Pith. sign in

REVIEW 4 major objections 6 minor 20 references

JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read JEPA-WAM proposes that a frozen V-JEPA space plus a shared predictor for transition prediction and action generation yields robust out-of-distribution robot control, reaching 79.2% on LIBERO-Plus without robot-policy pretraining and…

desk verdict A genuinely novel and well-ablated latent WAM architecture whose central LIBERO-Plus claim rests on a 2.2-point gain without variance; still worth sending to review. read the letter →

arxiv 2608.09381 v1 pith:MNIANWN2 submitted 2026-08-10 cs.RO

Yihan Lin , Jiawei He , Shifeng Bao , Chen Zhao , Yang Li , Xiaobo Wang , Yan Wang , Cheng Chi
show 1 more author
Jing Zhang
This is my paper · ORCID
classification cs.RO
keywords vision-language-actionmodelsworldactionjoint-embeddingpredictivearchitecturelatentout-of-distributiongeneralizationrobotmanipulationflowmatchingV-JEPA
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 sets out to establish that a robot policy can become more robust to visual and spatial changes by predicting, in latent space, how the scene will change between the current and a slightly later observation, without ever generating an image of the future. The proposed model, JEPA-WAM, builds on a frozen V-JEPA encoder and trains a shared predictor that both predicts a dense, per-image-region joint current-future target and produces representations used to generate continuous actions. The core claim is that this transition supervision directly shapes the same backbone that conditions action generation, so the policy learns task-shared visual temporal structure. On LIBERO-Plus the model reaches 79.2% average success without large-scale robot-policy pretraining, and the same supervision raises a pretrained pi0.5 policy to 86.3%, the best overall result, with similar gains on RoboTwin 2.0 and real-world bimanual tasks.

What carries the argument

The load-bearing mechanism is the spatially structured joint current-future target $Y_{t,t+\delta} = \mathrm{Concat}_{v}\,\mathrm{sg}\!\left(E_J(\mathrm{Stack}_{\mathrm{time}}(O_t^v, O_{t+\delta}^v))\right)$, a frozen V-JEPA encoding of the two-frame stack that keeps the same patch grid as the current representation. A shared predictor $F_\theta$ (a Qwen2.5-0.5B language model with LoRA) maps the projected current V-JEPA tokens, the task instruction, and dedicated action placeholders to hidden states; a lightweight prediction head maps the visual-token hidden states back into V-JEPA patch space, where they are supervised by mean patch-wise cosine distance to the joint target. The same forward pass yields dedicated action-conditioning tokens that feed a DiT flow-matching action expert. This coupling is what lets the transition loss update the action-generating backbone directly. The paper also transfers the same target to pretrained VLA policies by adding coarse future tokens, reshaping and upsampling their hidden states to match the target grid.

What would settle it

Run JEPA-WAM and its V-JEPA-only variant (transition prediction removed) with identical seeds, data, and rollouts on LIBERO-Plus; if the average success-rate gap between with and without transition supervision does not reproduce the reported 2.2-point difference (79.2 versus 77.0) with non-overlapping confidence intervals, the central claim that transition supervision drives the out-of-distribution gain would be undercut.

Watch

Extended reading notes

Core claim

The central discovery is that a transition target defined by stacking current and future observations and encoding them together with a frozen V-JEPA encoder, rather than encoding the future alone or subtracting endpoint encodings, carries more accessible temporal-relation information, and that supervising a shared Qwen-based predictor to predict it patch-wise improves the representations used for action conditioning. The paper demonstrates this through LIBERO-Plus ablations (joint target 79.2 versus future-only 77.3 versus endpoint-difference 70.9), frozen-probe experiments showing better temporal-gap decoding and residual trajectory prediction, and successful transfer to a pretrained VLA via auxiliary future tokens. On the paper's own terms, the claim is that latent transition prediction in the V-JEPA space, coupled with action generation through a shared predictor, yields strong generalization under visual and spatial shift while preserving in-distribution performance.

Load-bearing premise

The headline comparisons assume that the published baseline numbers were produced under training and evaluation protocols equivalent enough to the paper's own, including the same benchmark episodes, camera views, and rollout counts, since none of those baselines were re-run in the paper's codebase.

Editorial extensions

If this is right

  • Because the target encoder, prediction head, and transition loss are removed at deployment, the policy keeps the same inference cost as a non-world-model VLA, measured at 85 ms per inference on RoboTwin.
  • The same transition supervision transfers to pretrained VLAs as an auxiliary loss, improving pi0.5 from 84.5% to 86.3% on LIBERO-Plus and from 77.5% to 90.3% on real-world in-distribution tasks without altering its action pathway.
  • Patch-level transition supervision outperforms compressed or convolved targets, indicating that future latent world action models should preserve dense spatial structure rather than pool into global tokens.
  • The frozen V-JEPA representation space itself contributes to robustness: replacing DINOv2+SigLIP with V-JEPA raises LIBERO-Plus success from 73.2% to 77.0% even without transition prediction.
  • Joint current-future encoding beats both future-only targets and explicit endpoint differencing, suggesting that temporal-relation information, not just the future state, is the useful supervisory signal.

Reading between the lines

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

  • The paper's stated limitation suggests a testable extension: conditioning the transition target on language or other task context could close the gap when the same observation leads to different transitions under different instructions; a variant with instruction-conditioned targets should be compared on LIBERO-Plus.
  • Because patch-wise representation changes correlate only moderately with RGB change, the transition signal may encode object relations, contact, and configuration changes beyond low-level motion; an explicit study separating these factors could guide future target design.
  • The shared-predictor design implies that scaling the transition loss weight or using multiple temporal offsets might continue to improve out-of-distribution action conditioning, a prediction that could be checked by sweeping $\lambda_{wm}$ and $\delta$ on LIBERO-Plus.
  • The same joint-target formulation could be applied to other pretrained visual encoders or to action-chunking choices, potentially removing the need for a separate world-model module in existing VLA stacks.
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

4 major / 6 minor

Summary. The paper proposes JEPA-WAM, a latent world action model that couples transition prediction with continuous action generation through a shared predictor. The method operates in a frozen V-JEPA 2.1 representation space, constructs a joint current–future target by stacking observations at times t and t+δ, and supervises patch-level cosine alignment between the predicted and target representations. The same predictor produces action-conditioning representations for a flow-matching action expert. The authors evaluate on LIBERO, LIBERO-Plus, RoboTwin 2.0, and a real bimanual platform, reporting that JEPA-WAM achieves 79.2% on LIBERO-Plus without robot-policy pretraining and that instantiation of the same transition objective in π0.5 improves it to 86.3%, the best overall result. Controlled ablations compare visual encoders, transition targets, spatial supervision, and the placement of transition supervision.

Significance. If the empirical claims hold, the paper makes a meaningful contribution: it shows that a latent world model can provide OOD generalization without video generation, that a frozen V-JEPA space can serve as both representation and target space, and that a shared predictor lets transition supervision directly shape the action-generating backbone. The design is clearly specified, the ablation set is more comprehensive than typical for this area, and the use of a frozen, externally pretrained target avoids the circularity of fitting a self-produced target. The appendices add useful probe analyses and per-rollout real-world records. The main weaknesses are statistical: the central mechanistic claims rest on point estimates without variance or multiple seeds, and the headline comparisons against prior methods rely on published numbers produced under unknown protocols. These issues are fixable within the scope of the manuscript.

major comments (4)
  1. [§4.3, Table 4; §6] The central mechanistic claim that joint current–future transition supervision through the shared predictor improves OOD generalization rests on the comparison between JEPA-WAM (79.2%) and V-JEPA only (77.0%) in Table 4. No confidence intervals, multiple seeds, or per-category episode counts are reported. The advantage is reversed in four of seven categories: Language (68.2 vs. 70.9), Lighting (93.3 vs. 96.7), Noise (83.6 vs. 88.3), and Layout (76.1 vs. 79.3). The paper's own §6 acknowledges that language-agnostic transition targets may be less expressive when the same observation leads to different task-conditional transitions, but the paper does not present this as a trade-off. Please report variance estimates and discuss the category-level pattern; as it stands, the 2.2-point average gain is within plausible run-to-run variance and does not support the mechanistic claim.
  2. [§4.2, Tables 1–3] The headline comparisons ('best result without robot-policy pretraining', 'best overall') rely on baseline success rates taken from prior papers, and the manuscript re-runs none of these baselines under its own protocol. This is especially problematic for Table 3 on RoboTwin 2.0, where methods may differ in task grouping, camera views, training data, and number of evaluation rollouts. Without common-protocol re-evaluation or explicit evidence that the protocols are equivalent (e.g., same episodes, same camera views, same rollout counts), the SOTA claims are not supported.
  3. [§3.4, Tables 2 and 3] The π0.5+JEPA Obj. variant appends 64 learnable future tokens to the VLM prefix, and the reported gains (84.5→86.3 on LIBERO-Plus; 75.4→84.6 Clean on RoboTwin 2.0) could in principle come from the added parameters rather than from the transition supervision. Please include a control with the same future tokens inserted but the transition loss removed or ablated, to isolate the effect of the supervision. This is needed to support the transfer claim in §3.4.
  4. [§4.4, Table 15] The real-world conclusions are based on 10 rollouts per task and setting with normalized task-completion scores. While the per-rollout records in Appendix E are valuable, no uncertainty estimates or statistical tests are reported. The claim of 'strong generalization' would be better supported by bootstrap confidence intervals across rollouts or an explicit statement that the sample size is too small for statistical inference.
minor comments (6)
  1. [Figure 1] The caption states that 'the area represents the model backbone's parameter count', but no legend or scale is provided; please clarify how the reader should interpret the area.
  2. [Appendix A.1] Please clarify whether the visual projector is included in the reported 0.5B parameter count for the main backbone, since the projector (1024→896→896) contributes a non-negligible number of parameters.
  3. [Table 4] For consistency with the other tables, the final row should be labeled 'JEPA-WAM (Ours)' rather than just 'JEPA-WAM'.
  4. [Appendix C.1] The description of the iREPA align. variant says 'per-view 3×3 convolutional transformation and spatial target normalization'; please specify whether the normalization is applied to the prediction, the target, or both.
  5. [Abstract] The abstract states that a project page is available on GitHub, but no URL is provided in the manuscript.
  6. [§4.2] Please report the number of evaluation episodes used for each LIBERO-Plus average so that the reader can assess the reliability of the point estimates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the transition target is an external frozen V-JEPA encoding; transition supervision is an auxiliary loss against detached targets; and the reported gains are empirical results on held-out benchmarks, not outputs of the loss construction.

full rationale

JEPA-WAM's derivation chain is self-contained. The joint current–future target Y_{t,t+δ} (Eq. 2) is computed by a frozen, externally pretrained V-JEPA encoder from stacked current and future observations with stop-gradient; it is not defined in terms of the predictor's parameters or of the action loss. The transition objective L_wm (Eq. 5) is therefore a genuine external supervision signal rather than a renamed fit. The shared-predictor design in Eq. 3 is an architectural claim: the same F_θ produces Q^wm_t for transition prediction and C_t for action conditioning, and the paper tests this coupling with controlled ablations (Table 4) on held-out LIBERO-Plus categories. The headline results on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world rollouts are empirical evaluations against external benchmarks, with no fitted parameter being relabeled as a prediction. Self-citations in related work and implementation details (e.g., Li et al. 2026a, StarVLA) are contextual or engineering citations and are not load-bearing for the central transition-supervision mechanism; no uniqueness theorem or ansatz is imported from the authors' own prior work to force the design. The absence of confidence intervals for the 2.2-point LIBERO-Plus ablation gap and possible cross-paper protocol differences are statistical-validity and comparison-protocol concerns, not evidence that any prediction reduces by construction to its inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method relies on the frozen V-JEPA 2.1 encoder as an external perceptual anchor, on several hyperparameters chosen without sensitivity analysis, and on cross-paper baseline comparability. No new physical entities are postulated; the joint current-future target is a training target, not an independently observable world element.

free parameters (6)
  • world model loss weight lambda_wm = 0.5
    Balances L_act and L_wm in Eq. (7) and Eq. (10); no sensitivity study is reported.
  • target temporal offset delta = 31 for LIBERO, 50 for RoboTwin 2.0
    Defines the current-future pair for the joint target in Section A.1; chosen per benchmark, no sweep is reported.
  • action placeholder token count = 64
    Capacity of the dedicated action-conditioning readout in Section A.1; no ablation on the count is reported.
  • LoRA rank and scaling factor = rank 32, scale 64
    Adapter configuration for Qwen2.5-0.5B in Section A.1; fixed without tuning.
  • flow-matching Beta schedule parameters = alpha=1.5, beta=1.0
    Sampling distribution for flow time tau in Section A.3; fixed choice.
  • transition loss weight for pi0.5 transfer = 0.1 after 1K step warmup
    Auxiliary loss weight in Section A.4; chosen without sensitivity analysis.
assumptions (5)
  • domain assumption Frozen V-JEPA 2.1 dense patch features provide a meaningful space for transition supervision via patch-level cosine distance.
    Used in Eqs. (1) through (5) and Section 3.1; the entire method depends on this representation being informative about visual temporal structure.
  • domain assumption Two-frame joint encoding in V-JEPA 2.1 preserves the same spatial token grid as single-frame encoding, so patch correspondence holds.
    Stated in Section 3.1, relying on the temporal tubelet size of two in V-JEPA 2.1; if this property fails, the patch-level alignment in Eq. (5) is invalid.
  • domain assumption The joint current-future target captures task-shared transition structure that transfers across tasks despite being language-agnostic.
    Central premise in Section 3.1, and acknowledged as a limitation in Section 6 when the same observation yields different transitions under different instructions.
  • domain assumption Published baseline results in Tables 1 through 3 are generated under protocols comparable to the paper's own training and evaluation.
    The SOTA claims rely on external numbers; no baseline is re-run in the authors' codebase.
  • standard math Conditional flow matching with velocity prediction is a valid generative objective for continuous action chunks.
    Adopted from StarVLA and prior flow-matching policies; the paper does not re-derive this objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling." pith.science (2026). https://pith.science/paper/MNIANWN2

@misc{pith2026260809381,
  author       = {Pith},
  title        = {Pith review of: JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MNIANWN2}},
  note         = {Machine review of arXiv:2608.09381}
}
abstract

Robust robot control benefits from explicitly modeling state transitions, but video-generation world action models (WAMs) introduce substantial deployment cost. Existing latent WAMs avoid explicit future generation, but often compress predictive representations or separate predictive modeling from the representations used for action generation. We introduce JEPA-WAM, a latent WAM built in a pretrained V-JEPA space, which couples latent transition prediction with continuous action generation through a shared predictor. JEPA-WAM predicts a spatially structured joint current-future target that captures task-shared visual temporal structure between current and future observations, while preserving dense patch-level correspondence. Through the shared predictor, transition supervision directly shapes the backbone, from which dedicated representations are extracted for action prediction. The same design can also be instantiated in pretrained VLA policies while preserving their original perception and action pathways. On LIBERO-Plus, JEPA-WAM achieves 79.2%, the best result without large-scale robot-policy pretraining, while its pretrained $\pi_{0.5}$ instantiation reaches 86.3%, achieving the best overall performance. Experiments on RoboTwin 2.0 and real-world bimanual manipulation further demonstrate strong generalization under visual and spatial shifts.

Figures

Figures reproduced from arXiv: 2608.09381 by the authors.

Figure 1
Figure 1. Overview and performance of JEPA-WAM. Top: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of latent WAM paradigms. JEPA [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of JEPA-WAM. A frozen V-JEPA encoder constructs a spatially structured joint current–future target for [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Transfer of the proposed transition supervision to [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Real-world evaluation on five bimanual manipulation tasks under in-distribution (ID) and out-of-distribution (OOD) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Temporal-gap decoding with a fixed future observation. Rows denote true gaps and columns denote predicted gaps. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Spatial diagnostic of the joint target. Patch-wise representation changes are compared with image-space changes [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 10 canonical work pages

  1. [2]

    Black,K.;Brown,N.;Driess,D.;Esmail,A.;Equi,M.;Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; et al

    Motus: A unified latent action world model. Black,K.;Brown,N.;Driess,D.;Esmail,A.;Equi,M.;Finn, C.; Fusai, N.; Groom, L.; Hausman, K.; Ichter, B.; et al. 2024.π 0: A Vision-Language-Action Flow Model for Gen- eral Robot Control. arXiv:2410.24164. Cen,J.;Yu,C.;Yuan,H.;Jiang,Y.;Huang,S.;Guo,J.;Li,X.; Song,Y.;Luo,H.;Wang,F.;etal.2025.WorldVLA:Towards Autoreg...

    arXiv 2024

  2. [3]

    arXiv:2510.13626

    LIBERO-Plus: In-Depth Robustness Analysis of Vision-Language-Action Models. arXiv:2510.13626. Jeon, B.; Jeong, Y.; Lee, H.; Cho, M.; and Shin, J

    arXiv

  3. [5]

    0, .25, .50, .75,1 Three-block placement Place all three target blocks on the plate, with partial credit for completed placements

    Task Evaluation objective Recorded scores Bread placement Place the target bread on the plate.0,1 Fruit placement Placethetargetpeachesandbananasontheplatewithpartial credit for completed subgoals. 0, .25, .50, .75,1 Three-block placement Place all three target blocks on the plate, with partial credit for completed placements. 0, .33, .67,1 Three-block st...

    work page

  4. [6]

    Training Vision-Language-Action Models with Dense Embodied Chain-of-Thought Supervision

    OpenVLA: An Open-Source Vision- Language-Action Model. InConference on Robot Learn- ing, 6-9 November 2024, Munich, Germany, volume 270 ofProceedings of Machine Learning Research, 2679–2713. PMLR. Li, H.; Li, G.; Feng, Y.; Zhao, C.; Wang, Z.; Li, Y.; Wei, Q.; Bao, S.; Shen, H.; Zhao, Y.; et al. 2026a. Training Vision- Language-Action Models with Dense Emb...

    work page Pith review arXiv 2024

  5. [7]

    InInternational Conference on Learning Representations

    RDT-1B: A Diffusion Foun- dation Model for Bimanual Manipulation. InInternational Conference on Learning Representations. Luo, H.; Zhang, W.; Feng, Y.; Zheng, S.; Xu, H.; Xu, C.; Xi, Z.; Fu, Y.; and Lu, Z. 2026a. Being-H0.7: A Latent World-Action Model from Egocentric Videos. Luo, J.; Wen, Y.; Bai, Y.; Song, X.; Liu, Y.; and Lin, L. 2026b. RoVLA: Multi-Co...

    arXiv

  6. [8]

    Miao, S.; Feng, N.; Wu, J.; Lin, Y.; He, X.; Li, D.; and Long,M.2026

    Dit4dit: Jointly modeling video dynamics and actions for generalizable robot control. Miao, S.; Feng, N.; Wu, J.; Lin, Y.; He, X.; Li, D.; and Long,M.2026. JEPA-VLA:VideoPredictiveEmbeddingIs Needed for VLA Models. arXiv:2602.11832. Mur-Labadia, L.; Muckley, M.; Bar, A.; Assran, M.; Sinha, K.; Rabbat, M.; LeCun, Y.; Ballas, N.; and Bardes, A

    arXiv 2026

  7. [9]

    Physical Intelligence; Black, K.; Brown, N.; Darpinian, J.; Dhabalia,K.;Driess,D.;Esmail,A.;Equi,M.;Finn,C.;Fu- sai, N.; et al

    V-jepa2.1:Unlockingdensefeaturesinvideoself-supervised learning. Physical Intelligence; Black, K.; Brown, N.; Darpinian, J.; Dhabalia,K.;Driess,D.;Esmail,A.;Equi,M.;Finn,C.;Fu- sai, N.; et al. 2025.π0.5: A Vision-Language-Action Model with Open-World Generalization. In9th Annual Conference on Robot Learning. Singh, J.; Leng, X.; Wu, Z.; Zheng, L.; Zhang, ...

    work page 2025

  8. [10]

    arXiv:2604.05014

    StarVLA: A Lego-Like Codebase for Vision-Language-Action Model Developing. arXiv:2604.05014. Sun, J.; Zhang, W.; Qi, Z.; Ren, S.; Liu, Z.; Zhu, H.; Sun, G.; Jin, X.; and Chen, Z

    arXiv

Show all 20 references
  1. [11]

    arXiv:2602.10098

    VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model. arXiv:2602.10098. Wang,Y.;Ding,P.;Li,L.;Cui,C.;Ge,Z.;Tong,X.;Song,W.; Zhao, H.; Zhao, W.; Hou, P.; et al

  2. [12]

    Ye, A.; Wang, B.; Ni, C.; Huang, G.; Zhao, G.; Li, H.; Li, H.; Li, J.; Lv, J.; Liu, J.; et al

    ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning. Ye, A.; Wang, B.; Ni, C.; Huang, G.; Zhao, G.; Li, H.; Li, H.; Li, J.; Lv, J.; Liu, J.; et al. 2026a. GigaWorld-Policy: An Efficient Action-Centered World–Action Model. Ye, S.; Ge, Y.; Zheng,...

    arXiv

  3. [13]

    Zheng, R.; Wang, J.; Reed, S.; Bjorck, J.; Fang, Y.; Hu, F.; Jang,J.;Kundalia,K.;Lin,Z.;Magne,L.;etal.2025

    Frappe: Infusing world model- ingintogeneralistpoliciesviamultiplefuturerepresentation alignment. Zheng, R.; Wang, J.; Reed, S.; Bjorck, J.; Fang, Y.; Hu, F.; Jang,J.;Kundalia,K.;Lin,Z.;Magne,L.;etal.2025. Flare: Robot learning with implicit world modeling. InProceed- ings of ...

    2025 arXiv

  4. [14]

    arXiv:2604.21391

    From Noise to Intent: An- choring Generative VLA Policies with Residual Bridges. arXiv:2604.21391. A Implementation Details A.1 JEPA-WAM Architecture JEPA-WAM uses a frozen V-JEPA 2.1 ViT-L/16 encoder with an input resolution of384×384. For each camera view, the encoder produc...

    arXiv

  5. [19]

    We report the average task success rate over all 20 tasks

    using 20 manipulation tasks, organized into two groupsof10tasks.Foreachtask,thepolicyistrainedonlyondemonstrationsfromtheCleansetting.Thesametrainedpolicyis thenevaluatedonboththeCleanandRandomsettings,whereRandomintroduceschangestothesceneandobjectconfigurations. We report th...

    2025

  6. [50]

    Near the end of a trajectory, the future observation is clipped to the last available frame. A.2 Vision–Language Initialization Before robot-policy training, we initialize the vision-language interface following the single-stage finetuning setup of Pris- matic (Karamcheti et a...

    2024

  7. [128]

    Only the Qwen LoRA adapters, transition prediction head, and action expert are optimized

    During subsequent robot-policy training, the V-JEPA encoder, visual projector, and base Qwen weights are frozen. Only the Qwen LoRA adapters, transition prediction head, and action expert are optimized. A.3 Robot Policy Training During robot-policy training, JEPA-WAM jointly o...

    2026

  8. [896]

    During robot- policy training, the base predictor is adapted using LoRA with rank 32, scaling factor 64, and dropout 0.1. For latent transition prediction, the final predictor hidden states at the visual-token positions are mapped back to the V-JEPA representation space by a t...

  9. [2023]

    and LIBERO-Plus (Fei et al. 2025). For LIBERO, we jointlytrainonthefourstandardtasksuitesandevaluatetheresultingpolicyonthecorrespondingtesttasks.ForLIBERO-Plus, wedirectlytransferthepolicytrainedonLIBEROdemonstrationstotheperturbedenvironmentswithoutadditionalfine-tuning. We ...

    2025

  10. [2024]

    InForty-first International Conference on Machine Learn- ing,ICML2024,Vienna,Austria,July21–27,2024,volume 235 ofProceedings of Machine Learning Research, 23123– 23144

    Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models. InForty-first International Conference on Machine Learn- ing,ICML2024,Vienna,Austria,July21–27,2024,volume 235 ofProceedings of Machine Learning Research, 23123– 23144. Kim, M. J.; Gao, Y.;...

    2024

  11. [2025]

    arXiv:2506.09985

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Plan- ning. arXiv:2506.09985. Assran,M.;Duval,Q.;Misra,I.;Bojanowski,P.;Vincent,P.; Rabbat,M.;LeCun,Y.;andBallas,N.2023. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architect...

    2023 arXiv

  12. [2026]

    Karamcheti, S.; Nair, S.; Balakrishna, A.; Liang, P.; Kollar, T.; and Sadigh, D

    Vision-alignedLatentReasoningforMulti-modalLargeLan- guage Model.arXiv preprint arXiv:2602.04476. Karamcheti, S.; Nair, S.; Balakrishna, A.; Liang, P.; Kollar, T.; and Sadigh, D

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.