Yifan Yang, Shiqi Jiang, Qianxi Zhang, Hao Wu, Fucheng Jia, Liang Mi, Yuxuan Yan, Wei- jun Wang, et al. Adanav: Adaptive reasoning with un- certainty for vision-language navigation.arXiv preprint arXiv:2509.24387, 2025. 13

    2025

  6. [14]

    Unified world models: Memory-augmented planning and foresight for visual navi- gation.arXiv preprint arXiv:2510.08713, 2025

    Yifei Dong, Fengyi Wu, Guangyu Chen, Zhi-Qi Cheng, Qiyu Hu, Yuxuan Zhou, Jingdong Sun, Jun-Yan He, Qi Dai, and Alexander G Hauptmann. Unified world models: Memory-augmented planning and foresight for visual navi- gation.arXiv preprint arXiv:2510.08713, 2025. 3

  7. [15]

    Flying to image-specified objects: 3d quadrotor navigation via cross-graph memory and viewpoint planning.arXiv preprint arXiv:2606.29917,

    Junjie Gao, Yuqi Chen, Yongzhou Pan, Yaosheng Deng, Ji- aping Xiao, and Mir Feroskhan. Flying to image-specified objects: 3d quadrotor navigation via cross-graph memory and viewpoint planning.arXiv preprint arXiv:2606.29917,

    arXiv

  8. [16]

    Uncertainty-aware gaussian map for vision- language navigation

    Jianzhe Gao, Rui Liu, Yuxuan Xu, Tongtong Cao, Yingxue Zhang, Zhanguang Zhang, Sida Peng, Yi Yang, and Wen- guan Wang. Uncertainty-aware gaussian map for vision- language navigation. InThe Fourteenth International Con- ference on Learning Representations, 2026. 13

    2026

  9. [17]

    Openfly: A comprehensive plat- form for aerial vision-language navigation.arXiv preprint arXiv:2502.18041, 2025

    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 plat- form for aerial vision-language navigation.arXiv preprint arXiv:2502.18041, 2025. 5, 18, 19

    2025

  10. [18]

    Flownav: Combining flow matching and depth priors for efficient navigation

    Samiran Gode, Abhijeet Nayak, D ´ebora NP Oliveira, Michael Krawez, Cordelia Schmid, and Wolfram Burgard. Flownav: Combining flow matching and depth priors for efficient navigation. In2025 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), pages 1776...

    2025

  11. [19]

    Igl-nav: Incremental 3d gaus- sian localization for image-goal navigation

    Wenxuan Guo, Xiuwei Xu, Hang Yin, Ziwei Wang, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Igl-nav: Incremental 3d gaus- sian localization for image-goal navigation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 6808–6817, 2025. 2

    2025

  12. [20]

    Ur- banfly: Uncertainty-aware planning for navigation amongst high-rises with monocular visual-inertial slam maps

    Sudarshan S Harithas, Ayyappa Swamy Thatavarthy, Gurki- rat Singh, Arun K Singh, and K Madhava Krishna. Ur- banfly: Uncertainty-aware planning for navigation amongst high-rises with monocular visual-inertial slam maps. In2023 American Control Conference (ACC), pages 557–563. IEEE,

  13. [21]

    A baseline for detect- ing misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016

    Dan Hendrycks and Kevin Gimpel. A baseline for detect- ing misclassified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136, 2016. 14

    2016 arXiv

  14. [22]

    Gonet: A semi-supervised deep learning approach for traversability estimation

    Noriaki Hirose, Amir Sadeghian, Marynel V ´azquez, Patrick Goebel, and Silvio Savarese. Gonet: A semi-supervised deep learning approach for traversability estimation. In2018 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 3044–3051. IEEE, 2018. 21, 22

    2018

  15. [23]

    Learning when to ask for help: Efficient interactive navigation via implicit un- certainty estimation

    Ifueko Igbinedion and Sertac Karaman. Learning when to ask for help: Efficient interactive navigation via implicit un- certainty estimation. In2024 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 9593–9599. IEEE, 2024. 13

    2024

  16. [24]

    Dino-foresight: Looking into 9 the future with dino.Advances in Neural Information Pro- cessing Systems, 38:163779–163811, 2026

    Efstathios Karypidis, Ioannis Kakogeorgiou, Spyridon Gi- daris, and Nikos Komodakis. Dino-foresight: Looking into 9 the future with dino.Advances in Neural Information Pro- cessing Systems, 38:163779–163811, 2026. 1

    2026

  17. [25]

    A frame is worth one token: Efficient generative world mod- eling with delta tokens

    Tommie Kerssies, Gabriele Berton, Ju He, Qihang Yu, Wufei Ma, Daan de Geus, Gijs Dubbelman, and Liang-Chieh Chen. A frame is worth one token: Efficient generative world mod- eling with delta tokens. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2026

  18. [26]

    Pathdreamer: A world model for indoor navigation

    Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 14738–14748, 2021. 3

    2021

  19. [27]

    A simple unified framework for detecting out-of-distribution samples and adversarial attacks.Advances in neural infor- mation processing systems, 31, 2018

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks.Advances in neural infor- mation processing systems, 31, 2018. 14

    2018

  20. [28]

    Taihri: Task-aware 3d human keypoints localization for close-range human-robot interaction.arXiv preprint arXiv:2604.08921, 2026

    Ao Li, Yonggen Ling, Yiyang Lin, Yuji Wang, Yong Deng, and Yansong Tang. Taihri: Task-aware 3d human keypoints localization for close-range human-robot interaction.arXiv preprint arXiv:2604.08921, 2026. 3

    2026 arXiv

  21. [29]

    Last-meter precision navigation for uavs: A diffusion- refined aerial visual servoing approach.arXiv preprint arXiv:2607.04352, 2026

    Yaxuan Li, Jiarui Zeng, Shaofei Huang, and Zhedong Zheng. Last-meter precision navigation for uavs: A diffusion- refined aerial visual servoing approach.arXiv preprint arXiv:2607.04352, 2026. 3

    2026 arXiv

  22. [30]

    Aerialvln: Vision-and-language navigation for uavs

    Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yan- ning Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 15384– 15394, 2023. 5, 18, 19

    2023

  23. [31]

    Energy-based out-of-distribution detection.Advances in neural information processing systems, 33:21464–21475,

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection.Advances in neural information processing systems, 33:21464–21475,

  24. [32]

    Imagineuav: Aerial vision- language navigation via world-action modeling and kinody- namic planning.arXiv preprint arXiv:2606.01205, 2026

    Xuchen Liu, Jiawei Huang, Shihao Xia, Bingxi Liu, Jin- qiang Cui, and Jiankun Yang. Imagineuav: Aerial vision- language navigation via world-action modeling and kinody- namic planning.arXiv preprint arXiv:2606.01205, 2026. 3

    2026 arXiv

  25. [33]

    Manigaussian: Dynamic gaus- sian splatting for multi-task robotic manipulation

    Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Ji- wen Lu, and Yansong Tang. Manigaussian: Dynamic gaus- sian splatting for multi-task robotic manipulation. InEu- ropean Conference on Computer Vision, pages 349–366. Springer, 2024. 3

    2024

  26. [34]

    Gwm: Towards scalable gaussian world models for robotic manipulation

    Guanxing Lu, Baoxiong Jia, Puhao Li, Yixin Chen, Ziwei Wang, Yansong Tang, and Siyuan Huang. Gwm: Towards scalable gaussian world models for robotic manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9263–9274, 2025. 1

    2025

  27. [35]

    Safe-pruner: Semantic attention-guided future- aware token pruning for efficient vision-language-action ma- nipulation.arXiv preprint arXiv:2605.29662, 2026

    Shilin Ma, Chubin Zhang, Changyuan Wang, Yuji Wang, Yue Wu, Zixuan Wang, Jingqi Tian, Zheng Zhu, and Yan- song Tang. Safe-pruner: Semantic attention-guided future- aware token pruning for efficient vision-language-action ma- nipulation.arXiv preprint arXiv:2605.29662, 2026. 3

    2026 arXiv

  28. [36]

    Navwm: A unified navigation world model for foresight-driven planning.arXiv preprint arXiv:2606.24101, 2026

    Yanghong Mei, Longteng Guo, Ming-Ming Yu, Guiyu Zhao, Xingjian He, and Jing Liu. Navwm: A unified navigation world model for foresight-driven planning.arXiv preprint arXiv:2606.24101, 2026. 3

    2026 arXiv

  29. [37]

    World models that know when they don’t know: Controllable video generation with calibrated uncertainty.arXiv preprint arXiv:2512.05927, 2025

    Zhiting Mei, Tenny Yin, Micah Baker, Ola Shorinwa, and Anirudha Majumdar. World models that know when they don’t know: Controllable video generation with calibrated uncertainty.arXiv preprint arXiv:2512.05927, 2025. 14

    2025

  30. [38]

    Do deep gener- ative models know what they don’t know?arXiv preprint arXiv:1810.09136, 2018

    Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, Di- lan Gorur, and Balaji Lakshminarayanan. Do deep gener- ative models know what they don’t know?arXiv preprint arXiv:1810.09136, 2018. 14

    2018 arXiv

  31. [39]

    Uncertainty-aware visually-attentive naviga- tion using deep neural networks.The International Journal of Robotics Research, 43(6):840–872, 2024

    Huan Nguyen, Rasmus Andersen, Evangelos Boukas, and Kostas Alexis. Uncertainty-aware visually-attentive naviga- tion using deep neural networks.The International Journal of Robotics Research, 43(6):840–872, 2024. 13

    2024

  32. [40]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Jou...

    2024

  33. [41]

    See- ing through uncertainty: Robust task-oriented optimization in visual navigation.Advances in Neural Information Pro- cessing Systems, 38:25259–25286, 2026

    Yiyuan Pan, Yunzhe Xu, Zhe Liu, and Hesheng Wang. See- ing through uncertainty: Robust task-oriented optimization in visual navigation.Advances in Neural Information Pro- cessing Systems, 38:25259–25286, 2026. 13

    2026

  34. [42]

    Competency-aware plan- ning for probabilistically safe navigation under perception uncertainty

    Sara Pohland and Claire Tomlin. Competency-aware plan- ning for probabilistically safe navigation under perception uncertainty. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 15291–15298. IEEE, 2025. 13

    2025

  35. [43]

    Perception matters: Enhancing embodied ai with uncertainty-aware semantic segmentation.arXiv preprint arXiv:2408.02297, 2024

    Sai Prasanna, Daniel Honerkamp, Kshitij Sirohi, Tim Welschehold, Wolfram Burgard, and Abhinav Val- ada. Perception matters: Enhancing embodied ai with uncertainty-aware semantic segmentation.arXiv preprint arXiv:2408.02297, 2024. 13

    2024 arXiv

  36. [44]

    Prior does matter: Visual navigation via denoising diffusion bridge models

    Hao Ren, Yiming Zeng, Zetong Bi, Zhaoliang Wan, Junlong Huang, and Hui Cheng. Prior does matter: Visual navigation via denoising diffusion bridge models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12100–12110, 2025. 1, 3, 6, 7, 19

    2025

  37. [45]

    Likelihood ratios for out-of-distribution detec- tion.Advances in neural information processing systems, 32, 2019

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshmi- narayanan. Likelihood ratios for out-of-distribution detec- tion.Advances in neural information processing systems, 32, 2019. 14

    2019

  38. [46]

    Optimization of computer simulation models with rare events.European Journal of Operational Research, 99(1):89–112, 1997

    Reuven Y Rubinstein. Optimization of computer simulation models with rare events.European Journal of Operational Research, 99(1):89–112, 1997. 3, 6, 8

    1997

  39. [47]

    Uncertainty-aware latent safety filters for avoiding out-of- distribution failures

    Junwon Seo, Kensuke Nakamura, and Andrea Bajcsy. Uncertainty-aware latent safety filters for avoiding out-of- distribution failures. InConference on Robot Learning, pages 4442–4472. PMLR, 2025. 14

    2025

  40. [48]

    Rapid exploration for open- world navigation with latent goal models.arXiv preprint arXiv:2104.05859, 2021

    Dhruv Shah, Benjamin Eysenbach, Gregory Kahn, Nicholas Rhinehart, and Sergey Levine. Rapid exploration for open- world navigation with latent goal models.arXiv preprint arXiv:2104.05859, 2021. 21, 22

    2021 arXiv

  41. [49]

    Gnm: A general navigation model to drive any robot

    Dhruv Shah, Ajay Sridhar, Arjun Bhorkar, Noriaki Hirose, and Sergey Levine. Gnm: A general navigation model to drive any robot. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7226–7233. IEEE,

  42. [50]

    Vint: A foundation model for visual navigation

    Dhruv Shah, Ajay Sridhar, Nitish Dashora, Kyle Stachowicz, Kevin Black, Noriaki Hirose, and Sergey Levine. Vint: A foundation model for visual navigation. InConference on Robot Learning, pages 711–733. PMLR, 2023. 1, 3, 6, 7, 19

    2023

  43. [51]

    Airsim: High-fidelity visual and physical simula- tion for autonomous vehicles

    Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simula- tion for autonomous vehicles. InField and service robotics: Results of the 11th international conference, pages 621–635. Springer, 2017. 5, 6, 17, 19

    2017

  44. [52]

    An efficient and multi-modal nav- igation system with one-step world model.arXiv preprint arXiv:2601.12277, 2026

    Wangtian Shen, Ziyang Meng, Jinming Ma, Mingliang Zhou, and Diyun Xiang. An efficient and multi-modal nav- igation system with one-step world model.arXiv preprint arXiv:2601.12277, 2026. 1, 2, 3, 6, 7, 19

    2026

  45. [53]

    Latent diffusion model without variational autoen- coder

    Minglei Shi, Haolin Wang, Wenzhao Zheng, Ziyang Yuan, Xiaoshi Wu, Xintao Wang, Pengfei Wan, Jie Zhou, and Ji- wen Lu. Latent diffusion model without variational autoen- coder. InInternational Conference on Learning Representa- tions, pages 154506–154537, 2026. 1

    2026

  46. [54]

    Dinov3.arXiv preprint arXiv:2508.10104, 2025

    Oriane Sim ´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104, 2025. 1, 2, 5, 6, 14, 20

    2025 arXiv

  47. [55]

    Improved base- lines with representation autoencoders.arXiv preprint arXiv:2605.18324, 2026

    Jaskirat Singh, Boyang Zheng, Zongze Wu, Richard Zhang, Eli Shechtman, and Saining Xie. Improved base- lines with representation autoencoders.arXiv preprint arXiv:2605.18324, 2026. 8, 20

    2026 arXiv

  48. [56]

    Nomad: Goal masked diffusion policies for nav- igation and exploration

    Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for nav- igation and exploration. In2024 IEEE International Con- ference on Robotics and Automation (ICRA), pages 63–70. IEEE, 2024. 1, 3, 6, 7, 19, 21, 22

    2024

  49. [57]

    Dreamwalker: Mental planning for contin- uous vision-language navigation

    Hanqing Wang, Wei Liang, Luc Van Gool, and Wen- guan Wang. Dreamwalker: Mental planning for contin- uous vision-language navigation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10873–10883, 2023. 3

    2023

  50. [58]

    Towards realistic uav vision-language navigation: Platform, benchmark, and methodology

    Xiangyu Wang, Donglin Yang, Hohin Kwan, Jinyu Chen, Hongsheng Li, Yue Liao, Si Liu, et al. Towards realistic uav vision-language navigation: Platform, benchmark, and methodology. InInternational Conference on Learning Rep- resentations, pages 7292–7310, 2025. 5

    2025

  51. [59]

    Dreamnav: A trajectory-based imaginative frame- work for zero-shot vision-and-language navigation.arXiv preprint arXiv:2509.11197, 2025

    Yunheng Wang, Yuetong Fang, Taowen Wang, Yixiao Feng, Yawen Tan, Shuning Zhang, Peiran Liu, Yiding Ji, and Ren- jing Xu. Dreamnav: A trajectory-based imaginative frame- work for zero-shot vision-and-language navigation.arXiv preprint arXiv:2509.11197, 2025. 3

    2025 arXiv

  52. [60]

    Vg-refiner: Towards tool-refined refer- ring grounded reasoning via agentic reinforcement learning

    Yuji Wang, Wenlong Liu, Jingxuan Niu, Haoji Zhang, and Yansong Tang. Vg-refiner: Towards tool-refined refer- ring grounded reasoning via agentic reinforcement learning. arXiv preprint arXiv:2512.06373, 2025. 3

    2025

  53. [61]

    Iterprime: Zero-shot referring image segmen- tation with iterative grad-cam refinement and primary word emphasis

    Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yan- song Tang. Iterprime: Zero-shot referring image segmen- tation with iterative grad-cam refinement and primary word emphasis. InProceedings of the AAAI Conference on Artifi- cial Intelligence, pages 8159–8168, 2025

    2025

  54. [62]

    Sam2-love: Segment anything model 2 in language- aided audio-visual scenes

    Yuji Wang, Haoran Xu, Yong Liu, Jiaze Li, and Yansong Tang. Sam2-love: Segment anything model 2 in language- aided audio-visual scenes. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28932–28932. IEEE, 2025. 3

    2025

  55. [63]

    Image-goal navigation in complex environ- ments via modular learning.IEEE Robotics and Automation Letters, 7(3):6902–6909, 2022

    Qiaoyun Wu, Jun Wang, Jing Liang, Xiaoxi Gong, and Di- nesh Manocha. Image-goal navigation in complex environ- ments via modular learning.IEEE Robotics and Automation Letters, 7(3):6902–6909, 2022. 2

    2022

  56. [64]

    Uav-on: A benchmark for open-world object goal naviga- tion with aerial agents

    Jianqiang Xiao, Yuexuan Sun, Yixin Shao, Boxi Gan, Rongqiang Liu, Yanjin Wu, Weili Guan, and Xiang Deng. Uav-on: A benchmark for open-world object goal naviga- tion with aerial agents. InProceedings of the 33rd ACM In- ternational Conference on Multimedia, pages 13023–13029,

  57. [65]

    Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman fil- ter.IEEE Robotics and Automation Letters, 6(2):3317–3324,

    Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman fil- ter.IEEE Robotics and Automation Letters, 6(2):3317–3324,

  58. [66]

    Mwm: Mobile world models for action-conditioned consistent pre- diction.arXiv preprint arXiv:2603.07799, 2026

    Han Yan, Zishang Xiang, Zeyu Zhang, and Hao Tang. Mwm: Mobile world models for action-conditioned consistent pre- diction.arXiv preprint arXiv:2603.07799, 2026. 1, 2, 6, 7, 19

    2026

  59. [67]

    Sign: Safety-aware image-goal navigation for autonomous drones via reinforcement learning.IEEE Robotics and Au- tomation Letters, 11(2):1962–1969, 2025

    Zichen Yan, Rui Huang, Lei He, Shao Guo, and Lin Zhao. Sign: Safety-aware image-goal navigation for autonomous drones via reinforcement learning.IEEE Robotics and Au- tomation Letters, 11(2):1962–1969, 2025. 3, 5

    1962

  60. [68]

    Unigoal: Towards universal zero-shot goal- oriented navigation

    Hang Yin, Xiuwei Xu, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Unigoal: Towards universal zero-shot goal- oriented navigation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 19057–19066, 2025. 3

    2025

  61. [69]

    Manigaussian++: General robotic bimanual manipulation with hierarchical gaussian world model

    Tengbo Yu, Guanxing Lu, Zaijia Yang, Haoyuan Deng, Sea- son Si Chen, Jiwen Lu, Wenbo Ding, Guoqiang Hu, Yan- song Tang, and Ziwei Wang. Manigaussian++: General robotic bimanual manipulation with hierarchical gaussian world model. In2025 IEEE/RSJ International Conference on Int...

    2025

  62. [70]

    Navidiffusor: Cost-guided diffusion model for visual navigation

    Yiming Zeng, Hao Ren, Shuhang Wang, Junlong Huang, and Hui Cheng. Navidiffusor: Cost-guided diffusion model for visual navigation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 11994–12001. IEEE, 2025. 3

    2025

  63. [71]

    Clap: Contrastive latent action pretraining for learning vision- language-action models from human videos.arXiv preprint arXiv:2601.04061, 2026

    Chubin Zhang, Jianan Wang, Zifeng Gao, Yue Su, Tianru Dai, Cai Zhou, Jiwen Lu, and Yansong Tang. Clap: Contrastive latent action pretraining for learning vision- language-action models from human videos.arXiv preprint arXiv:2601.04061, 2026. 3

    2026

  64. [72]

    Rae-nwm: Navigation world model in dense visual representation space.arXiv preprint arXiv:2603.09241, 2026

    Mingkun Zhang, Wangtian Shen, Fan Zhang, Haijian Qin, Zihao Pei, and Ziyang Meng. Rae-nwm: Navigation world model in dense visual representation space.arXiv preprint arXiv:2603.09241, 2026. 1, 2, 3, 6, 7, 19, 21, 22

    2026 arXiv

  65. [73]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 3, 8, 18, 20

    2018

  66. [74]

    A survey of visual navigation: From geometry to 11 embodied ai.Engineering Applications of Artificial Intelli- gence, 114:105036, 2022

    Tianyao Zhang, Xiaoguang Hu, Jin Xiao, and Guofeng Zhang. A survey of visual navigation: From geometry to 11 embodied ai.Engineering Applications of Artificial Intelli- gence, 114:105036, 2022. 2

    2022

  67. [75]

    Aerial world model for long-horizon visual generation and navigation in 3d space.arXiv preprint arXiv:2512.21887, 2025

    Weichen Zhang, Peizhi Tang, Xin Zeng, Fanhang Man, Shi- quan Yu, Zichao Dai, Baining Zhao, Hongjin Chen, Yu Shang, Wei Wu, et al. Aerial world model for long-horizon visual generation and navigation in 3d space.arXiv preprint arXiv:2512.21887, 2025. 1, 3

    2025

  68. [76]

    Grounded vision-language navigation for uavs with open-vocabulary goal understanding.arXiv preprint arXiv:2506.10756, 2025

    Yuhang Zhang, Haosheng Yu, Jiaping Xiao, and Mir Fer- oskhan. Grounded vision-language navigation for uavs with open-vocabulary goal understanding.arXiv preprint arXiv:2506.10756, 2025. 3

    2025 arXiv

  69. [77]

    Latent-space autoregressive world model for efficient and robust image-goal navigation

    Zhiwei Zhang, Hui Zhang, Xieyuanli Chen, Kaihong Huang, Chenghao Shi, and Huimin Lu. Latent-space autoregressive world model for efficient and robust image-goal navigation. arXiv e-prints, pages arXiv–2511, 2025. 2, 3

    2025

  70. [78]

    Efficient image-goal navigation with representative latent world model.arXiv preprint arXiv:2511.11011, 2025

    Zhiwei Zhang, Hui Zhang, Kaihong Huang, Chenghao Shi, and Huimin Lu. Efficient image-goal navigation with representative latent world model.arXiv preprint arXiv:2511.11011, 2025. 1, 2, 3, 6

    2025

  71. [79]

    Worldvln: Autoregressive world action model for aerial vision-language navigation

    Baining Zhao, Jiacheng Xu, Weicheng Feng, Xin Zhang, Zhaolu Wang, Haoyang Wang, Shilong Ji, Ziyou Wang, Jian- jie Fang, Zhiheng Zheng, et al. Worldvln: Autoregressive world action model for aerial vision-language navigation. arXiv preprint arXiv:2605.15964, 2026. 3

    2026 arXiv

  72. [80]

    Diffusion transformers with representation autoen- coders.arXiv preprint arXiv:2510.11690, 2025

    Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representation autoen- coders.arXiv preprint arXiv:2510.11690, 2025. 8

    2025 arXiv

  73. [81]

    Dino-wm: World models on pre-trained visual features en- able zero-shot planning

    Gaoyue Zhou, Hengkai Pan, Yann Lecun, and Lerrel Pinto. Dino-wm: World models on pre-trained visual features en- able zero-shot planning. InInternational Conference on Ma- chine Learning, pages 79115–79135. PMLR, 2025. 1

    2025

  74. [82]

    Segment anything with motion, geometry, and semantic adaptation for complex nonlinear vi- sual object tracking.arXiv preprint arXiv:2605.22538, 2026

    Deyi Zhu, Yuji Wang, Yong Liu, Yansong Tang, Bingyao Yu, Jiwen Lu, and Jie Zhou. Segment anything with motion, geometry, and semantic adaptation for complex nonlinear vi- sual object tracking.arXiv preprint arXiv:2605.22538, 2026. 3

    2026 arXiv

  75. [83]

    Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi

    Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforce- ment learning. In2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3357–3364. IEEE,

  76. [84]

    Astra: General interactive world model with autoregressive denoising.arXiv preprint arXiv:2512.08931, 2025

    Yixuan Zhu, Jiaqi Feng, Wenzhao Zheng, Yuan Gao, Xin Tao, Pengfei Wan, Jie Zhou, and Jiwen Lu. Astra: General interactive world model with autoregressive denoising.arXiv preprint arXiv:2512.08931, 2025. 1 12 Uncertainty-Aware World Model for Aerial Image-Goal Navigation Supple...

    2025

  77. [85]

    Image-Goal Navigation

    Related Work 2 2.1. Image-Goal Navigation . . . . . . . . . . . . 2 2.2. World Models for Navigation . . . . . . . . 3

  78. [86]

    Uncertainty-Aware Trajectory Scoring

    Method 3 4.1. Uncertainty-Aware Trajectory Scoring . . . . 3 4.2. Hierarchical Error Projection (HEP) . . . . . 4 4.3. Deterministic Baseline and Training . . . . . 5

  79. [87]

    Experimental Settings

    Experiments 5 5.1. Experimental Settings . . . . . . . . . . . . 5 5.1.1 . Benchmark . . . . . . . . . . . . . . 5 5.1.2 . Baselines . . . . . . . . . . . . . . . 6 5.1.3 . Evaluation Metrics . . . . . . . . . . 6 5.1.4 . Implementation Details . . . . . . . 6 5.2. Offline Exp...

  80. [88]

    # 𝑜! 𝑜!"# 𝑧! 𝑧!

    Conclusion 8 A . Additional Related Work 13 A.1 . Uncertainty-Aware Visual Navigation . . . . 13 A.2 . OOD Detection . . . . . . . . . . . . . . . . 14 B . Implementation Details of UA-NWM 14 B.1. Model Structure . . . . . . . . . . . . . . . 14 B.2. Training and Inference Pro...

  81. [89]

    The loss weights areβ rec = 0.05, ρ= 0.2, andγ 1 =γ 2 =γ 3 = 1.0

    Stage 1 is trained for 10k steps with learning rate 10−4, weight decay0.05, 600 warmup steps, and minimum learning-rate ratio0.01. The loss weights areβ rec = 0.05, ρ= 0.2, andγ 1 =γ 2 =γ 3 = 1.0. We inject Gaussian noise with standard deviation0.01into compressed latents and0...

  82. [90]

    !!"= "+e∥ Obs. & Act. r1 = e r2 r7 r14 e⊥=r14+

    Ifτ 2 is not substantially larger thanσ 2, a calibrated quadratic score would also re- tain a non-negligible penalty one ∥; our score should then 22 be viewed as a conservative compatibility approximation that treats variation insideSas acceptable for navigation. This yields t...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.