acceptodds
Under review as a conference paper at ICLR 2027

Copper-Policy: Focus on the Representation for Robust Robot Manipulation

Abstract

World Action Models (WAMs) acquire behavioral priors by modeling future scene evolution, but predicting detailed futures in pixel or latent space incurs substantial cost. Recent evidence that co-training gains persist without test-time generation raises a question: what must a WAM learn to improve control? We introduce ***Copper-Policy***, which learns a compact World representation with the policy rather than relying on a predefined target space. Through temporal joint-embedding prediction, it predicts future observation embeddings conditioned on task intention without reconstructing pixels. This prediction and action decoding shape the representation jointly, while the policy retains access to current-frame spatial detail for execution. Representation analyses show that the learned features better separate task-driven change from perturbations and provide complementary information for control. Compact prediction targets reduce training tokens per sample, **enabling a 2B-parameter model trained in 9.67 hours on 8 RTX 5090 GPUs and 6 faster than Fast-WAM on matched A100 GPUs**. *Copper-Policy* outperforms every compared method without embodied pretraining on RoboTwin and several embodied-pretrained VLAs on LIBERO-Plus (80.85%). On three challenging real-robot tasks, it performs comparably to and attains a higher average score. Together, these results show that Copper-Policy combines strong control performance with efficient training.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, Joint Embedding Predictive Architecture, Robotics, Self-Supervised Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_u539OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Copper-Policy learns a compact, task-conditioned world representation through joint future prediction and action supervision. An EMA compactor provides future targets, while an asymmetric interface preserves current-frame visual details for action generation. Future prediction is omitted at deployment. Evaluations on LIBERO, LIBERO-Plus, RoboTwin, and three real-robot tasks demonstrate competitive control performance with substantially reduced training costs.

Critical strengths

1.A practically useful combination of predictive representation learning and efficient control. The compact prediction pathway and separate current-frame visual pathway provide a sensible balance between temporal representation learning and precise execution. The reported training cost(9.67 hours on eight RTX 5090 GPUs) makes the approach practically relevant.

2.Broad empirical evidence supporting the complete system. Evaluation spans standard manipulation, distribution shifts, and physical robot tasks. Component ablations support the usefulness of both predictive supervision and current-frame features, while representation diagnostics provide additional evidence of complementary information in the two pathways. Together, these results offer useful insights beyond aggregate benchmark performance.

Critical weaknesses

1.The specific methodological contribution could be better isolated. FLARE already employs compact, action-aware representations and EMA target adaptation. Although this is acknowledged, the additional benefit of Copper-Policy’s target-space formulation remains insufficiently established. Removing prediction does not distinguish jointly learned targets from fixed or staged targets, and removing current-frame features does not isolate the asymmetric attention design. These gaps limit the strength of the novelty and attribution claims.

2.Some empirical conclusions require stronger qualification. Independent training seeds are not reported for the main policy results, limiting interpretation of both small comparative gains and large ablation drops. Robustness improvements also vary across perturbation categories. The representation diagnostics are informative, but their connection to improved closed-loop control remains primarily correlational.

Questions for authors

1.What is the precise distinction between your jointly learned target space and FLARE’s action-aware encoder with EMA adaptation? Is there controlled evidence demonstrating the benefit of this distinction?

2.Do you have comparisons isolating the asymmetric attention mask or comparing future flow matching against direct feature regression or cosine alignment?

3.How reproducible are the real-robot ablation results? In particular, why does removing prediction severely degrade towel folding and banana placement while leaving bowl stacking unchanged?

4 comments
AuthorsOpenReview ↗

# Response to Reviewer u539 — Part 1/3

Response to Reviewer u539 — Part 1/3

We sincerely thank the reviewer for recognizing the practical value of the representation design, the broad empirical evaluation, and the complementary information in the two visual pathways. We have clarified the description of our representation design to better reflect our central research question: how to learn and use a compact World representation for control. We have also added attention-mask, action-gradient, learnable-input regression, and multi-seed results.

W1 / Q1: Relationship to FLARE and the representation design

FLARE [1] indeed shares EMA adaptation and action-aware representations with Copper. The distinction lies in the complete representation design, including what prediction shapes and how action decoding accesses the resulting information.

FLARE uses a Q-former-based vision-language encoder to provide current-observation context to an action-denoising DiT. Learnable future tokens interact with action and state tokens within that DiT. The cosine alignment is applied to the future-token hidden representations at an internal Transformer layer of the action DiT, using future-observation embeddings from an EMA policy encoder as targets.

Copper's online and EMA compactors map current and future observations into the same compact, task-conditioned World space. A separate Future Expert predicts future-view embeddings from the current World representation, while action supervision shapes both the compact input and its contextual pathway. The asymmetric interface additionally supplies current-frame visual features to action decoding while restricting future queries from directly reading them. Action tokens never read predicted future tokens. These learning and information-flow relationships distinguish the formulation; EMA adaptation alone is a shared ingredient. We have clarified these relationships in the Introduction and Related Work.

A matched FLARE reproduction requires architectural and training changes across its coupled prediction and action pathways; we have not undertaken that transfer. Our controlled evidence comes from interventions within Copper, described below; these directly test predictive shaping, auxiliary vision, masking, and action-gradient flow.

The original component ablations test the World and visual pathways and predictive shaping. The original diagnostics examine temporal organization and complementary action information in these pathways. Together, they support the representation's usefulness for control.

To further explain the mechanism, we added three simulation variants: w/o Asymmetric Mask and the two action-gradient interventions. The mask variant complements the original w/o Visual Features ablation to examine information routing. The action-gradient results are:

ConfigurationLIBERO-PlusRoboTwin Avg.
Copper-Policy80.8541.91
w/o Action Shaping World Input73.7341.06
w/o Action Shaping World Branch62.4233.44

Success rates (%), with training seed 42.

The larger losses after additionally blocking the contextual World pathway support the role of action shaping throughout that pathway. This intervention retains joint-embedding prediction, also testing whether a World space shaped by prediction alone can support the policy. The performance drop shows that predictive shaping alone is insufficient to sustain the full policy’s performance. Section 4.5 distinguishes the purposes of these comparisons.

References

[1] R. Zheng et al., “FLARE: Robot Learning with Implicit World Modeling,” in Proc. 9th Conf. Robot Learning (CoRL), ser. Proceedings of Machine Learning Research, vol. 305, 2025, pp. 3952–3971. [Online]. Available: https://proceedings.mlr.press/v305/zheng25a.html

AuthorsOpenReview ↗

# Response to Reviewer u539 — Part 2/3

Response to Reviewer u539 — Part 2/3

W1 / Q2: Auxiliary vision, asymmetric mask, and direct regression

Asymmetric attention mask

The original visual-feature ablation tests the auxiliary pathway; the new mask ablation tests its access constraint while retaining that pathway. The asymmetric interface has two parts: action decoding receives auxiliary current-frame features, and the mask restricts future queries from directly accessing them. Removing visual features tests the first part; allowing all Future-branch tokens to attend to them tests the second.

ConfigurationLIBERO Avg.LIBERO-Plus TotalRoboTwin CleanRoboTwin RandomRoboTwin Avg.
Copper-Policy97.2580.8570.8412.9841.91
w/o Visual Features95.4574.1167.8210.2239.02
w/o Asymmetric Mask96.6581.0365.7210.4838.10

All entries are success rates (%) using training seed 42. Removing the mask reduces LIBERO by 0.60 points and RoboTwin average by 3.81, while increasing LIBERO-Plus by 0.18. Camera, Robot, and Noise improve without the mask, while Light, Background, and Layout decline. Section 4.5 reports these category-dependent effects. Removing the auxiliary visual pathway is especially harmful on the real robot, reducing the score from 96.3 to 20.7. The two interventions therefore examine complementary aspects of the interface.

Direct future-embedding regression

We also evaluated direct future-embedding regression. Specifically, we replaced the Future Expert's noise input with a learnable input and trained it with L1 loss:

ConfigurationLIBERO Avg.LIBERO-Plus TotalRoboTwin Avg.
Copper-Policy97.2580.8541.91
Learnable input + L1 loss96.6577.0433.55

The predictor struggled to reduce L1 training loss. We hypothesize that future multimodality contributes to this fitting difficulty; investigating this explanation remains future work. This variant changes both input and objective; Appendix E, Table A6 reports it as that specific regression configuration.

W2: Independent seeds and the robustness claim

Independent training seeds

We have added three independent training seeds, 0, 7, and 42, for the main simulation policy:

BenchmarkMean ±\pm sample std (%)
LIBERO Avg.97.0 ±\pm 0.2
LIBERO-Plus Total80.8 ±\pm 0.2
RoboTwin Clean70.87 ±\pm 0.08
RoboTwin Random13.00 ±\pm 0.05
RoboTwin Avg.41.94 ±\pm 0.03

Tables 2–3 report these comparison statistics separately from the ablations, which use a common seed 42 reference.

Robustness across perturbation categories

The effects vary by category:

ConfigurationCameraRobotNoise
Copper-Policy78.6773.7494.44
w/o JE Pred.76.1778.5891.63

LIBERO-Plus success rates (%), with training seed 42. Removing prediction lowers Camera and Noise performance while improving Robot. Removing the mask also has mixed effects, improving Camera, Robot, and Noise but lowering Light, Background, and Layout (Section 4.5).

Changes in the learned representation

The original representation diagnostics characterize changes associated with predictive shaping: temporal organization, sensitivity to perturbations, and complementary action information in the two pathways (Section 4.6; Appendix C):

Diagnosticw/o JE Pred.Copper-Policy
Perturbation–temporal overlap (rank 8)0.7980.187
Paired perturbation distance0.3440.134
Action-probe inputR2R^2
World features0.802
Current-frame features0.910
Both pathways0.927

These correlational analyses describe how the representation changes with predictive shaping. We leave further analysis of how these changes causally affect closed-loop success to future work.

AuthorsOpenReview ↗

# Response to Reviewer u539 — Part 3/3

Response to Reviewer u539 — Part 3/3

Q3: Real-robot evaluation and unchanged bowl stacking

All policies are evaluated on the same 50 initial scenes per task, with the same platform, observations, action space, and execution protocol. Appendix D.3 and Table A5 provide the full/half/failure counts. Copper-Policy scores 96.3 overall, compared with 94.3 for π0.5\pi_{0.5} [1] and 84.0 for Fast-WAM [2]; we describe its performance relative to π0.5\pi_{0.5} as comparable.

The prediction ablation has different effects across tasks:

ConfigurationFold TowelPut BananaStack BowlsAverage
Copper-Policy939610096.3
w/o Joint-Embedding Prediction42610043.3
w/o World Branch84828884.7

We thank the reviewer for highlighting this task difference. We checked the prediction ablation with three independent training seeds, 0, 7, and 42, after observing the original degeneration. Across all three seeds, towel folding and banana placement degraded, while bowl stacking remained successful. The manuscript retains the seed 42 models as the common ablation reference, and the table above reports those scores.

The original RoboTwin per-task results show the same contrast among tasks on which the full model already succeeds at high rates:

RoboTwin task (Clean)Copper-Policyw/o Joint-Embedding Prediction
adjust bottle100100
shake bottle horizontally100100
place cans plasticbox9229
place empty cup9544

All entries are success rates (%), with 100 rollouts per task. The first two retain their high performance, while the latter two lose 63 and 51 percentage points. The task-dependent effect therefore also occurs in simulation among tasks with high full-model performance. The scores are reported in the original supplementary RoboTwin per-task results.

This pattern is consistent with tasks differing in fitting difficulty and control-information requirements. It supports a task-dependent role for predictive shaping, including tasks where removing prediction preserves performance and tasks where it causes a large decline.

References

[1] K. Black et al., “π0.5\pi_{0.5}: A Vision-Language-Action Model with Open-World Generalization,” in Proc. 9th Conf. Robot Learning (CoRL), ser. Proceedings of Machine Learning Research, vol. 305, 2025, pp. 17–40. [Online]. Available: https://proceedings.mlr.press/v305/black25a.html

[2] T. Yuan, Z. Dong, Y. Liu, and H. Zhao, “Fast-WAM: Do World Action Models Need Test-time Future Imagination?,” arXiv preprint arXiv:2603.16666, 2026.

AuthorsOpenReview ↗

Final Response to Reviewer u539

As the discussion period draws to a close, we would like to sincerely thank you once again for recognizing the value of our representation design and empirical evaluation. Your constructive comments helped us strengthen the experimental evidence, clarify our methodological contribution, and improve the manuscript. We will ensure that these improvements are incorporated into the final version of the paper. We sincerely appreciate your careful review and valuable feedback, and look forward to further academic exchanges with you.

Reviewer_8bh4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(−1)
Soundness
3
Read review
Paper summary

This paper introduces Copper-Policy, which is an improved World Action Models (WAMs). The model is designed to predict future latent embeddings instead of pixels or latent representations which are not as compact as the proposed embeddings. The proposed method defines a compactor that maps image and text embeddings into a compact target space (world representation). In this space, a flow matching based Future Expert is trained to predict the future representations in the same episode, encoded by EMA version of the compactor. Meanwhile, a flow matching based Action Expert is jointly trained using robot actions, conditioned on the future expert representations world input, current observation and robot state. During inference, the future expert only runs once and the action expert decodes over 10 flow steps based on future expert's latent tokens.

The method was validated on LIBERO, RoboTwin and real-world experiments, shows a relatively strong performance with only 2B trainable parameters and resonable training budget.

Critical strengths

The proposed method runs efficiently in the latent world space and the performance is strong considering the training budget and model size. The experiments are reasonable and the presentation is generally clear and provides useful details.

Critical weaknesses

The main concern is that the novelty and claimed contribution of this paper needs further revision.

Predicting future in latent space is intuitive, DynaMo[1] and VLA-JEPA[2] share the same idea. This paper differs from them by using predicted future as a part of inputs of action generation, while LiLa-WAM and StageWAM mentioned in the paper also has a similar idea. This narrows the contribution of this paper to the joint-embedding prediction. I appreciate the ablations on this but very confused on the config which does not have this joint-embedding prediction (w/o JE Pred.) Does that mean the future prediction loss is completely dropped? If so, that only supports the importance of future prediction instead of joint-embedding prediction. To further clarify the conclusions, an ablation of multiple stage training is suggested: train Future Expert first using only future prediction loss, then apply action loss with and without future prediction loss in the second stage.

At the same time, in w/o JE Pred. setting, does the future expert still take the noised zt+1:t+Hfz_{t+1:t+H_f} tokens as the input? How much the change of input token sizes will affect the performance?

A detailed clarification and further experiments should make the claims in paper better supported.

Finally, the boundary between the compactor and a typical encoder is not clear, the author should claify more what does a compact World representation really different than the a predefined target space, as mentioned in the abstract.

Reference:

[1] Cui, Zichen J., et al. "Dynamo: In-domain dynamics pretraining for visuo-motor control." Advances in Neural Information Processing Systems 37 (2024): 33933-33961.

[2] Sun, Jingwen, et al. "Vla-jepa: Enhancing vision-language-action model with latent world model." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

Questions for authors

Same as the weakness: what's the exact config of w/o Joint-Embedding Prediction?

3 comments
AuthorsOpenReview ↗

# Response to Reviewer 8bh4 — Part 1/2

Response to Reviewer 8bh4 — Part 1/2

We thank the reviewer for recognizing the efficiency, performance, and implementation detail of Copper-Policy. The questions helped us clarify the representation–prediction–action relationship and what the ablations establish.

W1: Predicted futures and the contribution

We would first like to clarify the information used by action decoding. The Action Expert does not consume predicted future tokens. It reads the current World representation, the current-frame visual features, robot state, and instruction. The Future Expert predicts future-view embeddings during training to shape the current World pathway. At inference, it runs once on current tokens to prefill the contextual keys and values; no future tokens are instantiated. The distinction between current-token contextual features and predicted future tokens is now explicit in the Introduction, Section 2, and Section 3.7.

Our contribution concerns the jointly shaped World representation and its asymmetric control interface. LiLa-WAM and StageWAM are concurrent works with different representation and control designs; the detailed comparison is given in our separate comment, “General Response: Contribution and Relationship to Related Work”.

W2 / Q1: What w/o Joint-Embedding Prediction tests

This variant removes the predictive objective while retaining the World input. It tests predictive shaping in the compact space supplied by the online/EMA compactor pair, where the Future Expert predicts future-view embeddings.

The original ablations show that the predictive objective affects control differently across benchmarks:

ConfigurationLIBERO-PlusRoboTwin Avg.Real robot Avg.
Copper-Policy80.8541.9196.3
w/o JE Pred.79.1128.7743.3
w/o World Branch74.5632.7484.7

Success rates/scores (%), using seed 42 models. On RoboTwin and the real robot, retaining the World pathway without predictive shaping performs worse than removing it.

Staged training and action shaping

Thank you for suggesting this comparison. Predictive pretraining followed by action learning, with or without the predictive loss, would help distinguish predictive initialization from continued predictive shaping. Our earlier contrast with a “staged representation” placed too much emphasis on training order. We have clarified the wording to better reflect our central question: how to learn and use a compact World representation for control.

Joint training provides the setting in which predictive and action objectives shape the proposed World pathway. The contribution concerns this representation design and its usefulness for control, rather than the superiority of joint over staged training. We leave the comparison of training schedules for future study. Within our computational budget, we prioritized gradient interventions that directly examine action shaping and whether a World space shaped by joint-embedding prediction alone can sustain the full policy’s performance:

ConfigurationLIBERO-PlusRoboTwin Avg.
Copper-Policy80.8541.91
w/o Action Shaping World Input73.7341.06
w/o Action Shaping World Branch62.4233.44

Success rates (%), with training seed 42.

Blocking action gradients first to the compactor and then additionally to the contextual World pathway produces progressively larger losses, while retaining joint-embedding prediction. Predictive learning alone is insufficient to sustain the full policy's performance: action supervision also shapes the World space and its contextual representations for control (Section 4.5).

Future/noise tokens and sequence length

Yes, the future/noise tokens remain as inputs, with the same forward sequence length. The isolation mask is present throughout training, including in this ablation. No current-context or action tokens attend to the future/noise tokens. Once the predictive loss is disabled, the future-token computation has no loss of its own and contributes no gradient. It also has no forward information path into action generation. Consequently, the presence of future/noise tokens does not change the action outputs or their training gradients; it does not enlarge the context available to action decoding. The Future Expert can still receive action gradients through its current-token contextual pathway, which is separate from its future-token computation. This is precisely the pathway examined by the new action-gradient ablations.

AuthorsOpenReview ↗

# Response to Reviewer 8bh4 — Part 2/2

Response to Reviewer 8bh4 — Part 2/2

W3: Compactor, encoder, and compact World representation

By “predefined target space,” we referred to predicting representations supplied by pretrained video or visual encoders, such as VAE, V-JEPA, or DINO features. We have revised the wording to clarify how Copper’s compact, task-conditioned World space is jointly shaped by predictive and action objectives.

The compactor is an encoding module; its role in the learning formulation distinguishes it from a predefined target space. Frozen V-JEPA supplies patch features, while the trainable compactor defines compact, task-conditioned World tokens jointly shaped by predictive and action objectives. Its EMA counterpart supplies future targets in the same space. These tokens condition both prediction and action decoding, while the additional visual pathway supplies execution detail (Section 3.2; Appendix A.2). The revised Abstract and Introduction state this representation design explicitly.

The original representation analysis (Section 4.6 in the revision) also examines the learned features directly:

Diagnosticw/o JE Pred.Copper-Policy
Perturbation–temporal overlap (rank 8)0.7980.187
Paired perturbation distance0.3440.134
Action-probe inputR2R^2
World features0.802
Current-frame features0.910
Both pathways0.927

Predictive shaping reduces perturbation–temporal overlap, while joint action readout supports the complementarity of the two pathways.

I appreciate the authors' clarification which corrects my understanding and provides additional evidence. The revised paper helps clarify a lot and provides useful details not fully covered by the rebuttal. One suggestion is to replace the terms like world branch, world input, world representation with a more explicit version to reduce confusion, also let the figure 2 reflect these terms. The presentation of this paper could be made clearer. But in genearl, I'm happy to change the score from weak reject to weak accept.

Reviewer_XritOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Copper-Policy, a VLA built on Fast-WAM’s two-stream design. A trainable “compactor” on top of a frozen V-JEPA encoder maps observations into a small set of task-conditioned tokens (the “World representation”). A Future Expert predicts EMA-target future tokens by flow matching, with BYOL-style stop-gradient. An Action Expert decodes actions from these tokens plus current-frame features through a DINOv2-based grid sampler. No future is generated at test time. The authors claim this beats methods without embodied pretraining on RoboTwin, is competitive on LIBERO-Plus and the real robot, and trains faster than Fast-WAM.

Critical strengths
  1. A practical, efficient recipe that remains competitive without robot-data pretraining. Copper-Policy is a 2B-parameter model trained in 9.67 hours on 8 consumer RTX 5090 GPUs with plain data parallelism. Predicting compact tokens instead of generating video shortens the training sequence (244 vs 392 tokens per sample compared with Fast-WAM). Despite this low cost, it reaches 80.85 on LIBERO-Plus, above several embodied-pretrained VLAs (ABot-M0 80.5, VLA-JEPA 77.9, X-VLA 71.4) and far above its closest baseline Fast-WAM (51.5). It is also the best method without embodied pretraining on the RoboTwin average (41.91). Why it is consequential: The training and inference cost of video-generative world action models is a real barrier to adoption, and the paper shows that a much cheaper design keeps most of their robustness. That practical value is the main reason I rate the paper as weak reject, even though the margins over the strongest baselines (π0.5, Cosmos-Policy) are small or absent. I credit the absolute efficiency more than the “6× faster than Fast-WAM” figure, which is partly confounded by different memory and parallelism setups.

  2. A broad and transparent empirical evaluation. The paper evaluates on LIBERO, LIBERO-Plus (10,030 perturbed tasks), all 50 RoboTwin tasks with 100 rollouts each, and three real-robot tasks. The real-robot comparison uses shared seeded initial scenes and the same protocol for π0.5 and Fast-WAM, with episode-level confidence intervals and outcome counts in the appendix. The authors also report weaknesses openly, including Language, Background and Layout on LIBERO-Plus and the low RoboTwin randomized accuracy compared with π0.5. Why it is consequential: This breadth and openness make the claims checkable and show where the method works and where it does not. For example, the real-robot ablations reveal how strongly performance depends on the prediction objective and the current-frame features. It gives a solid empirical base that the authors can strengthen with the ablations I request (frozen target space, gradient-blocked compactor, matched comparison to FLARE and VLA-JEPA, multiple seeds). This is why I regard the weaknesses as fixable and not fatal, and why my recommendation is a weak reject.

Critical weaknesses
  1. Limited novelty, and the one claimed novelty is untested. The method is primarily an integration of existing components. Its building blocks all have direct precedents: a frozen V-JEPA encoder, BYOL/I-JEPA-style EMA targets with stop-gradient, a learned-query compactor (as in LiLa-WAM), Fast-WAM’s two-stream flow-matching design, and future-embedding prediction for policies (FLARE, WoG, Being-H0.7, VLA-JEPA). The paper’s stated difference, and its first contribution, is that the predictive target space is learned jointly with the policy instead of being pretrained or frozen. None of the ablations isolates this. They compare against removing prediction, the world input, or the current-frame features. There is no comparison against a frozen target space (for example V-JEPA features), none against blocking the action-loss gradient into the compactor, and no matched comparison to FLARE or VLA-JEPA. The experiments therefore show that an auxiliary self-predictive loss helps, which is already known. They do not show that learning the target space is what matters. This is consequential because, with the one claimed novelty unsupported, there is little left that the paper demonstrates as new.

  2. The empirical gains are small, not clearly reliable, and not cleanly attributable. Against the strongest baselines the method does not win. On LIBERO-Plus it scores 80.85 vs 85.7 for π0.5 and 82.2 for Cosmos-Policy. On RoboTwin it is about 1–2 points above the best no-pretraining baselines, with worse clean accuracy than Fast-WAM (70.84 vs 77.80) and randomized accuracy far below π0.5 (12.98 vs 46.0). On the real robot it is 96.3 vs 94.3 for π0.5 over three tasks, likely within noise. The largest gains are over Fast-WAM, but that comparison is uncontrolled: Copper-Policy adds frozen V-JEPA and DINOv2 encoders that Fast-WAM does not use. The ablations are also inconsistent, with no training seeds. Removing JE prediction costs under 2 points on LIBERO-Plus but drops the real-robot average from 96.3 to 43.3. This is consequential because the paper’s conclusions rest on these numbers. Small margins, no variance estimates and a confounded main comparison mean the results cannot show that the proposed design, and not the stronger encoders or training variance, explains the improvements.

Together, the lack of demonstrated novelty and the weakly attributable gains are why I recommend rejection at the margin and not acceptance.

Questions for authors

Question:

  1. Learned vs fixed target space. The main claimed contribution is that the target space is learned jointly with the policy. Could you report, under the same training budget: (a) the full model with the target replaced by frozen V-JEPA features or a frozen pretrained compactor, and (b) the full model with the action-loss gradient blocked from the compactor?
  2. Controlled comparison with Fast-WAM. Copper-Policy uses frozen V-JEPA and a DINOv2-initialized spatial encoder that Fast-WAM does not. How much of the gain over Fast-WAM (especially 80.85 vs 51.5 on LIBERO-Plus) remains if Fast-WAM gets the same encoders? Were the Fast-WAM LIBERO-Plus numbers produced under the same evaluation protocol as yours?
  3. Matched comparison to prior future-embedding methods. Could you compare against a FLARE-style or VLA-JEPA-style target under the same backbone, encoders, data and budget? The current comparison uses numbers reported in other papers with different setups. Without this, it is unclear whether gains come from the target design or from the backbone and encoders.
  4. Seed variance and ablation consistency. Removing JE prediction costs about 1.7 points on LIBERO-Plus but drops the real-robot average from 96.3 to 43.3 (Towel 93→4). On RoboTwin, removing both JE prediction and visual features (30.54) outperforms removing JE prediction alone (28.77). Could you report results over multiple training seeds, at least for the real-robot ablations, and explain these discrepancies? Was each ablation tuned, or did all variants share the same 30K-step recipe? If the large real-robot gaps reflect training variance, the case for the design weakens considerably.

Clarification:

  • Efficiency. Is the “6× faster than Fast-WAM” figure measured under matched parallelism and memory settings? Does the 2B parameter count include the frozen 300M V-JEPA encoder, and what is the total parameter count at inference?
4 comments
AuthorsOpenReview ↗

# Response to Reviewer Xrit — Part 1/3

Response to Reviewer Xrit — Part 1/3

We thank the reviewer for the careful assessment and for recognizing the practical training efficiency and broad evaluation of Copper-Policy. We have clarified the description of our representation design to better reflect our central research question: how to learn and use a compact World representation for control. We have also added action-gradient and attention-mask ablations and evaluated the main policy with three independent training seeds.

W1: Contribution and evidence in the original manuscript

Our contribution is the integrated representation design: a compact task-conditioned space jointly shaped by prediction and action, with an asymmetric interface for execution detail (see our separate comment, “General Response: Contribution and Relationship to Related Work”).

The benefit of predictive training for control is our starting point. The experiments examine the proposed representation's role in control and the changes induced by predictive shaping. The original component ablations test the World pathway, predictive shaping, current-frame vision, and their combination. The original results include:

ConfigurationRoboTwin Avg.Real robot Avg.
Copper-Policy41.9196.3
w/o JE Pred.28.7743.3
w/o World Branch32.7484.7
w/o Visual Features39.0220.7

Success rates/scores (%), using seed 42 models. Retaining the World input without prediction performs worse than removing the branch on both benchmarks; removing current-frame vision is especially harmful on the real robot.

The original paper also analyzes what predictive training changes in the representation, beyond policy success rates:

Diagnosticw/o JE Pred.Copper-Policy
Perturbation–temporal overlap (rank 8)0.7980.187
Paired perturbation distance0.3440.134
Action-probe inputR2R^2
World features0.802
Current-frame features0.910
Both pathways0.927

Together with attention focused on manipulated objects, these diagnostics characterize the learned temporal structure, response to perturbations, and complementary control information (Section 4.6 and Appendix C).

To clarify these learning mechanisms, we added an asymmetric-mask ablation and two action-gradient interventions, examining information routing and action shaping within the same representation design. The gradient results are detailed in Q1(b) below.

Q1(a): Alternative target designs

By “predefined target space,” we referred to prediction targets supplied by pretrained video or visual encoders. Our contribution concerns how the compact World space is jointly shaped and used for control, as clarified in the revised Introduction and Related Work.

Thank you for suggesting these alternatives. We have not evaluated either frozen-target configuration: the dense-feature version exceeds our training budget, while a frozen pretrained compactor requires a separate pretraining objective and stage. These are alternative representation designs; our study examines how predictive and action supervision shape the jointly learned World pathway.

The dense-feature configuration exceeds our training budget. For a 256 ×\times 256, 32-frame clip, VAE compression of 32 ×\times 32 ×\times 4 gives 512 tokens; ViT-style patching of 16 ×\times 16 ×\times 1 gives 8,192. Even fourfold token downsampling leaves 2,048 tokens and a 16-fold quadratic attention term relative to the VAE example. Given this computational cost and our available computational budget, we prioritized gradient interventions within the existing joint-training formulation. They directly test how action supervision shapes the World pathway and whether a space shaped by joint-embedding prediction alone can sustain the full policy’s performance.

W1 / Q1(b): Action-gradient ablations

The original experiments established the role of predictive shaping and visual-pathway complementarity. Following this question, we added two nested interventions to examine where action learning shapes the World pathway.

ConfigurationLIBERO Avg.LIBERO-Plus TotalRoboTwin CleanRoboTwin RandomRoboTwin Avg.
Copper-Policy97.2580.8570.8412.9841.91
w/o Action Shaping World Input95.6073.7368.7613.3641.06
w/o Action Shaping World Branch88.7062.4258.448.4433.44
AuthorsOpenReview ↗

# Response to Reviewer Xrit — Part 2/3

Response to Reviewer Xrit — Part 2/3

Action-gradient ablations (continued)

All entries are success rates (%) using training seed 42. The first intervention blocks action gradients to the compactor while preserving action updates to the contextual World representations within the DiT. It reduces LIBERO and LIBERO-Plus performance, with a smaller change in RoboTwin average and a slight increase on Random. The second intervention additionally blocks action shaping of that contextual pathway and produces substantially larger losses across all three benchmarks. On RoboTwin Clean, the further decrease is 10.32 percentage points.

These results support the role of action supervision throughout the World pathway, including the contextual representations used by action decoding. They also explain why blocking only the compactor does not remove all action shaping. With action gradients blocked throughout the World branch and joint-embedding prediction retained, the second intervention also tests whether a World space shaped by prediction alone can support the policy. Its lower scores show that this space is insufficient to sustain the full policy’s performance. The definitions and results are included in Section 4.5 and Tables 2–3.

Q2: Fast-WAM with the same visual encoders

Fast-WAM [1] is built around video generation, so using Copper's encoders requires redesigning their connection to its video and action streams and training the resulting architecture. We have not carried out this transfer and therefore cannot quantify the remaining gain under matched encoders. Our existing comparisons evaluate the complete policies, while internal interventions examine Copper's representation design.

LIBERO-Plus evaluation protocol. The 51.5% Fast-WAM score is an externally reported benchmark result, not a rerun in our evaluation pipeline. Section 4.2 identifies the comparison sources. We therefore cannot verify identical rollout implementation from our own runs; the real-robot comparison, in contrast, uses our shared evaluation setup.

Q3: Matched FLARE and VLA-JEPA configurations

FLARE couples prediction within the action DiT, while VLA-JEPA [2] includes an additional latent-action world model; the methodological relationships are detailed in our separate comment, “General Response: Contribution and Relationship to Related Work”. A matched reproduction requires redesigning the coupled prediction and action pathways, so we have not undertaken these transfers. Our added internal interventions instead examine action shaping and information routing within Copper.

VLA-JEPA is included in Table 2 using published LIBERO-Plus results, as specified in Section 4.2. Appendix C additionally compares dense and token-matched V-JEPA features through representation probes.

W2: Performance in the context of embodied pretraining

The comparison with the strongest policies should also account for their embodied pretraining. Copper uses no embodied-pretrained components and no data augmentation. Under these conditions, it achieves the highest RoboTwin average among the compared methods without embodied pretraining, exceeds several embodied-pretrained policies on LIBERO-Plus, and achieves real-robot performance comparable to π0.5\pi_{0.5}. These results demonstrate the value of the proposed World representation for competitive control under a modest task-training budget.

Q4: Training seeds and task-dependent ablation effects

We thank the reviewer for asking whether the reported differences reflect training variance. We examined main-policy stability and the recurring real-robot ablation pattern separately.

Main-policy stability

We added three independent training seeds, 0, 7, and 42, to the comparison blocks of Tables 2–3:

BenchmarkMean ±\pm sample std (%)
LIBERO Avg.97.0 ±\pm 0.2
LIBERO-Plus Total80.8 ±\pm 0.2
RoboTwin Clean70.87 ±\pm 0.08
RoboTwin Random13.00 ±\pm 0.05
RoboTwin Avg.41.94 ±\pm 0.03

The main simulation scores are consistent across these seeds. All ablation variants follow the shared 30K-step training recipe in Section 4.1; the manuscript retains seed 42 as their common comparison reference.

References

[1] T. Yuan, Z. Dong, Y. Liu, and H. Zhao, “Fast-WAM: Do World Action Models Need Test-time Future Imagination?,” arXiv preprint arXiv:2603.16666, 2026.

[2] J. Sun et al., “VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model,” in Proc. European Conf. Computer Vision (ECCV), 2026, pp. 478–497.

AuthorsOpenReview ↗

# Response to Reviewer Xrit — Part 3/3

Response to Reviewer Xrit — Part 3/3

Reproducibility of the real-robot pattern

After observing severe degradation on towel folding and banana placement but preserved bowl stacking, we trained the prediction ablation with seeds 0, 7, and 42. The same task-specific pattern appeared in all three: towel folding and banana placement degraded, while bowl stacking remained successful. This task-dependent pattern recurs across the three training seeds. The paper reports the seed 42 results consistently across variants.

The original supplementary RoboTwin results show the same phenomenon among tasks that the full model already performs well:

RoboTwin task (Clean)Copper-Policyw/o Joint-Embedding Prediction
adjust bottle100100
shake bottle horizontally100100
place cans plasticbox9229
place empty cup9544

These tasks use 100 rollouts each. Prediction removal preserves two scores at 100%, but lowers the other two by 63 and 51 points. Thus, the contrast between maintained and degraded performance also occurs within one simulation benchmark. Tasks differ in fitting difficulty and control-information requirements, so the effect of predictive shaping need not be uniform across tasks or benchmark averages.

On 30.54% versus 28.77% in RoboTwin

The two visual pathways supply different information and benefit different tasks. Current-frame vision supplies execution detail and shares the modeling burden with the World pathway, so removing it changes what the World representation must support. Its effect can therefore differ with predictive shaping: removing vision lowers the full policy from 41.91% to 39.02%, while the variant without prediction changes from 28.77% to 30.54%. This non-monotonic aggregate is consistent with the task-dependent effects in the per-task results above. The full policy combines both pathways with predictive shaping and achieves 41.91%.

Q5: Efficiency and parameter accounting

Training speed and memory

The approximately 6$\times$ speedup is measured under the same GPU model, per-GPU memory threshold, and global batch, using each method's fastest feasible parallelism or sharding strategy. Appendix B.3 and Table A2 give the settings: eight A100 80GB GPUs, an 80GB per-GPU memory threshold, and global batch 128. Copper-Policy uses DDP at 88.0 samples/s, while Fast-WAM uses full sharding at 14.36 samples/s, giving approximately 6.13$\times$ throughput. Their corresponding peak memory is 31.83GB versus 62.42GB per GPU. This comparison captures training speed under a common resource budget, including the sharding overhead required by each method. Appendix B.3 provides the fuller configuration list. Copper-Policy trains on eight RTX 5090 32GB GPUs with plain DDP in 9.67 hours. We also tested the two baselines on this hardware: under the same training setup, neither Fast-WAM nor π0.5\pi_{0.5} [1] fits with any of the evaluated sharding strategies.

MethodFastest feasible strategyPeak memory/GPU (GB)Throughput (samples/s)Inference latency (ms)
Copper-PolicyDDP31.8388.085
Fast-WAMFull sharding62.4214.3680
π0.5\pi_{0.5}FSDP38.6544.9175

Training measurements use eight A100 80GB GPUs and global batch 128 (Appendix B.3).

Trainable and total parameters

Appendix B.2 now specifies that 2B counts trainable parameters, including the visual feature encoder: approximately 870M in the Future Expert, 790M in the Action Expert, 25M in the compactor, and 300M in the visual encoder, together with the associated modules. It excludes the frozen 300M V-JEPA encoder, the frozen text encoder, and the EMA copy. The text encoder is Wan2.2-5B's encoder-only UMT5-XXL, approximately 5.68B parameters according to its official architecture [2]. Including these frozen encoders, the deployed model contains approximately 8B parameters; 2B is the trainable count. T5 encodes the instruction, the Future Expert prefills current-token context once, and only the Action Expert is iterated over the action flow steps. The EMA compactor and future-token generation are absent at inference. Inference latency is separately reported as approximately 85ms, compared with 80ms for Fast-WAM and 75ms for π0.5\pi_{0.5}.

References

[1] K. Black et al., “π0.5\pi_{0.5}: A Vision-Language-Action Model with Open-World Generalization,” in Proc. 9th Conf. Robot Learning (CoRL), ser. Proceedings of Machine Learning Research, vol. 305, 2025, pp. 17–40. [Online]. Available: https://proceedings.mlr.press/v305/black25a.html

[2] Wan-Video, “Wan2.2: T5 text encoder implementation,” GitHub. Accessed: Nov. 14, 2026. [Online]. Available: https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/t5.py

AuthorsOpenReview ↗

Final Response to Reviewer Xrit

Thank you for the time and effort you have devoted to reviewing our work and engaging in the discussion. We greatly appreciate your constructive comments and thoughtful suggestions, which have helped us reflect more deeply on our research and further improve the manuscript. We will incorporate these improvements into the final version of the paper. We sincerely appreciate your time, thoughtful feedback, and valuable contributions to improving our work.

Comments on the paper

3 comments
AuthorsOpenReview ↗

# General Response: Contribution and Relationship to Related Work — Part 1/2

General Response: Contribution and Relationship to Related Work — Part 1/2

We thank the reviewers for their comments on the contribution and its relationship to prior work. We have refined the Introduction and Related Work to make the representation design and its positioning more precise.

Integrated representation design

Our main contribution is the integrated design for learning and using the World representation.

  • Compact task-conditioned space. The online compactor maps current observations into compact World tokens; its EMA counterpart maps future observations into the same space to provide prediction targets.
  • Predictive shaping. A separate Future Expert predicts future-view embeddings from the current World representation and instruction, shaping the representation used for action generation.
  • Action shaping. Action supervision shapes both the compact World input and its contextual representations within the transformer, anchoring the World pathway to control.
  • Asymmetric prediction–control interface. Auxiliary current-frame vision supplies execution detail to action decoding, while masking restricts future queries from directly accessing that detail channel. Action tokens never read predicted future tokens.

These relationships determine what information the representation learns and how control uses it. The representation design cannot be summarized by a single component. It is defined jointly by the representation space, learning objectives, gradient paths, and information routing. Learning the targets jointly is one part of this design; fixed or pretrained target spaces are alternative designs, whose relative superiority is outside the question studied here.

Relationship to prior work

Fast-WAM [1]. Its finding that training-time future modeling can benefit control without test-time future imagination motivates our study. Fast-WAM uses a video-generative world stream; Copper directly learns a compact, task-conditioned representation through joint-embedding prediction and action supervision.

FLARE [2] and WoG [3]. FLARE shares action-aware representations and EMA adaptation. It aligns future-token hidden representations at an internal Transformer layer of the action-denoising DiT to EMA future-observation embeddings, while Copper uses a separate Future Expert to shape the current World pathway and supplies action decoding with auxiliary current-frame vision. WoG first learns future conditions through action generation, then freezes their encoder and trains current-context queries to predict those conditions. Copper jointly shapes its current World pathway through predictive and action objectives.

VLA-JEPA [4]. It constructs latent actions through world-state prediction in a frozen V-JEPA2 space, combining world-model and action objectives on robot data. Copper's compact observation representation, prediction target construction, and action interface form a different learning formulation. Transferring these methods into a common architecture requires changes across their coupled prediction and control pathways, beyond substituting one component.

LiLa-WAM [5] and StageWAM [6]. These are concurrent works in the shared direction of future-informed control, with distinct designs. LiLa-WAM decodes predicted latents into frozen DINOv3 features. StageWAM freezes a next-stage predictor to condition video–action generation. The broad description “latent future prediction” leaves out the learning objectives and information-flow relationships that distinguish these methods.

References

[1] T. Yuan, Z. Dong, Y. Liu, and H. Zhao, “Fast-WAM: Do World Action Models Need Test-time Future Imagination?,” arXiv preprint arXiv:2603.16666, 2026.

[2] R. Zheng et al., “FLARE: Robot Learning with Implicit World Modeling,” in Proc. 9th Conf. Robot Learning (CoRL), ser. Proceedings of Machine Learning Research, vol. 305, 2025, pp. 3952–3971. [Online]. Available: https://proceedings.mlr.press/v305/zheng25a.html

[3] Y. Su et al., “World Guidance: World Modeling in Condition Space for Action Generation,” arXiv preprint arXiv:2602.22010, 2026.

[4] J. Sun et al., “VLA-JEPA: Enhancing Vision-Language-Action Model with Latent World Model,” in Proc. European Conf. Computer Vision (ECCV), 2026, pp. 478–497.

[5] F. Yang et al., “LiLa-WAM: Lightweight Latent Reasoning World-Action Model for Robotic Manipulation,” arXiv preprint arXiv:2608.03701, 2026.

[6] X. Liu et al., “StageWAM: Joint-Embedding Stage Prediction for World-Action Models in Robot Manipulation,” arXiv preprint arXiv:2608.10780, 2026.

AuthorsOpenReview ↗

# General Response: Contribution and Relationship to Related Work — Part 2/2

General Response: Contribution and Relationship to Related Work — Part 2/2

Original evidence and additional experiments

The original experiments evaluate the proposed representation's usefulness and characterize what predictive shaping changes. Component ablations examine the World pathway, joint-embedding prediction, auxiliary vision, and their combination. On RoboTwin and the real robot, retaining the World pathway without prediction performs worse than removing the branch. The representation diagnostics examine temporal organization, sensitivity to perturbations, and complementary action information in the World and visual pathways. Joint action readout reaches R2R^2 = 0.927, above either pathway alone.

The additional simulation ablations clarify the roles of information routing and action shaping. Removing the asymmetric mask complements the original visual-feature ablation, separating the presence of auxiliary vision from its access constraint. Two nested gradient interventions block action shaping first at the compactor and then throughout the contextual World pathway. LIBERO-Plus success changes from 80.85% to 73.73% and 62.42%, respectively. The latter retains joint-embedding prediction while blocking action gradients throughout the World branch, testing whether the prediction-shaped space alone can support the policy. Its substantial performance drop shows that predictive shaping alone is insufficient to sustain the full policy’s performance and supports the role of action supervision in learning the World representation. The reviewer-specific replies give the full results and directly address each requested comparison.

AuthorsOpenReview ↗

# Author Final Remarks

Dear Reviewers and Area Chair,

We sincerely thank you for your valuable suggestions and insights. We appreciate your supportive comments highlighting our work's strengths:

  1. A novel, unified representation-learning methodology (u539). Copper learns control-relevant representations through future prediction without costly future generation. u539 recognized its "practically useful combination of predictive representation learning and efficient control" and "sensible balance between temporal representation learning and precise execution."
  2. Strong performance at low cost without embodied pretraining (Xrit, 8bh4, u539). Xrit noted Copper is "the best method without embodied pretraining on the RoboTwin average (41.91)" and called generative WAM cost "a real barrier to adoption"; 8bh4 found performance "strong considering the training budget and model size"; u539 recognized practical relevance. Copper achieves 6.13× Fast-WAM's training throughput.
  3. Broad evaluation and insights (Xrit, u539). Xrit praised a "broad and transparent empirical evaluation"; u539 highlighted "broad empirical evidence supporting the complete system" and results that "offer useful insights beyond aggregate benchmark performance."
  4. Clarity and detail (Xrit, 8bh4). Xrit rated the paper “Clear and easy to follow”; 8bh4 acknowledged that “the revised paper helps clarify a lot and provides useful details.”

Major Improvements During Rebuttal

We strengthened the manuscript in three ways:

  1. New controlled experiments on the proposed mechanisms. We added two nested action-gradient interventions, an asymmetric attention-mask ablation, and an alternative prediction configuration. These experiments provide more direct evidence for the roles of action shaping, information routing, and predictive learning.
  2. Stronger experimental robustness and reproducibility. We added three independent training seeds for the main policy, verified the reproducibility of key real-robot ablation findings, and provided more detailed analyses and experimental clarifications to strengthen the reliability of our conclusions.
  3. Clearer methodological contribution and architecture. We substantially revised the Introduction, Related Work, and architectural explanations to clarify Copper's unified representation-learning methodology, its distinctions from prior work, and the prediction–action information flow.

Responses to Specific Reviewer Concerns

  1. Action-gradient ablations (Xrit, 8bh4). We added two nested interventions to examine action supervision of the compactor and contextual World pathway, providing direct evidence that action shaping matters beyond predictive supervision alone.
  2. Additional controlled prediction and routing experiments (u539). We added an asymmetric attention-mask ablation and a learnable-input L1 regression configuration, examining information-access constraints and an alternative future-prediction formulation.
  3. Multi-seed reproducibility (Xrit, u539). We evaluated the main policy with three independent training seeds and verified that the key real-robot prediction-ablation patterns recur across seeds, strengthening confidence in the reported findings.
  4. Methodological novelty and positioning (Xrit, 8bh4, u539). We substantially revised the Introduction and Related Work to clarify Copper's contribution as a unified representation-learning methodology, and more precisely position it within the development of World Action Models.
  5. Architectural clarity and terminology (Xrit, 8bh4, u539). We clarified the prediction–action information flow, distinguished the roles of World input and contextual World representations, and refined terminology that previously required additional domain knowledge to interpret.

Reviewer Feedback and Closing Remarks

We are encouraged that 8bh4 updated the rating from Weak Reject to Weak Accept, explicitly acknowledging that our clarifications "correct my understanding and provide additional evidence" and that "the revised paper helps clarify a lot and provides useful details." Xrit acknowledged that some concerns had been addressed, although the score remained unchanged; we subsequently provided a further response clarifying our methodological contribution and the complete nested action-gradient evidence. u539 provided positive feedback in the initial review but has not yet provided follow-up feedback; we have addressed the specific concerns through additional controlled experiments, multi-seed evaluations, and manuscript revisions.

We respectfully ask the AC to consider the complete rebuttal discussion, the additional experimental evidence, and the positive reassessment following our revisions. We believe Copper's unified representation-learning methodology and strong performance at substantially lower training cost can make World Action Models more accessible and contribute to their broader development within the robotics community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.