acceptodds
Withdrawn from ICLR 2027

Juno: Taming Predictive Latents for Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Joint-embedding predictive architectures (JEPAs) predict masked or future observations in representation space, offering a natural source of predictive latents for vision–language–action (VLA) models. Yet making these latents useful across pretraining, policy learning, and deployment requires addressing three failures: mismatch with embodiment-specific control, interference with action learning, and teacher miscalibration under distribution shifts. We introduce **_Juno_**, a unified framework built around one action-conditioned JEPA that serves as a control-aligned representation backbone, a predictive teacher, and an adaptable dynamics model. During **pretraining**, we train it on embodiment-matched trajectories and use a dynamic CLS loss to transfer motion-weighted patch dynamics to a compact global state. During **policy learning**, we fuse current-frame JEPA patches into VLA perception and use a decoupled reasoning branch with separate transformation parameters to distill future latent states for action generation. During **deployment**, we adapt the world model on all observed transitions, including failed rollouts, freeze the adapted teacher, and re-align the policy on verified executions using LoRA adapters and a trainable action head, without expert corrections or task rewards. On SimplerEnv, **_Juno_** raises average success from to over Qwen3GR00T, the strongest baseline, and test-time adaptation further reaches ; on a real robot, it retains – success under background, height, and object shifts where the base policy collapses to .

Primary area:
applications to robotics, autonomy, planning
Keywords:
Jepa, vla, world model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GhCVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Juno, a framework that integrates an action-conditioned JEPA world model into the pretraining, policy learning, and deployment adaptation of VLA models. Juno introduces a Dynamic CLS loss to transfer local motion changes into a global representation, fuses patch features from the current frame, and uses a separate predictive branch to align with future latent states. During deployment, Juno-TTT first updates the world model using valid trajectories, then freezes the adapted teacher and updates the policy using only verified successful trajectories. In the experiments section, Juno achieves average success rates of 68.5% on SimplerEnv and 59.6% on RoboCasa, compared with 60.9% and 47.8% for the base model. Real-robot results further demonstrate generalization and adaptation to several environmental changes.

Critical strengths
  1. The method is comprehensive, with clearly connected components. The paper organizes its method around three challenges: aligning representations with control, coordinating prediction with action learning, and adapting the model after deployment. The JEPA module provides visual features, future-state supervision, and a dynamics learning objective at different stages, resulting in a coherent design. Dynamic CLS also preserves a compact global-state interface, facilitating integration with existing VLA models.
  2. The Juno outperforms the base model on both simulation benchmarks, improving performance on 21 of the 24 RoboCasa tasks. In real-robot experiments, the frozen policy maintains success rates of 70%–75% under cumulative changes to the background, height, and object. Component ablations further indicate that feature fusion and future-state alignment jointly contribute to the final performance.
  3. Juno-TTT makes use of both successful and failed trajectories. During deployment, all valid trajectories help update the world model, while only successful ones are used to update the policy. This allows the model to learn from failures without copying failed actions.
Critical weaknesses
  1. The paper does not separately test the contribution of updating the world model or using failed trajectories. It also does not compare Juno-TTT with simply continuing action imitation on the same successful trajectories. Thus, it is unclear how much of the improvement comes from adapting the dynamics model and how much comes from further training on successful data. This weakens the evidence for updating the world model before the policy.
  2. Explicit tests of distribution shifts are mainly conducted on a real robot; comparable simulation tests of shifts during deployment are lacking. The real-robot experiments also focus mainly on placing an object into a bowl, with only 20 trials per condition. Under dynamic lighting, the increase from 55% to 70% amounts to three additional successes. Without results from multiple independent training seeds or uncertainty estimates, it is difficult to assess the stability of smaller gains and how broadly the adaptation method applies.
  3. The second stage relies on verified successful trajectories, but the paper does not state whether success is judged by a human, task-specific rules, or a detector. It also does not describe how invalid segments are removed in enough detail. If all deployment trajectories fail, the second stage has no successful data for action imitation, and the paper does not explain how it handles this case. These details affect the method’s autonomy, deployment cost, and reproducibility.
  4. The effect of Dynamic CLS is not fully explained. Its loss gives more weight to patches with larger visual changes, but such changes may come from lighting or background motion rather than action-relevant objects. The ablation shows an overall performance gain from Dynamic CLS, but does not establish that the learned representation captures the intended control-relevant dynamics. A direct analysis of which patches receive high weights and how CLS changes relate to actions would strengthen this claim.
Questions for authors

Refer to the Weakness

2 comments

Thank you for the detailed comments. The additional experiments are now in the revised manuscript, with new material shown in blue\color{blue}{\text{blue}}. We answer your questions in turn.


W1. The contribution of updating the world model or using failed trajectories / Continuing action imitation on the same successful trajectories

A1. We added two ablations on SimplerEnv. First, we used the same successful episodes to compare three policy updates: action imitation alone, alignment to the original teacher, and alignment after updating the teacher (Table 12). Second, we kept Stage B fixed to successful episodes and varied the episode types used in Stage A, matching the amount of data across conditions (Table 11).

  1. Full Juno-TTT reached 72.7%, 4.8 percentage points above imitation alone. It also outperformed alignment to the original teacher by 3.4 points, showing the benefit of updating the world model before policy alignment.
  2. With Stage B fixed, Stage A reached 72.4% using failures alone and 73.4% using successes alone. Using both yielded 72.7%. Failed episodes provided useful supervision for the world model, although adding them did not improve on successes alone.

Updating the teacher (revised Appendix D.3, p. 23, Table 12\color{blue}{\text{revised Appendix D.3, p. 23, Table 12}})

Adaptation on SimplerEnvAvg. success (%)Change (pts)Future alignment during the policy update
Before adaptation68.5—No policy update
Stage B* only: action imitation67.9−0.6None
Stage B only69.3+0.8Original teacher
Stage A + B (Juno-TTT)72.7+4.2Adapted teacher

Selecting episodes by stage (revised Appendix D.3, p. 23, Table 11\color{blue}{\text{revised Appendix D.3, p. 23, Table 11}})

Stage A (world model)Stage B (policy)Avg. success (%)
Success + failureFailure59.1
SuccessSuccess73.4
FailureSuccess72.4
Success + failure (ours)Success72.7

W2. Comparable simulation tests of shifts during deployment are lacking

A2. Our simulation and real-robot experiments cover different shifts.

  1. Simulation. The policy trains on real Bridge/Fractal demonstrations and adapts on its own SimplerEnv rollouts. This changes both the observation domain and the source of the trajectories. In the standard simulator setting, success rises from 68.5% to 72.7% across four tasks (Table 1).
  2. Real robot. On the object-to-bowl task, adaptation raises success from 8/20 to 13/20 under Gaussian observation noise and from 11/20 to 14/20 under dynamic lighting (Table 4). Adaptation improves success under both deployment shifts.

W3. Whether success is judged by a human, task-specific rules, or a detector / How invalid segments are removed / If all deployment trajectories fail

A3. Success is determined by the simulator in SimplerEnv and by a human observer on the real robot.

  1. Labels and filtering. On the robot, the observer marks success when the designated object is placed in the designated bowl without intervention. A tracking fault requiring an operator reset counts as a failure (Appendix C, real-world protocol). For dynamics training, a usable transition contains an observation, the executed action, and its next observation. In SimplerEnv we trim trailing idle actions while retaining aligned observations. Valid transitions from unsuccessful attempts remain available to Stage A.
  2. No successful episodes. Thank you for pointing out this limitation of Juno-TTT. If all deployment episodes fail, Stage A can still update the world model from the observed consequences of the robot’s actions. Stage B requires successful episodes to provide a suitable alignment signal for the policy, so its update cannot proceed. However, none of the VLA studies we reviewed has demonstrated that test-time parameter updates using only failed deployment episodes can improve success on the original task without additional demonstrations or human corrections. Your question is an important problem, and we will investigate it in future work, thank you.

W4. Which patches receive high weights and how CLS changes relate to actions

A4. We compared where LeWM and V-JEPA 2 patch features change between frames (Fig. 2a, p. 3). We then examined a ten-frame sequence in which the gripper first moves right and then closes. For each pair of consecutive frames, we computed the change in CLS and compared the directions of the nine resulting vectors (revised Appendix A.3, p. 18, Figure 6\color{blue}{\text{revised Appendix A.3, p. 18, Figure 6}}).

  1. Patch weights. In Fig. 2a, LeWM feature changes concentrate around the gripper–object interaction, while V-JEPA 2 responds more diffusely. The figure visualizes temporal cosine distance. Dynamic CLS assigns higher weights to patches with larger L2 feature changes (Eq. 3), so lighting or background motion can also receive high weight.
  2. CLS changes. In Fig. 6, CLS changes point in more similar directions within each motion than across the two motions. In this sequence, the direction of change in the global CLS token distinguishes gripper translation from closing.

Downstream effect. Using Dynamic CLS instead of standard LeWM CLS improves average success by 2.9 percentage points on SimplerEnv and 4.9 on RoboCasa-GR1 (revised Appendix A.1, p. 17, Tables 6–7\color{blue}{\text{revised Appendix A.1, p. 17, Tables 6–7}}).

Teacher / targetAvg.Put Spoon on TowelPut Carrot on PlateStack Green Block on Yellow BlockPut Eggplant in Yellow Basket
V-JEPA 263.567.752.135.499.0
LeWM (patch mean)62.281.360.48.399.0
LeWM (CLS)65.676.061.528.196.9
LeWM (CLS) + Dynamic CLS (ours)68.586.559.431.396.9
Teacher / targetAvg.PnP + ClosePnP From CuttingboardPnP From PlacematPnP From PlatePnP From Tray
LeWM (CLS)54.752.758.853.059.550.4
LeWM (CLS) + Dynamic CLS (ours)59.657.364.855.568.053.6

Thank you again for raising these points. They led us to separate the effect of updating the world model from that of continued imitation.

Reviewer_v7AGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses a JEPA-style world model as a alignment target for VLA training and execution across the entire deployment pipeline, in pretraining, policy learning, and test-time adaptation, where every step used to incur failure modes such as distributional shift. They introduce JEPA training with a dynmaic CLS loss that helps distill motion-weighted features into global representations. They also fuse the JEPA features into the VLA and train a predictive branch to predict the same. Finally, they use adaptation for further downstream task gains.

Critical strengths

(1) There are substantial improvements across simulation and real-world tasks, especially for the generalization experiments with real-world robots.

(2) I think the motivation itself for utilizing dyamic CLS objective is concrete, and the proposed method is a elegant way to reduce the suppressing of local changes in the global token.

Critical weaknesses

(1) Overall, I don't think the experiments showed by the authors, even with the ablation studies, sufficiently show (or isolate) that their contributions (multi-step method / motivations, etc.) actually were crucial in the overall performance increase. I think this issue is prevalant across the paper, but for example, the authors claim that matching the teacher to the embodiment is more important than pretraining scale, and show a few representation analyses, but do not show downstream policy performance comparisons with models that use internet-scale pretrained JEPAs as predictive targets. So the reader is left with only intuitive grasp of the motivation, but not why these additions might help, or why specific designs are necessary, or whether a much simpler method would be better. Also, there is no sufficient ablations for the Juno-TTT part as well, like flow-matching on only successful trajectories, policy realignment with org. teacher, etc. so no real explanation for benefit of the method.

(2) This is mostly an extension of the first problem, but the methodology of this paper has limited novelty, especially in its individual components, such as future representation prediction for VLAs, or the MoT specialization, etc., etc. The parts where the authors claim novelty, such as the CLS loss or the two-stage adaptation seem marginal at best, and if there are some points for persuasion, they should be emphasized more in the writing, and also shown more thoroughly experimentally to be the actual components driving the performance gains.

Questions for authors

In the ablation tables can you provide more info on what exactly is being compared? For example, how is the MoT branch replaced exactly, and how is the alignment loss being applied?

2 comments

We appreciate your assessment and the concrete questions. The corresponding results are included in the revised manuscript and marked in blue\color{blue}{\text{blue}}. Our answers follow.


W1.1. Do not show downstream policy performance comparisons with models that use internet-scale pretrained JEPAs as predictive targets

A1.1. On SimplerEnv, the policy trained with standard LeWM CLS reaches 65.6% average success, compared with 63.5% using V-JEPA 2. Adding Dynamic CLS raises the LeWM result to 68.5%. Downstream closed-loop success for all four targets is reported in revised Appendix A.1 (p. 17, Table 6)\color{blue}{\text{revised Appendix A.1 (p. 17, Table 6)}}.

We consider this sufficient to show that, in this setting, an embodiment-matched teacher trains a better policy than a web-scale video teacher.

Teacher / future targetAvg. (%)Spoon on towelCarrot on plateStack blocksEggplant in basket
V-JEPA 263.567.752.135.499.0
LeWM, mean of patch tokens62.281.360.48.399.0
LeWM, standard CLS65.676.061.528.196.9
LeWM, CLS + Dynamic CLS loss (ours)68.586.559.431.396.9

W1.2. Flow-matching on only successful trajectories, policy realignment with org. teacher

A1.2. No. Imitation alone does not explain the gain. Using the same successful episodes, action flow matching alone reaches 67.9% (−0.6 points) and alignment to the original teacher reaches 69.3% (+0.8 points). Both stay within one point of the 68.5% before adaptation, so we read them as essentially unchanged given ordinary evaluation variation. Updating the teacher first and then training the policy with future alignment reaches 72.7% (+4.2 points) (revised Appendix D.3, p. 23, Table 12\color{blue}{\text{revised Appendix D.3, p. 23, Table 12}}).

Adaptation on SimplerEnvAvg. success (%)Change (pts)Future alignment during the policy update
Before adaptation68.5—No policy update
Stage B* only: action imitation67.9−0.6None
Stage B only69.3+0.8Original teacher
Stage A + B (Juno-TTT)72.7+4.2Adapted teacher

W2. Limited novelty / The CLS loss or the two-stage adaptation seem marginal at best

A2.

  1. Gains from Dynamic CLS and TTT. Adding the Dynamic CLS loss raises average policy success from 65.6% to 68.5% on SimplerEnv (+2.9 percentage points, Table 5) and from 54.7% to 59.6% on RoboCasa-GR1 (+4.9 points, Table 7). At deployment, Juno-TTT raises SimplerEnv success from 68.5% to 72.7% (+4.2 points, Table 1). On the real robot, adaptation increases successful trials from 8/20 to 13/20 under Gaussian observation noise and from 11/20 to 14/20 under dynamic lighting (Table 4). We therefore see a practical benefit from both designs in the settings tested.
  2. What Juno adds. We agree that MoT specialization and the use of future information for action generation have precedents [1–4]. Our contribution is to use the same action-conditioned JEPA throughout pretraining, policy learning, and deployment. During JEPA pretraining, Dynamic CLS teaches the global token to retain local changes. During policy learning, early fusion uses cross-attention to bring current-frame JEPA patches into the VLA’s visual tokens before reasoning. The predictive branch then learns to predict future CLS tokens from the same encoder. This lets the policy use spatial detail from the current observation and learn how that scene will change in the teacher’s representation. At deployment, we update this teacher on valid transitions from successful and failed episodes, freeze it, and realign the policy on successful episodes. We will make this role of JEPA across the three stages clearer in the paper.

Q1. How is the MoT branch replaced exactly, and how is the alignment loss being applied?

A. The two rows in Table 5 are set up as follows (Sec. 4.5, p. 9).

  1. Without MoT (65.4%). We keep the eight reasoning queries and process them with the same Qwen layers as the vision–language tokens. The action decoder still reads both sets of final hidden states. Current-frame fusion and future alignment remain enabled. The comparison with 68.5% therefore tests separate parameters for the reasoning queries.
  2. Without future alignment (60.8%). We keep fusion, the MoT branch, and the same action-decoder inputs, but train the policy with action flow matching alone. With alignment enabled, we project each of the eight final reasoning states and match it to the frozen teacher’s CLS token at the corresponding future step using cosine distance and Smooth L1 loss. The loss trains the projection and the policy layers that produce those states; it does not update the teacher. The action decoder is trained by flow matching.

We thank you again for your professional comments.


References

[1] Physical Intelligence et al. π0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities. arXiv:2604.15483, 2026.

[2] BeingBeyond Team. Being-H0.7: A Latent World-Action Model from Egocentric Videos. arXiv:2605.00078, 2026.

[3] W. Zhang et al. DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge. In NeurIPS, 2025.

[4] D. M. Nguyen et al. FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation. In ICML, 2026.

Reviewer_U6imOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Juno, a framework that uses a JEPA world model to improve VLA training and adaptation. It first trains the global CLS token to capture local changes using motion-weighted patch features. Second, it fuses current-frame JEPA patches into the VLA and adds a separate predictive Transformer branch, jointly optimized with action generation to align reasoning tokens with future observation embeddings. During deployment, Juno updates the world model using both successful and failed rollouts, then freezes it and adapts the policy using successful executions. Experiments on SimplerEnv, RoboCasa-GR1, and a real robot show improved task success and greater robustness to environmental changes.

Critical strengths

Reviewer found the following strengths for the paper regarding the method's novelties:

  • A promising approach to preserving local changes. The Dynamic CLS objective provides a simple way to transfer local patch changes into a compact global state. Within the proposed architecture, removing this objective reduces average success from 68.5% to 65.6%, supporting its practical value.

  • Promising results in simulation and on a real robot. Juno improves average success over its base VLA from 60.9% to 68.5% on SimplerEnv and from 47.8% to 59.6% on RoboCasa-GR1. The real-robot results also show improved robustness under changes in background, height, and objects.

Critical weaknesses

The paper is still unclear or not convincing on the following aspects:

1. Pre-training latent world model phase:

  • The need for the additional Dynamic CLS training stage is not fully established. The paper compares the proposed CLS targets with ordinary CLS targets, but not with direct future-patch supervision or simpler pooled-patch targets. The latter could retain the same number of reasoning queries while avoiding the additional encoder-training stage. The reported 2.9-point gain therefore supports Dynamic CLS within the chosen design, but does not establish its advantage over simpler alternatives. Its additional training cost is also not reported.

  • Dynamic CLS may emphasize task-irrelevant changes. Patch weights depend on the magnitude of temporal feature changes, without explicitly identifying which changes matter for the task. Moving distractors or camera motion could dominate small robots–object changes and bias the global-state supervision. Although the pretrained encoder may suppress some irrelevant changes, this behavior is not guaranteed by the objective. Controlled moving-distractor experiments are needed to test this limitation.

2. Decoupled predictive reasoning phase

  • Joint training may disrupt the pretrained policy and require substantial data and training to stabilize. The action decoder is conditioned on both the original VLM features and the newly added predictive branch from the start of training. Even with copied pretrained weights, the new branch is not initially trained for its future-prediction role or its interaction with the action decoder. Updating these components together may disturb the pretrained policy and require more data and optimization to recover and improve performance. Alignment-loss warm-up does not directly control the branch’s contribution to action generation. The paper provides no learning curves or comparison with staged training to assess this risk.

  • Stage 2 overlooks closely related prior work. FOCA [1] and DreamVLA [2] both use future-supervised representations to support action generation, yet neither is cited, discussed, and compared in related works, or included in the experiments. Their omission leaves the contribution, novelty, and added value of Juno’s predictive branch insufficiently explained. Additional comparisons to those works are required to validate the efficiency of future predictions.

[1] Nguyen, Duc Minh, et al. "FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation." ICML 2026

[2] Zhang, Wenyao, et al. "Dreamvla: A vision-language-action model dreamed with comprehensive world knowledge." NeurIPS 2025

3. Dynamics-First Test-Time Adaptation

Deployment adaptation is not shown to preserve the representation learned in Stage 1. Dynamic CLS is introduced because the standard JEPA objective may overlook local interaction changes. However, deployment adaptation updates the full encoder using prediction and regularization losses without retaining the Dynamic CLS objective or an explicit representation-preservation constraint. This could weaken the control-relevant features learned earlier. The paper does not evaluate whether sensitivity to local interactions or performance in previously supported conditions is preserved after adaptation.

Questions for authors

Please address my critical weakness above. I would revise my rating of the key concerns, which are well discussed with evidence.

2 comments

We sincerely thank you for your thoughtful and constructive comments. We have added the key experiments to the revised manuscript, with revisions marked in blue\color{blue}{\text{blue}}. Our responses to each point follow.


W1.1. The need for the additional Dynamic CLS training stage is not fully established.

A1.1. Using Dynamic CLS as the prediction target improves average success over standard LeWM CLS on both benchmarks: +2.9 percentage points on SimplerEnv and +4.9 on RoboCasa-GR1.

  1. SimplerEnv (revised Appendix A.1, p. 17, Table 6\color{blue}{\text{revised Appendix A.1, p. 17, Table 6}}) We compared four targets. Following your suggestion, we include the direct use of LeWM patch tokens. Dynamic CLS outperforms this simpler alternative by 6.3 percentage points.
Prediction targetAverage success (%)
V-JEPA 263.5
LeWM patch tokens62.2
LeWM CLS65.6
LeWM CLS + Dynamic CLS loss (ours)68.5
  1. RoboCasa-GR1 (revised Appendix A.1, p. 17, Table 7\color{blue}{\text{revised Appendix A.1, p. 17, Table 7}}) We compared LeWM CLS with and without the Dynamic CLS loss. Adding the loss raises average success from 54.7% to 59.6%, with gains on 18 of 24 tasks.
Prediction targetAverage success (%)
Standard LeWM CLS54.7
LeWM CLS + Dynamic CLS loss (ours)59.6
  1. Additional training cost: Stage 1 of LeWM pretraining used 300 RTX 4090 GPU-hours. The subsequent Dynamic CLS stage used an additional 150 RTX 4090 GPU-hours, a 50% increase over Stage 1 alone. We include this breakdown to help assess the gains relative to the additional compute.

W1.2. Dynamic CLS may emphasize task-irrelevant changes.

A1.2. We also considered this issue during development, and we agree with your observation. An abrupt change in lighting or shadows can change many patch features at once, giving those changes more weight than a small gripper movement. However, our design reflects the following considerations.

  1. Pretraining and deployment. In practice, this issue almost never arises during pretraining. Robot demonstration datasets are collected under stable lighting by standard practice, and episodes with severe visual disturbances are discarded as low quality. Abrupt lighting changes are therefore virtually absent from robot pretraining data, and the concern applies essentially only at deployment. Our flickering-light test introduces such changes deliberately, and in that setting we omit the Dynamic CLS loss during test-time adaptation, as we discuss in our response to W3.
  2. Robustness to ordinary appearance changes. This robustness is a key strength of action-conditioned JEPAs and the main reason we chose one from the outset. Because the model learns to predict the consequences of robot actions in latent space, it is encouraged to keep features that reflect those consequences and to discard appearance details, such as lighting and background, that do not. Hegde et al. [1] likewise found that LeWM features retain task state in the presence of visual distractors. In our own background-shift test, the frozen Juno policy succeeded in 15/20 trials, compared with 0/20 for the base VLA (Table 3).

W2.1. Joint training may disrupt the pretrained policy and require substantial data and training to stabilize.

A2.1. We clarify that, following the StarVLA setup, we initialize the VLM backbone from pretrained Qwen3-VL weights and train the predictive branch and action decoder jointly from the first step of VLA training. We gradually increase the weight of the future-alignment loss over the first 5,000 updates (Appendix B.2, pp. 19–20, Table 9\color{blue}{\text{Appendix B.2, pp. 19–20, Table 9}}). We have revised the manuscript to state this initialization and training schedule explicitly (revised Sec. 4.1, p. 7\color{blue}{\text{revised Sec. 4.1, p. 7}}).


W2.2. Stage 2 overlooks closely related prior work.

A2.2. We overlooked FOCA and DreamVLA and thank you for pointing them out. The main difference is how each method defines what the policy should predict: FOCA focuses on future interaction and goal regions, DreamVLA predicts motion, depth, and semantics, and Juno uses one action-conditioned world model both to represent the current scene and to define the future latent states its predictive branch learns to anticipate. We will cite both papers and clarify these differences in Related Work.

  1. FOCA vs. Juno. FOCA [2] asks which future interactions can guide policy adaptation when demonstrations are scarce. It uses auxiliary tokens to predict features of future gripper–object interactions and align with future goal regions. FOCA also uses the same visual encoder for current and future frames. The distinction is what the policy learns to anticipate. FOCA selects task-relevant regions as targets. Juno predicts a sequence of global latent states from a world model trained to capture local changes, with one CLS target per future step.
  2. DreamVLA vs. Juno. DreamVLA [3] asks which complementary aspects of the future scene can support action planning. It encodes current images with an MAE-pretrained ViT and trains separate heads to predict dynamic regions, depth, and semantics. Juno instead organizes prediction around one world model’s representation of the scene over time. The action decoder directly conditions on both the VLM tokens enriched with current JEPA features and the reasoning tokens trained to predict future JEPA states.

FOCA and DreamVLA report results under different evaluation protocols from ours. Given the time and compute available during rebuttal, we did not evaluate either method under Juno’s protocol. We plan to compare all three methods under a common protocol in future work. Thank you for suggesting this comparison.


W3. Deployment adaptation is not shown to preserve the representation learned in Stage 1.

A3. Juno-TTT starts from the pretrained encoder and adapts it to the current task. We agree that updating the full encoder can change its sensitivity to local interactions. We have not directly measured how much of that sensitivity is retained after adaptation.

  1. Why we omit Dynamic CLS during TTT. As discussed in A1.2, even an action-conditioned JEPA can respond to strong visual changes unrelated to the robot’s actions. In our real-robot experiment, flickering colored light changes the appearance of the robot, object, and tabletop at once. Weighting patches by the magnitude of their feature changes could then give the lighting more influence than a local gripper–object interaction. This motivated our choice to update the encoder with action-conditioned latent prediction and regularization during TTT (Eq. 8).
  2. TTT is independent of Dynamic CLS and targets the current task. The TTT objective does not include Dynamic CLS. Unlike continual learning, which explicitly aims to retain performance on earlier tasks, TTT is not designed to preserve the local interaction features learned during pretraining. Its sole purpose is successful control under the current deployment conditions. Our SimplerEnv results show that the benefit of TTT does not depend on Dynamic CLS: with the same TTT objective, success improves whether or not Dynamic CLS was used during pretraining. Without Dynamic CLS during pretraining, success rises from 65.6% to 68.4% (+2.8 percentage points). With Dynamic CLS during pretraining, it rises from 68.5% to 72.7% (+4.2 points).

We thank you again for your careful reading and constructive feedback, which have helped us strengthen the paper.


References

[1] S. Hegde, A. Popov, E. Aljalbout, and N. Smolyanskiy. Diffusion-Steering-Based MPC with a JEPA World Action Model. arXiv:2610.12407, 2026.

[2] D. M. Nguyen et al. FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation. In ICML, 2026.

[3] W. Zhang et al. DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge. In NeurIPS, 2025.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your initial feedback. We have revised the manuscript and highlighted the changes in blue. We added two sets of ablation experiments to examine how future prediction targets affect policy performance and what accounts for the gains from test-time adaptation.

  1. Dynamic CLS and alternative prediction targets (Appendix A.1, p. 17, Tables 6–7). On SimplerEnv, we compared four prediction targets: V-JEPA 2 features, mean-pooled LeWM patch features, standard LeWM CLS tokens, and LeWM CLS tokens trained with the Dynamic CLS loss. Dynamic CLS achieved the highest average success rate (68.5%). Compared with standard LeWM CLS, it improved success by 2.9 percentage points on SimplerEnv and 4.9 points on RoboCasa-GR1.

  2. Juno-TTT ablations (Appendix D.3, p. 23, Tables 11–12). On SimplerEnv, we tested how the choice of successful or failed episodes at each adaptation stage affects task success. We also compared three policy updates using the same successful episodes. Action imitation alone reached 67.9% success, alignment to the original teacher reached 69.3%, and updating the teacher before policy alignment reached 72.7%.

We also clarified the policy’s initialization and joint training setup in Section 4.1 (p. 7). We welcome further questions and look forward to discussing these results with you.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.