Pith. sign in

REVIEW 4 major objections 5 minor 22 references

A training-only depth prior makes a compact JEPA world model more transferable on real outdoor robot video: visual odometry error drops 33%, surprise separation roughly doubles in-domain and triples out of domain, at zero inference overhead

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 05:21 UTC pith:Q6YGRF64

load-bearing objection A clearly written proof-of-concept that depth regularization can help compact JEPAs on real outdoor data, but the two simultaneous interventions leave the headline claim untested. the 4 major comments →

arxiv 2607.16314 v1 pith:Q6YGRF64 submitted 2026-07-15 cs.CV cs.LGcs.RO

Depth-Regularized JEPA World Models Learn More Transferable Representations from Real Outdoor Robot Data

Usman M. Khan This is my paper
classification cs.CV cs.LGcs.RO
keywords JEPAworld modelsdepth regularizationgeometric priorvisual odometry probeout-of-domain transferrobot videolatent dynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a compact JEPA-style world model, trained directly on real agricultural-robot video, becomes substantially more useful and more transferable when a depth-alignment loss is added during training only. The depth signal acts as a geometric teacher, organizing the latent diversity that the SIGReg regularizer maximizes toward physically meaningful directions rather than leaving it isotropic. On the paper's own benchmarks, this yields a 33% reduction in frozen-embedding visual odometry error, roughly a doubling of surprise-score separation on in-domain data, and a nearly three-fold increase on the out-of-domain TartanGround benchmark, with multi-step rollout fidelity improving under domain shift as the horizon grows. The paper also reports that the geometric prior sharpens expectations about lighting and appearance, which have no direct 3D content. If the claim holds, it gives a cheap, inference-free recipe for injecting physical structure into RGB-only world models for real outdoor robots.

Core claim

The paper's central claim: a stop-gradient depth-alignment loss added to LeWM's SIGReg-regularized JEPA objective yields the highest-entropy latent representation consistent with scene geometry, and this representation transfers better than a purely isotropic one. In evaluation, the 18M-parameter model lowers VO probe MSE from 0.0022 to 0.0015, raises surprise separation from 0.067 to 0.120 in-domain and from 0.035 to 0.099 on TartanGround, and improves OOD rollout goal similarity at h=20 from 0.458 to 0.537. Gains also appear on brightness and color-swap perturbations, suggesting geometry-grounded representations impose consistency expectations on appearance. Depth and added predictor capac

What carries the argument

The core mechanism is the depth-alignment term added to the LeWM objective: L = L_pred + 0.09 L_sigreg + 0.1 L_depth, with L_depth = 1 − cos(f_θ(x_RGB), sg[f_θ(x_depth)]). During training, paired RGB and stereo depth frames pass through the same encoder–projector; the depth branch uses stop-gradient, so only the RGB branch is pushed toward the geometry-informed embedding. SIGReg continues to prevent collapse by regularizing latents toward isotropic Gaussians, so the combined objective targets 'the highest-entropy latent representation consistent with scene geometry.' To keep the predictor from exerting back-pressure that would flatten the richer latent structure, the paper adds rank-16 RepLi

Load-bearing premise

The claim that depth regularization improves representation quality rests on the assumption that the reported gains come from the depth-alignment loss itself rather than from the simultaneously added predictor overparameterization, since the paper changes both at once and defers the capacity ablation.

What would settle it

Train the same LeWM base under four conditions—vanilla, overparameterization only, depth-loss only, and both—and compare VO probe MSE and TartanGround surprise separation. If the overparameterization-only condition reproduces the bulk of the gains (or the depth-only condition does not), the depth-as-geometric-prior explanation is refuted. Complementarily, if a non-geometric auxiliary matching loss produces the same brightness/color-swap separation improvements, the physical-scaffold interpretation is weakened.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Compact JEPA world models can be trained end-to-end on real outdoor robot video (~31K frames, single platform), a regime not previously reported for this architecture family.
  • The training-only depth prior provides a way to inject physical structure into RGB-only world models without task labels, pretrained encoders, or any inference-time depth input.
  • The pattern of slightly lower in-domain but markedly better out-of-domain rollout similarity, with gains growing with horizon, implies the depth-shaped latent dynamics are more transferable and more resistant to compounding error.
  • Improved surprise separation on non-geometric perturbations suggests that geometry-grounded representations naturally enforce appearance-consistency expectations, which could strengthen anomaly detection.
  • Because the added predictor capacity is merged away after training, the deployed model is architecturally identical to the baseline, so the benefit is entirely paid for at training time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A factorial ablation—depth loss only, predictor overparameterization only, and both together—would settle whether the gains are attributable to the geometric prior or to added capacity; the paper defers this, so the two explanations remain entangled.
  • A non-geometric auxiliary control (e.g., aligning RGB embeddings to a fixed random target or to a second RGB view) could test whether the brightness/color-swap gains reflect geometry-grounded physical expectations or merely any well-chosen auxiliary objective that tightens the latent space.
  • The PCA mid-rank subspace-alignment result suggests a practical transferability diagnostic: if the top-k (k≈5–50) principal subspaces per domain become more shared under depth regularization, this metric could guide model selection or early stopping when depth supervision is available.
  • The offline rollout results are latent-similarity metrics, not closed-loop robot performance; testing the depth-regularized latent space in a real planner would show whether the transfer gains persist under control.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a depth-regularized JEPA world model built on LeWorldModel (LeWM), adding a training-only depth alignment loss (L_depth, Eq. 2) to the base SIGReg-regularized latent prediction objective, together with predictor-side overparameterization via rank-16 RepLinear branches. An 18M-parameter model is trained on ~31K frames of real agricultural robot video and evaluated against vanilla LeWM with frozen-embedding visual odometry probes, predictor-based surprise detection, and multi-step latent rollout on in-domain and TartanGround out-of-domain data. The headline claims are a 33% VO probe MSE reduction (0.0022 to 0.0015), substantially increased surprise separation (0.067 to 0.120 in-domain; 0.035 to 0.099 OOD), and improved OOD rollout goal similarity at h=20 (0.458 to 0.537), with the interpretation that depth provides a geometric prior that improves representation quality and transferability.

Significance. If the attribution to depth regularization can be established, the result is practically valuable: it shows a compact, end-to-end trainable JEPA can be trained directly on noisy outdoor robot data, and a cheap training-time depth signal (no depth at inference, no task labels, no frozen teacher) improves geometric and transferable properties of the latent space. The paper is honest about several limitations and includes useful diagnostic analyses (principal subspace overlap, limitation of rollout similarity comparisons). However, the central claim is currently underdetermined because the method changes two components at once and the authors explicitly defer the capacity ablation to future work. The significance will rise considerably if the depth-only effect is isolated with appropriate controls, uncertainty quantification, and scale-normalized surprise metrics.

major comments (4)
  1. [§3.1, Eq. (2), §5, Table 1] The paper changes two training-time components at once: the depth alignment loss (Eq. 2) and predictor-side overparameterization (rank-16 RepLinear branches, §3.1). Section 5 states that isolating the capacity contribution is left to future work. Since every headline gain in Table 1 is measured for the joint intervention, the evidence cannot distinguish the depth-prior explanation from an enlarged-predictor explanation (or an interaction). A predictor with more capacity can by itself reduce pressure on the encoder to keep latents overly simple, which could improve surprise margins and OOD rollout without any geometric supervision. A capacity-only ablation (and ideally a depth-only ablation without overparameterization) is required to support the title claim that depth regularization is responsible for the reported gains.
  2. [§3.4, §4.2, Table 1] The surprise separation metric is a difference of mean squared predictor errors in latent space. No normalization is reported for the scale or covariance of the latent embeddings, and SIGReg and the depth term could change the average latent norm or per-dimension variance between the two models. If the depth-regularized embeddings have larger scale, the separation difference (0.067→0.120 in-domain; 0.035→0.099 OOD) could partly reflect a mechanical scaling effect rather than sharper predictive expectations. The paper should report latent norm/variance statistics or use a scale-invariant separation metric (e.g., normalized by per-dimension variance of the plausibles distribution), and ideally show distributions rather than only means.
  3. [§3.1, §4.1, Table 1] The central quantitative comparison relies on a single representative run. The text says 'Repeated training runs yielded consistent results; we report a representative run,' but no standard deviations, confidence intervals, or per-seed results are provided anywhere, including Table 1. Without uncertainty estimates, the reader cannot tell whether the VO gap (0.0022 vs 0.0015), the surprise separation gains, or the OOD rollout improvements are robust. At minimum, means and standard deviations over at least 3 seeds are needed; the in-domain rollout degradation (0.903→0.890) also needs a paired comparison to assess whether the trade-off is significant.
  4. [§4.4, Table 1] The OOD rollout claims rest on a comparison where the depth model is worse in-domain at h=10 (goal 0.903→0.890; trajectory 0.967→0.956) but better OOD. The authors interpret this as a transfer trade-off, which is plausible. However, given the absence of error bars (Major Comment 3), the OOD advantage could be within run-to-run noise. In addition, the OOD setting changes platform, environment, and sim-vs-real simultaneously (as stated in §3.3), so any single benchmark cannot identify which shift matters. A per-episode breakdown or at least a paired significance test across episodes would strengthen this part of the evidence.
minor comments (5)
  1. [Abstract / Title] The abstract and title attribute the gains to depth regularization, but the method also includes overparameterization. Please qualify the claim (e.g., 'depth-regularized training with predictor overparameterization') until the ablation is available.
  2. [§4.3] The paper correctly notes that a non-geometric auxiliary control would be needed to distinguish 'depth as geometric scaffold' from 'any auxiliary objective sharpens prediction-error margins.' This caveat should appear in the abstract or conclusion if the lighting gains are highlighted as a headline result.
  3. [§3.3 / Table 1] The description of depth truncation ('truncated at a distance of 1 meter') is ambiguous: does it mean depths are clipped at 1m, or only depths within 1m are used? Please clarify the preprocessing.
  4. [Figure 3] The figure caption mentions 'silhouette 0.170→0.137' and 'centroid 9.96σ→8.96σ' but the main text does not define the silhouette computation or the σ units. Please define these metrics explicitly.
  5. [§3.1] Minor wording: 'the same encoder–projector network processes both modalities' with a stop-gradient on depth — it would help to state whether the depth branch uses the same input resolution and augmentation as RGB, and whether any modality-specific normalization is applied.

Circularity Check

0 steps flagged

No circularity found: depth alignment is genuine auxiliary supervision with stop-gradient; evaluation is on held-out/OOD frozen embeddings; the capacity confound is explicitly flagged as an ablation gap, not a fitted-input prediction.

full rationale

The derivation chain is empirical, not definitional. L_depth (Eq. 2) is a cosine alignment between RGB and stop-gradient depth embeddings; it is not defined in terms of the outcome metrics (VO probe MSE, surprise separation, rollout similarity), and none of those metrics are fitted or used to set the depth weight (0.1 chosen heuristically, Sec. 3.2). Evaluation is on held-out frames and the external TartanGround benchmark with frozen embeddings, so the headline numbers are independent measurements, not algebraic consequences of the training objective. The method adds two interventions simultaneously — the depth loss and rank-16 RepLinear predictor branches (Sec. 3.1) — and Sec. 5 states: 'Isolating its contribution with a dedicated capacity ablation is left to future work.' This is a causal attribution/ablation gap: the reported gains could in principle be driven by capacity rather than depth, but that underdetermination is not a circular reduction; the paper never renames a fitted parameter as a prediction and never defines the method in terms of the evaluation. Citations to LeWM [12], RepVGG [4], MobileOne [19], and Kabra et al. [10] are external prior work by other authors; none is a self-citation used to forbid alternatives or justify a uniqueness claim. The rollout and surprise metrics are self-consistency measures in the model's own latent space, but they are applied symmetrically to both variants, and Sec. 4.4 itself cautions that rollout similarity 'partly rewards smooth or slowly-varying latents,' so the comparison is a soundness caveat, not a circularity. No load-bearing step reduces by construction to its inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on two tuned/added components (depth term, overparameterization) that are not separately ablated; both are training-only privileges. The evaluation metrics (surprise separation, rollout similarity) are not normalized across models, so cross-model comparisons are only partially controlled.

free parameters (5)
  • depth loss weight (lambda_depth) = 0.1
    Set heuristically 'to be of comparable magnitude to the SIGReg term' (Section 3.2, Eq. 1); not tuned or ablated.
  • SIGReg weight (lambda_sigreg) = 0.09
    Taken from LeWM baseline; the paper states the value follows the one used in LeWM (Section 3.2).
  • RepLinear overparameterization rank = 16
    Rank of low-rank branches added to each predictor MLP up/down-projection (Section 3.1); capacity intervention not ablated separately.
  • Depth truncation distance = 1 m
    Depth frames truncated at 1 meter (Section 3.3); determines the geometric signal range.
  • Perturbation magnitudes for surprise probes = not specified
    The magnitudes of teleportation, temporal reversal, brightness jump, and color swap are not reported (Section 3.4); these directly control the measured surprise separation.
axioms (5)
  • domain assumption LeWM's SIGReg regularizer prevents collapse and produces meaningful latent dynamics (Maes et al., 2026).
    The base training recipe and anti-collapse guarantee are taken as given from the cited prior work (Sections 2.2, 3.1).
  • domain assumption Depth is a reliable geometric prior; aligning RGB embeddings to depth embeddings induces useful structure without harming RGB-only inference.
    Central modeling premise; the paper provides no analysis that the stop-gradient alignment produces the intended geometry rather than, e.g., a texture-based shortcut (Eq. 2).
  • domain assumption A frozen-embedding linear probe from a world model's latent to ego-motion measures representation quality.
    Evaluation protocol (Section 3.3, 4.1) assumes linear decodability reflects downstream utility.
  • domain assumption Surprise separation (mean perturbed minus mean plausible predictor error) is comparable across models without normalizing latent scales.
    Section 4.2 reports separation as an absolute difference; if the two models' latents have different scales, the margins are not directly comparable.
  • standard math Structural re-parameterization of linear branches is algebraically exact, so the merged predictor matches the training-time predictor for the same inputs.
    RepVGG/MobileOne-style merging (refs [4,19]) assumes linearity of the added branches, which holds here (Section 3.1).

pith-pipeline@v1.3.0-alltime-deepseek · 9089 in / 12060 out tokens · 104440 ms · 2026-08-02T05:21:40.415309+00:00 · methodology

0 comments
read the original abstract

World models, especially based on JEPA architectures, have been shown to learn robust dynamics of various environments. However, learning from visually complex real-world data remains a challenge, especially in unpredictable outdoor environments. We introduce depth as a geometric prior during training in learning more robust latent dynamics directly from robot video data and handling visual complexity. This combines depth supervision with an isotropy-inducing latent regularizer (SIGReg), maximizing task-agnostic latent diversity while constraining how that diversity is organized, with the combined objective targeting the highest-entropy representation consistent with scene geometry. To satisfy this greater complexity without increasing inference time, we also add training-only overparameterization. Training an 18M-parameter model on video from a real agricultural robot, we evaluate with frozen-representation visual odometry probes, predictor-based surprise detection, and multi-step latent rollout fidelity. Compared to the baseline LeWM, our method lowers visual odometry probe error by 33%, substantially increases surprise-score separation both in-domain and on the out-of-domain TartanGround benchmark, and improves multi-step rollout fidelity under domain shift, with gains that grow with rollout horizon. Notably, we also see improvements in surprise-score separation on physics understanding that is not directly tied to 3D geometry, such as lighting and shadows. These results show that a lightweight training-time geometric prior makes a compact JEPA world model more useful and more transferable on real outdoor data with strong underlying representations, without adding inference overhead. Our work suggests that depth as a physically grounded prior can enhance world model generalization on a variety of tasks.

Figures

Figures reproduced from arXiv: 2607.16314 by Usman M. Khan.

Figure 1
Figure 1. Figure 1: Method overview. RGB history is encoded by the shared encoder–projector [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Example data sample collected from the Aigen Element robot. (a) An RGB image from [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-domain latent geometry. Top: principal-subspace overlap and mean principal angle vs. k (PCA fit per domain per model, full 192-D space). The depth-regularized model is more cross-domain aligned across the mid-rank band k ≈ 5–50; large-k overlap is dimensionally forced (floor (2k − d)/k), so the informative regime is k ≤ 96. Bottom: 2-D PCA projections of Aigen validation (blue) and TartanGround (oran… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 17 linked inside Pith

  1. [1]

    Self-supervised learning from images with a joint- embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint- embedding predictive architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. arXiv:2301.08243

    Pith/arXiv arXiv 2023

  2. [2]

    V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    Mahmoud Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Revisiting feature prediction for learning visual repre- sentations from video.arXiv preprint arXiv:2404.08471, 2024

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual repre- sentations from video.arXiv preprint arXiv:2404.08471, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Repvgg: Making vgg-style convnets great again

    Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. arXiv:2101.03697

    Pith/arXiv arXiv 2021

  5. [5]

    seq-jepa: Autoregressive predictive learning of invariant-equivariant world models

    Hafez Ghaemi et al. seq-jepa: Autoregressive predictive learning of invariant-equivariant world models. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. arXiv:2505.03176

    arXiv 2025

  6. [6]

    World models.arXiv preprint arXiv:1803.10122, 2018

    David Ha and Jürgen Schmidhuber. World models.arXiv preprint arXiv:1803.10122, 2018

    Pith/arXiv arXiv 2018

  7. [7]

    Dream to control: Learning behaviors by latent imagination.arXiv preprint arXiv:1912.01603, 2019

    Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination.arXiv preprint arXiv:1912.01603, 2019

    Pith/arXiv arXiv 1912

  8. [8]

    Learning latent dynamics for planning from pixels

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. InProceedings of the International Conference on Machine Learning (ICML), 2019. arXiv:1811.04551. 12

    Pith/arXiv arXiv 2019

  9. [9]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  10. [10]

    Hudson, Ye Xia, Skanda Koppula, Andre Araujo, Joao Carreira, and Niloy J

    Rishabh Kabra, Maks Ovsjanikov, Drew A. Hudson, Ye Xia, Skanda Koppula, Andre Araujo, Joao Carreira, and Niloy J. Mitra. A mixed diet makes dino an omnivorous vision encoder. arXiv preprint arXiv:2602.24181, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Enhancing end-to-end autonomous driving with latent world model.arXiv preprint arXiv:2406.08481, 2024

    Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model.arXiv preprint arXiv:2406.08481, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Leworld- model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. Leworld- model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    MaximeOquab, TimothéeDarcet, ThéoMoutakanni, HuyVo, MarcSzafraniec, VasilKhalidov, PierreFernandez, DanielHaziza, FranciscoMassa, AlaaeldinEl-Nouby, etal. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Tartanground: A large-scale dataset for ground robot perception and navigation

    Manthan Patel, Fan Yang, Yuheng Qiu, Cesar Cadena, Sebastian Scherer, Marco Hutter, and Wenshan Wang. Tartanground: A large-scale dataset for ground robot perception and navigation. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),

  15. [15]

    Learning from reward-free offline data: A case for planning with latent dynamics models.arXiv preprint arXiv:2502.14819, 2025

    Vlad Sobal et al. Learning from reward-free offline data: A case for planning with latent dynamics models.arXiv preprint arXiv:2502.14819, 2025

    arXiv 2025

  16. [16]

    Understanding self-supervised learning dynamics without contrastive pairs

    Yuandong Tian, Xinlei Chen, and Surya Ganguli. Understanding self-supervised learning dynamics without contrastive pairs. InProceedings of the International Conference on Machine Learning (ICML), 2021. arXiv:2102.06810

    Pith/arXiv arXiv 2021

  17. [17]

    Oliehoek, and Max Welling

    Elise van der Pol, Daniel Worrall, Herke van Hoof, Frans A. Oliehoek, and Max Welling. Mdp homomorphic networks: Group symmetries in reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2020. arXiv:2006.16908

    Pith/arXiv arXiv 2020

  18. [18]

    A new learning paradigm: Learning using privileged information.Neural Networks, 22(5–6):544–557, 2009

    Vladimir Vapnik and Akshay Vashist. A new learning paradigm: Learning using privileged information.Neural Networks, 22(5–6):544–557, 2009

    2009

  19. [19]

    Mobileone: An improved one millisecond mobile backbone

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. arXiv:2206.04040

    Pith/arXiv arXiv 2023

  20. [20]

    World4drive: End-to-endautonomousdrivingviaintention-awarephysical latent world model.arXiv preprint arXiv:2507.00603, 2025

    YupengZhengetal. World4drive: End-to-endautonomousdrivingviaintention-awarephysical latent world model.arXiv preprint arXiv:2507.00603, 2025

    Pith/arXiv arXiv 2025

  21. [21]

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

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning. InProceedings of the International Con- ference on Machine Learning (ICML), 2025. arXiv:2411.04983

    Pith/arXiv arXiv 2025

  22. [22]

    Ad-l-jepa: Self- supervised spatial world models with joint embedding predictive architecture for autonomous driving with lidar data

    Haoran Zhu, Zhenyuan Dong, Klodian Topollai, and Anna Choromanska. Ad-l-jepa: Self- supervised spatial world models with joint embedding predictive architecture for autonomous driving with lidar data. InProceedings of the AAAI Conference on Artificial Intelligence, 2026. arXiv:2501.04969. 13

    arXiv 2026