Pith. sign in

REVIEW 2 major objections 7 minor 69 references

UniNav: A Unified World-Action Diffusion Model for Visual Navigation

T0 review · 2 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A single diffusion model predicts future views and waypoints together, beating the strongest navigation baselines on all four benchmarks.

desk verdict Solid, clearly-written method paper joining the world-action navigation line; the benchmark margins are plausible but the baseline protocol needs spelling out before the headline claim is verifiable. read the letter →

arxiv 2608.03244 v1 pith:KVERJQKA submitted 2026-08-04 cs.AI

Changqing Zhou , Yueru Luo , Zeyu Jiang , Changhao Chen This is my paper · ORCID
classification cs.AI
keywords visualnavigationworld-actionmodeldiffusiontransformerimage-goalflowmatchingwaypointpredictionforesightcamerageometrytokens
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

The paper sets out to show that image-goal navigation can be one generative problem instead of two: a single diffusion model can predict both what the robot will see from future viewpoints and the continuous waypoint trajectory that takes it there. The payoff, if the claim holds, is that robots no longer have to trade visual foresight for speed—the same model gives interpretable future views and a ready-to-execute trajectory, with a stripped mode that skips image decoding and runs at 0.1 s per step. UniNav reports lower Absolute Trajectory Error than the strongest baseline on all four benchmarks tested: 3.7% lower on RECON, 24.7% lower on SaCSoN, 18.2% lower on GO Stanford, and 6.8% lower on SCAND. It also claims that geometric camera tokens are the bridge that lets unlabeled video data help waypoint prediction; without those tokens, added video data slightly hurts accuracy.

What carries the argument

The load-bearing object is a heterogeneous token sequence processed by a single transformer: future video latents from a frozen VAE, K continuous waypoint tokens, normalized camera-geometry tokens from a frozen 3D foundation model, history and goal tokens, and register tokens. All prediction tokens are noised under a flow-matching schedule and jointly denoised with three losses (video, waypoint, camera). The camera-geometry token is the bridge the argument depends on: it captures the 3D structure of each future view, so unlabeled video (which has no waypoints) can still supervise geometry, and waypoint denoising learns to anticipate visual and geometric change rather than only matching pixel

What would settle it

Re-run the comparison with each baseline fine-tuned on the same trajectory-labeled splits (C=3 history, K=6 waypoints) and the same ATE/RPE evaluator, or run UniNav-Fast closed-loop on a real robot against a fine-tuned ViNT under identical delays; if a fine-tuned baseline matches or beats UniNav, or closed-loop success does not improve, the central claim fails its own terms.

Watch

Extended reading notes

Core claim

UniNav is a world-action diffusion model built on a pretrained 1.3B video diffusion transformer. Given C history frames and a goal image, it constructs a token sequence of future video latents, waypoint tokens, camera-geometry tokens, history/goal context tokens, and learnable register tokens, and denoises the prediction tokens together under one flow-matching objective. The paper's central claim is that denoising future views, waypoints, and 3D camera tokens in a shared transformer is a better formulation for image-goal navigation than either waypoint-only policies or planning-based world models: jointly predicting the future view and the trajectory grounds the action in anticipated visual

Load-bearing premise

The reported margins rest on the assumption that the baselines were evaluated under the same protocol as UniNav—same data for training or fine-tuning, same waypoint parameterization, same ATE computation—and the paper does not state whether they were; if they were run zero-shot while UniNav was fine-tuned, the gaps would measure adaptation rather than architecture.

Editorial extensions

If this is right

  • A robot can switch between interpretive and fast modes from the same weights: UniNav-Full emits future frames for inspection, UniNav-Fast drops them and still predicts trajectories.
  • Since the mixed-training ablation shows video-only data only helps with camera-token denoising, future data needs can be fed by unlabeled video plus a frozen geometry model instead of collecting more waypoint-labeled trajectories.
  • The efficiency table says action-only inference at one step (0.109 s) is nearly as accurate as two-step sampling, so the model is viable for real-time closed-loop control if the offline errors carry over.
  • The joint generation distills world-model reasoning into the policy, avoiding the planning baseline's roughly 12,000 function evaluations per decision.
  • Future-frame quality becomes a byproduct of the same architecture, giving operators interpretable rollouts while the model also predicts actions.

Reading between the lines

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

  • Beyond the paper, the same 'geometric bridge' recipe—denoising a frozen 3D token alongside video and action tokens—could port world-action training to driving or manipulation, where future-view geometry is equally predictive of feasible actions.
  • The RECON ablation suggests a testable scaling law: as unlabeled video volume grows, trajectory accuracy should keep improving only if the camera-token objective scales too; pure video-prediction pretraining would plateau or regress.
  • If the one-step UniNav-Fast result holds under closed-loop control, the multi-step diffusion process is mostly a training-time device, and a single-step distilled trajectory head may recover most of the benefit.
  • The paper leaves implicit that its camera-token denoising design is a deployable substitute for oracle geometric conditioning: it captures most of the gain of conditioning on ground-truth future geometry without needing future frames at inference.
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

2 major / 7 minor

Summary. UniNav proposes a unified world-action diffusion model for image-goal navigation. It builds on the pretrained Wan2.1-1.3B video diffusion transformer and represents navigation as a joint flow-matching problem over three modalities: future video latents, continuous waypoint tokens, and camera-geometry tokens derived from a frozen geometry foundation model. The token sequence is processed by a shared transformer with separate decoding heads, and a mixed training strategy uses trajectory-labeled data plus video-only data. Two variants are introduced: UniNav-Full (joint future-frame and trajectory generation) and UniNav-Fast (asymmetric attention and removal of future-image tokens at inference for low-latency action prediction). Experiments on RECON, SaCSoN, GO Stanford, and SCAND report the best ATE among the tested methods on all four datasets, with the Fast variant reaching 0.109 s one-step inference latency. The paper also provides ablations on camera-token supervision and video-only data.

Significance. If the reported results hold up, the contribution is significant: it demonstrates that a single diffusion transformer can jointly predict interpretable future views and executable waypoints, and that the resulting action-only mode is fast enough for closed-loop use. The formulation is mostly clear, the flow-matching objective is standard, and the efficiency profile (1-2 steps on a 1.3B transformer) is plausible. The camera-token ablation (Table 2) and mixed-data ablation (Table 3) are informative. The central weakness is empirical verification: the baseline evaluation protocol is underspecified and no uncertainty estimates accompany the small headline margins. These issues are fixable and should be addressed before the benchmark claims can be accepted.

major comments (2)
  1. [§4.2, Table 1] The sentence 'all results in the main comparison are obtained using only trajectory-labeled navigation data' constrains UniNav's training data but does not specify the baseline protocol. Please state whether ViNT, NoMaD, NavDP, FlowNav, and NWM were (a) fine-tuned on the same train splits with the same K=6 waypoint parameterization and ATE/RPE computation, or (b) evaluated from their original checkpoints without adaptation. If (b), the headline reductions of 3.7% (RECON) and 6.8% (SCAND) over NWM may reflect domain adaptation rather than architectural benefit. This is load-bearing for the central claim.
  2. [§4.2, Table 4] No error bars, multiple seeds, or statistical tests are reported. The ATE gaps over the strongest baselines on RECON (0.314 vs 0.326) and SCAND (0.463 vs 0.497) are small, and Table 4 shows the same model's ATE varying by up to ~0.03 across sampling steps (e.g., UniNav-Fast 0.300–0.331). Run-to-run variance across training seeds could therefore be comparable to the claimed margins. Please report variance and, where feasible, significance tests.
minor comments (7)
  1. [§4.1, Eq. (8)] The total objective writes λ_traj, but Implementation details set 'λ_act = 5'. Align the notation.
  2. [§3.4] For UniNav-Fast, it is not stated whether camera-geometry tokens are also removed at inference or still sampled. Since camera tokens encode future-view geometry, clarify their status in the reported latency/accuracy.
  3. [§2.3/§4.2] UniWM and WAM-Nav are discussed as closely related world-action navigation models, but they are absent from Table 1. Please state availability or provide a comparison/discussion.
  4. [§4.1] PSNR/SSIM/LPIPS are listed as future-frame quality metrics, but no quantitative future-frame results appear in the main text; the appendix is referenced but not included in this version.
  5. [Table 1, RECON row] The RPE of both UniNav variants (0.121/0.122) is worse than NWM (0.115); the text's claim of 'consistent' improvement across metrics should be softened.
  6. [Table 4, NWM row] The latency (~900 s, NFE 12,000) needs a one-sentence description of how it is measured (e.g., planning rollouts, candidates per step) to be comparable.
  7. [§4.1, Metrics] Specify how ATE/RPE are computed from K=6 local waypoints (integration/alignment), since TUM metrics are normally defined for global trajectories.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: targets are defined from ground-truth data, results are held-out, and no load-bearing self-citation exists.

full rationale

UniNav is an empirical benchmark paper whose headline quantities (ATE/RPE on RECON, SaCSoN, GO Stanford, SCAND) are computed on held-out test data against ground-truth trajectories, not derived from the training losses (Eqs. 5–8). The flow-matching targets in Eq. 4 are defined from ground-truth future frames, waypoint annotations, and DA3 camera tokens; the model is trained to denoise these targets and then evaluated on unseen data. No fitted parameter is renamed as a prediction: the loss weights (lambda_vid=1, lambda_act=5, lambda_cam=0.01) and the default two-step sampler are fixed, conventional choices, and Table 4 reports accuracy across sampling steps rather than selecting the best result for the main comparison. The camera-token denoising target is extracted from ground-truth future frames during training, but the paper explicitly labels clean camera-token conditioning an oracle setting (Table 2) and uses denoised camera tokens at inference, so this is auxiliary supervision, not a circular shortcut. I found no load-bearing self-citation: references to Wan2.1, DepthAnything3/DA3, flow matching, and prior navigation methods are external prior work, none by the present authors, and none is invoked as a uniqueness theorem or as the sole justification for a core claim. The reader-flagged concern about whether baselines are zero-shot or fine-tuned is an evaluation-fairness and verifiability issue, not circularity: even if adaptation explains part of the margin, UniNav's reported ATE values do not reduce analytically to the baseline values or to the training objective. The paper also acknowledges its own scope in 'Limitations and Future Work' (offline evaluation, short-horizon prediction) and states in Section 4.2 that the main comparison uses only trajectory-labeled data, which further separates the central results from the video-only training contribution. No circular step meets the evidentiary standard of the review.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central benchmark claim depends on five unproved assumptions: the flow-matching objective is a valid generative formulation, the Wan2.1 video prior transfers to navigation, DA3 camera tokens carry usable geometry, asymmetric attention makes image-token removal safe, and video-only data helps waypoint prediction through geometry supervision. The first is standard math; the last four are domain assumptions supported only by this paper's internal ablations, with the mixed-training effect being small.

free parameters (6)
  • lambda_vid = 1
    Hand-chosen weight for the future-video flow-matching loss; no sweep is reported (Sec 4.1, Eq 8).
  • lambda_traj = 5
    Hand-chosen weight for the trajectory loss, called lambda_act in the implementation details; no sweep reported (Sec 4.1).
  • lambda_cam = 0.01
    Hand-chosen weight for the camera-token denoising loss; no sweep reported (Sec 4.1).
  • num_inference_steps = 2
    Default number of denoising steps used for all main results; chosen as a speed-accuracy tradeoff (Sec 4.1).
  • history_frames_C = 3
    Number of past frames fed as context (Sec 4.1).
  • future_horizon_K = 6
    Number of future frames/waypoints predicted (Sec 4.1).
assumptions (5)
  • standard math Flow matching (Lipman et al. 2022) provides a valid objective for generative modeling of all three token modalities.
    Used throughout Sec 3.3, Eq (4)-(8), to define the denoising targets; the paper inherits the convergence and training properties of flow matching without proof.
  • domain assumption The pretrained Wan2.1-1.3B video diffusion transformer and its frozen VAE provide a useful prior for egocentric navigation futures.
    The model is initialized from Wan2.1 and the VAE is frozen (Sec 4.1); transfer to navigation is justified only by the empirical gains.
  • domain assumption The frozen geometry foundation model (Depth Anything 3, Lin et al. 2025) emits camera tokens that capture metric 3D structure useful for waypoint and video prediction.
    Eq (2) in Sec 3.2 uses these tokens as supervision; the only support is the internal ablation (Table 2), not external evidence.
  • domain assumption Asymmetric attention lets trajectory tokens ignore future-video tokens during training, so removing image tokens at inference does not break the policy.
    Sec 3.4 relies on Fast-WAM-style asymmetric attention (Yuan et al. 2026) without analysis of residual interference.
  • domain assumption Geometry-supervised video-only data transfers to waypoint prediction.
    Sec 3.3 and Table 3 assume that camera-token denoising creates a shared representation where unlabeled video helps actions; the reported effect is small (0.012 ATE).
invented entities (1)
  • camera-geometry tokens
    purpose: An auxiliary denoising target injected into the shared token sequence to give the model geometric supervision from a frozen depth-based foundation model.
    The paper introduces this token type and its denoising objective (Sec 3.2). No external benchmark or falsifiable prediction supports that these tokens encode usable geometry; the only support is the paper's own ablations (Tables 2 and 3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniNav: A Unified World-Action Diffusion Model for Visual Navigation." pith.science (2026). https://pith.science/paper/KVERJQKA

@misc{pith2026260803244,
  author       = {Pith},
  title        = {Pith review of: UniNav: A Unified World-Action Diffusion Model for Visual Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KVERJQKA}},
  note         = {Machine review of arXiv:2608.03244}
}
read the original abstract

Image-goal visual navigation is a fundamental capability for embodied agents. Existing navigation policies efficiently predict waypoint trajectories but lack visual foresight, while navigation world models can anticipate future observations but often require costly planning rollouts. We present UniNav, a unified world-action model that generates future visual observations and continuous waypoint trajectories through a single diffusion process. Given history frames and a goal image, UniNav jointly denoises visual and waypoint tokens within a single transformer, unifying future prediction and action generation in a shared framework. To improve spatial grounding, we incorporate geometry-aware camera tokens. We also train on both trajectory-labeled navigation data and video-only data, enabling the model to benefit from diverse videos without waypoint annotations. Based on this unified framework, we introduce two variants: UniNav-Full jointly predicts interpretable future observations and their corresponding trajectories, while UniNav-Fast removes future-image tokens at inference for efficient trajectory prediction. Experiments on navigation benchmarks show that UniNav outperforms the strongest baseline in ATE across all datasets. With one-step inference, UniNav-Fast achieves a latency of 0.1s without a substantial accuracy drop. Code will be released.

Figures

Figures reproduced from arXiv: 2608.03244 by the authors.

Figure 1
Figure 1. Comparison of visual navigation paradigms. (a) Policy-only methods directly map the observation history and goal image to a waypoint trajectory, without explicitly mod￾eling future observations. (b) Action-conditioned navigation world models generate visual rollouts for candidate action sequences and select actions by comparing the imagined out￾comes with the goal. (c) Our unified world-action diffusion model jointl… view at source ↗
Figure 2
Figure 2. Overview of UniNav training and inference. During training, visual tokens from a pretrained VAE, camera￾geometry tokens extracted by a frozen geometry model, trajectory tokens, history and goal tokens, and learnable register tokens are jointly processed by the world-action model. The model jointly denoises future-video, camera, and trajectory tokens using the corresponding objectives Lvid, Lcam, and Ltraj. At infere… view at source ↗
Figure 3
Figure 3. Qualitative joint prediction of future observations and waypoint trajectories. The predicted trajectories are shown [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative future-frame predictions on video-only datasets trained without waypoint annotations. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 19 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 69 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2017 , eprint=

    Attention Is All You Need , author=. 2017 , eprint=

    2017

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Nomad: Goal masked diffusion policies for navigation and exploration , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  5. [13]

    arXiv preprint arXiv:2306.14846 , year=

    ViNT: A foundation model for visual navigation , author=. arXiv preprint arXiv:2306.14846 , year=

    arXiv

  6. [14]

    2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Gnm: A general navigation model to drive any robot , author=. 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2023 , organization=

    2023

  7. [15]

    arXiv preprint arXiv:2603.05438 , year=

    Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model , author=. arXiv preprint arXiv:2603.05438 , year=

  8. [16]

    arXiv preprint arXiv:2603.09241 , year=

    RAE-NWM: Navigation World Model in Dense Visual Representation Space , author=. arXiv preprint arXiv:2603.09241 , year=

    arXiv

  9. [17]

    arXiv preprint arXiv:2505.08712 , year=

    Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance , author=. arXiv preprint arXiv:2505.08712 , year=

  10. [18]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Navigation world models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  11. [19]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Navidiffusor: Cost-guided diffusion model for visual navigation , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  12. [20]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Flownav: Combining flow matching and depth priors for efficient navigation , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  13. [21]

    arXiv preprint arXiv:2512.19629 , year=

    Logoplanner: Localization grounded navigation policy with metric-aware visual geometry , author=. arXiv preprint arXiv:2512.19629 , year=

  14. [22]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Beings: Bayesian embodied image-goal navigation with gaussian splatting , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  15. [23]

    arXiv preprint arXiv:2602.09765 , year=

    NavDreamer: Video Models as Zero-Shot 3D Navigators , author=. arXiv preprint arXiv:2602.09765 , year=

  16. [24]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    UrbanNav: Learning Language-Guided Embodied Urban Navigation from Web-Scale Human Trajectories , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  17. [25]

    arXiv preprint arXiv:2601.00844 , year=

    Value-guided action planning with JEPA world models , author=. arXiv preprint arXiv:2601.00844 , year=

  18. [26]

    arXiv preprint arXiv:2511.11011 , year=

    Efficient Image-Goal Navigation with Representative Latent World Model , author=. arXiv preprint arXiv:2511.11011 , year=

  19. [27]

    arXiv preprint arXiv:2602.15922 , year=

    World action models are zero-shot policies , author=. arXiv preprint arXiv:2602.15922 , year=

    arXiv

  20. [28]

    arXiv preprint arXiv:2601.16163 , year=

    Cosmos policy: Fine-tuning video models for visuomotor control and planning , author=. arXiv preprint arXiv:2601.16163 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2601.21998 , year=

    Causal World Modeling for Robot Control , author=. arXiv preprint arXiv:2601.21998 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2603.16666 , year=

    Fast-WAM: Do World Action Models Need Test-time Future Imagination? , author=. arXiv preprint arXiv:2603.16666 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2603.17240 , year=

    GigaWorld-Policy: An Efficient Action-Centered World--Action Model , author=. arXiv preprint arXiv:2603.17240 , year=

  24. [32]

    arXiv preprint arXiv:2604.11135 , year=

    AIM: Intent-Aware Unified world action Modeling with Spatial Value Maps , author=. arXiv preprint arXiv:2604.11135 , year=

    arXiv

  25. [33]

    arXiv preprint arXiv:2604.03181 , year=

    Multi-View Video Diffusion Policy: A 3D Spatio-Temporal-Aware Video Action Model , author=. arXiv preprint arXiv:2604.03181 , year=

    arXiv

  26. [34]

    arXiv preprint arXiv:2503.00200 , year=

    Unified video action model , author=. arXiv preprint arXiv:2503.00200 , year=

    arXiv

  27. [35]

    arXiv preprint arXiv:2504.02792 , year=

    Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets , author=. arXiv preprint arXiv:2504.02792 , year=

    arXiv

  28. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Motus: A unified latent action world model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [37]

    arXiv preprint arXiv:2603.07799 , year=

    MWM: Mobile World Models for Action-Conditioned Consistent Prediction , author=. arXiv preprint arXiv:2603.07799 , year=

  30. [38]

    arXiv preprint arXiv:2503.20314 , volume=

    Wan: Open and advanced large-scale video generative models , author=. arXiv preprint arXiv:2503.20314 , volume=

    arXiv

  31. [39]

    arXiv preprint arXiv:2104.05859 , year=

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

    arXiv

  32. [40]

    IEEE Robotics and Automation Letters , volume=

    Sacson: Scalable autonomous control for social navigation , author=. IEEE Robotics and Automation Letters , volume=. 2023 , publisher=

    2023

  33. [41]

    IEEE Robotics and Automation Letters , volume=

    Deep visual mpc-policy learning for navigation , author=. IEEE Robotics and Automation Letters , volume=. 2019 , publisher=

    2019

  34. [42]

    IEEE Robotics and Automation Letters , volume=

    Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  35. [43]

    arXiv preprint arXiv:2511.10647 , year=

    Depth anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

    arXiv

  36. [44]

    arXiv preprint arXiv:2210.02747 , year=

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    arXiv

  37. [45]

    arXiv preprint arXiv:2512.15692 , year=

    mimic-video: Video-action models for generalizable robot control beyond vlas , author=. arXiv preprint arXiv:2512.15692 , year=

    arXiv

  38. [46]

    arXiv preprint arXiv:2508.05635 , year=

    Genie envisioner: A unified world foundation platform for robotic manipulation , author=. arXiv preprint arXiv:2508.05635 , year=

    arXiv

  39. [47]

    arXiv preprint arXiv:2604.26694 , year=

    Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising , author=. arXiv preprint arXiv:2604.26694 , year=

    arXiv

  40. [48]

    2026 , eprint=

    Light-WAM: Efficient World Action Models with State-Fusion Action Decoding , author=. 2026 , eprint=

    2026

  41. [49]

    arXiv preprint arXiv:2606.04907 , year=

    WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation , author=. arXiv preprint arXiv:2606.04907 , year=

    arXiv

  42. [50]

    arXiv preprint arXiv:2510.08713 , year=

    Unified world models: Memory-augmented planning and foresight for visual navigation , author=. arXiv preprint arXiv:2510.08713 , year=

  43. [51]

    International conference on learning representations , volume=

    Vision transformers need registers , author=. International conference on learning representations , volume=

  44. [52]

    arXiv preprint arXiv:2605.15195 , year=

    Wang, Jianyuan and Chen, Minghao and Zhang, Shangzhan and Karaev, Nikita and Sch. arXiv preprint arXiv:2605.15195 , year=

    arXiv

  45. [53]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    arXiv

  46. [54]

    The International Journal of Robotics Research , volume=

    Diffusion policy: Visuomotor policy learning via action diffusion , author=. The International Journal of Robotics Research , volume=. 2025 , publisher=

    2025

  47. [55]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Prior does matter: Visual navigation via denoising diffusion bridge models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  48. [56]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Scannet: Richly-annotated 3d reconstructions of indoor scenes , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  49. [57]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  50. [58]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Citywalker: Learning embodied urban navigation from web-scale videos , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  51. [59]

    arXiv preprint arXiv:2402.14281 , year=

    A landmark-aware visual navigation dataset , author=. arXiv preprint arXiv:2402.14281 , year=

    arXiv

  52. [60]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Tartanground: A large-scale dataset for ground robot perception and navigation , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  53. [61]

    Evaluating egomotion and structure-from-motion approaches using the TUM RGB-D benchmark , author=. Proc. of the Workshop on Color-Depth Camera Fusion in Robotics at the IEEE/RJS International Conference on Intelligent Robot Systems (IROS) , volume=

  54. [62]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    arXiv

  55. [63]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  56. [64]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  57. [65]

    arXiv preprint arXiv:2508.00795 , year=

    Video generators are robot policies , author=. arXiv preprint arXiv:2508.00795 , year=

    arXiv

  58. [66]

    Advances in neural information processing systems , volume=

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras , author=. Advances in neural information processing systems , volume=

  59. [67]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  60. [68]

    Neurocomputing , volume=

    Roformer: Enhanced transformer with rotary position embedding , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  61. [69]

    arXiv preprint arXiv:2606.15869 , year=

    Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation , author=. arXiv preprint arXiv:2606.15869 , year=

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.