acceptodds
Under review as a conference paper at ICLR 2027

CL-WAM: Closing the Observation Loop in World–Action Models for Humanoid Manipulation

Abstract

World-action models (WAMs) predict video and generate actions, but refreshing an image context does not explicitly correct a persistent predicted state when contact outcomes diverge from a rollout. We introduce the Closed-Loop World–Action Model (CL-WAM), which closes this prediction–observation–action loop inside the world model. At each policy call, the previous belief, memory, and executed action produce a prior; the new observation provides an innovation that corrects the belief and updates memory before the next action chunk and, when scheduled, the next candidate rollout. A port-Hamiltonian-inspired latent field separates a skew/dissipative autonomous term from action-conditioned and predicted-contact terms, while scheduled rollout training feeds real and self-generated visual features through the same belief-update path. At deployment, slow candidate planning periodically refreshes a token that conditions a faster action head. Across four LIBERO suites, CL-WAM achieves 99.2% average success versus 98.6% for the officially released DiT4DiT checkpoint. On 24 RoboCasa-GR1 tasks, it reaches 67.50% (2430/3600) versus 56.7% (680/1200) for official DiT4DiT Run 1, a descriptive 10.8-point difference. Removing belief correction or scheduled rollout training lowers RoboCasa-GR1 success by 3.7 or 4.1 points, respectively. On seven real-world Unitree G1 tasks, observed success is 109/140 (77.9%) versus 101/140 (72.1%) for a locally matched DiT4DiT baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world-action models, robot learning, humanoid manipulation, closed-loop control, belief state estimation, port-Hamiltonian dynamics, writable memory, flow matching

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_niF5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CL-WAM, a closed-loop world-action model that combines belief correction, writable memory, structured latent dynamics, and closed-loop rollout training. It is evaluated on LIBERO, RoboCasa-GR1, and Unitree G1, with consistent improvements over the DiT4DiT baseline.

Critical strengths
  1. Strong empirical performance. CL-WAM shows clear gains on RoboCasa-GR1 and also improves results on real-world G1 tasks.

  2. Careful controlled evaluation. The paper includes matched baselines, parameter controls, planning controls, and component ablations, making the empirical results relatively convincing.

Critical weaknesses
  1. Limited methodological novelty.

Most components of CL-WAM are closely related to existing ideas, including closed-loop training, belief correction, memory, and structured dynamics. The main contribution is their integration into one WAM framework.

  1. Unclear contribution of each component.

The closed-loop rollout schedule already brings a large improvement, while the gains from belief, memory, and the port-Hamiltonian design are smaller. This makes it less clear which components are truly essential.

Questions for authors
  1. Could the authors clarify the main novelty of CL-WAM compared with the closest closed-loop WAM and memory-based methods, beyond integrating these components into one framework?

  2. The closed-loop rollout schedule alone already brings a large improvement. How should we understand the additional contribution of belief correction, memory, and the port-Hamiltonian dynamics beyond this training strategy?

  3. Could the authors clarify the specific benefit of the port-Hamiltonian design? Given that a capacity-matched unconstrained MLP performs relatively similarly, what does the pH structure provide beyond a generic latent dynamics module?

3 comments
AuthorsOpenReview ↗

Clarifying CL-WAM’s novelty and component evidence

We thank the reviewer for recognizing the empirical gains and controlled comparisons. The revised manuscript adds a Mechanism Tests subsection and summary table to the main text. Appendices H, I, and N give the full protocols and results.

Q1. What is methodologically new? CL-WAM uses the discrepancy between the incoming observation embedding and its prediction from the belief prior to correct a persistent state shared by world prediction and action generation. At each policy call, the previous corrected belief, memory, and last executed action produce a belief prior, from which the model predicts the observation embedding. The observation innovation corrects the belief before memory is updated and the next action is generated. The corrected belief also initializes port-Hamiltonian (pH) candidate rollouts when slow planning is due. On the scheduled-mixing training path, real and self-generated visual features enter this same prediction–correction update, linking training exposure to the deployed feedback rule.

World-VLA-Loop uses policy rollouts to refine the world model and policy across training rounds. MemoryVLA updates memory that conditions an action expert. CL-WAM closes the feedback loop within each deployed policy call by using the incoming observation to correct the predictive belief used for subsequent prediction, planning, and control.

Q2. What contributes beyond the rollout schedule? In Appendix N, DiT4DiT +58M obtains 55.08%, adding only the schedule reaches 62.33%, and submitted Full CL-WAM reaches 67.50% on RoboCasa-GR1. The last 5.17-point gap compares whole systems. We therefore ran a separate five-seed, pH-free study with 527M trainable parameters per variant, the same rollout schedule, training budget, eight-candidate planner and non-pH transition, a shared frozen contact predictor, and matched evaluation resets. Across 6,000 trials per variant, success is 61.83% without persistent belief, 63.93% with GRU belief, and 66.40% with innovation-corrected belief. Innovation correction exceeds GRU by 2.47 points (five-seed paired-t 95% CI [2.31, 2.62], conditional on the fixed tasks and resets). In the matched evaluation, 458 trials succeed only with innovation correction versus 310 only with GRU. At horizon 8, future-observation MSE falls from 0.0106 to 0.0082 on 1,138 valid paired event windows (22.6% reduction). In a separate three-seed matched candidate-scoring study (Appendix N.1), memory raises success from 64.17% to 65.00% when added to innovation belief plus a non-pH transition. Replacing that transition with pH, with memory disabled, reaches 65.83% (+1.67 points over the same belief baseline). The task-cost heads share architecture and supervision but are trained separately. The contact predictor shares frozen weights. We interpret only within-study differences as component effects.

Q3. What does the pH parameterization provide over an unconstrained MLP? The complete-field MLP replacement reported in Appendix A reaches 65.3%, versus 67.50% for Full CL-WAM. Table 3’s intervention disables the two pH auxiliary regularizers while retaining the structured field. To examine the field directly, we added same-state action-intervention results in Appendix I and summarized them in Table 4 in Sec. 4.5. Across 360 held-out simulator snapshots, each variant evaluates the same eight candidate actions from the same restored initial state. At planning position 24, Full pH obtains action-effect MSE 0.00190, effect cosine 0.71, and candidate-rank correlation 0.64. The retrained capacity-matched unconstrained MLP obtains 0.00243, 0.59, and 0.53. The pH field reduces measured action-effect MSE by 21.8% (paired MLP-minus-pH difference 0.00053, 95% CI [0.00035, 0.00072]), while candidate-rank correlation improves by 0.11 (95% CI [0.06, 0.16]). Ranking uses independently measured simulator task progress. The Spearman average covers the 324 snapshots with nonconstant candidate progress. Shuffling candidate actions increases effect MSE to 0.00410, and zeroing the control field gives the zero-effect reference of 0.00465. Candidate actions are executed directly in this diagnostic. At deployment, the selected candidate is converted into a plan token that conditions executable actions sampled from fresh noise.

These direct controls show where CL-WAM closes the loop, distinguish innovation correction from generic recurrence, and identify the structured field’s contribution to action-dependent latent effects and candidate selection.

Thanks for the detailed response and additional experiments. The new results help clarify the contributions of the individual components and provide a better understanding of the benefits of the port-Hamiltonian design. These address my main concerns, and I will raise my score accordingly.

Thank you for taking the time to review our response and the additional experiments.

Reviewer_6H5EOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CL-WAM, which aims to improve the execution capabilities of World-Action Models in humanoid robot manipulation through a closed-loop observation mechanism. CL-WAM introduces a persistent belief state that updates state estimates using the residual between predicted and actual observations, as well as a writable memory module for retaining task-relevant information. In addition, it adopts structured latent dynamics inspired by Port-Hamiltonian systems and employs closed-loop rollout training to improve robustness during continuous execution. During inference, CL-WAM uses a dual-rate architecture in which a slow planner searches for candidate plans while a fast action module generates actions, with plans periodically refreshed based on new observations. Experiments demonstrate that CL-WAM outperforms matched baselines on LIBERO, the RoboCasa-GR1 simulation environment, and real-world manipulation tasks using the Unitree G1 humanoid robot, achieving success rates of 67.5% on RoboCasa-GR1 and 77.9% in real-world experiments.

Critical strengths
  1. This paper is logically coherent, with sufficient and sound mathematical proofs.

  2. This paper proposes a novel method that integrates port-Hamiltonian latent flows into WAM, introducing the modeling of latent dynamics. Furthermore, by employing closed-loop rollout training, the model achieves significant performance improvements in both simulation and real-world hardware experiments.

  3. This paper presents very thorough experiments and analysis.

Critical weaknesses
  1. In the experiments presented in Tables 1 and 2, it would be beneficial to include comparisons with other methods based on the WAM framework, such as Fast-WAM[1] and MaskWAM[2], to better demonstrate the capabilities of CL-WAM.

  2. In the ablation study, the authors could include experiments on certain parameters, for example, the size of the memory bank and the number of candidates refreshed during each slow-loop step.

[1] Yuan, Tianyuan, et al. "Fast-wam: Do world action models need test-time future imagination?." arXiv preprint arXiv:2603.16666 (2026). [2] Yu, Hanyang, et al. "Maskwam: Unifying mask prompting and prediction for world-action models." arXiv preprint arXiv:2606.13515 (2026).

Questions for authors
  1. The overall training objective consists of a total of eight auxiliary loss functions. Wouldn't it be difficult to train the model when it is being supervised by so many loss functions simultaneously?

  2. What specific physical information does this Port-Hamiltonian framework incorporate into flow matching?

2 comments
AuthorsOpenReview ↗

Additional Comparisons, Joint Training, and pH Structure

We thank the reviewer for the constructive assessment and specific suggestions.

Additional comparisons and design sensitivity. Table 1 now includes the published LIBERO results of Fast-WAM and MaskWAM, whose four-suite averages are 97.6% and 98.4%, respectively, alongside CL-WAM’s 99.2%. These results broaden the WAM comparison. DiT4DiT (matched), at 98.7%, remains our shared-protocol control. We have added both methods to Related Work and documented the sources of their results in Appendix E. Neither reports results under the RoboCasa-GR1 or Unitree G1 protocols of Table 2.

We also bring the design controls into Sec. 4.4. Memory banks with 16, 32, 64, and 128 slots obtain 63.7%, 66.8%, 67.5%, and 67.7% RoboCasa-GR1 success, respectively (Appendix G). With the trained Full CL-WAM checkpoints held fixed, reducing the slow planner from eight candidates to one lowers RoboCasa-GR1 success from 67.50% to 66.50% (Sec. 4.4 and Appendix A).

Q1. How do we train the eight auxiliary terms jointly? We jointly optimize the video and action flow-matching losses and the eight auxiliary terms using fixed weights (Eq. 12). Supervision follows the rollout’s prediction–correction order. At each of the 24 transitions, the corrected belief, memory read, and demonstration action first produce a belief prior and visual prediction. The closed-loop visual, geometry, and observation losses use predictions computed before correction and stop-gradient targets from the real observation. Mixed real and generated visual features then correct the belief and update memory. A separate free-recursive path starts from the same initial state and uses generated observations without correction. Its rollout loss therefore exposes multi-step errors that immediate observation correction could conceal. The mixing schedule progressively increases exposure to generated inputs over training.

The visual MSE within the observation loss is identical to the closed-loop visual MSE, giving the shared visual MSE a combined weight of 0.6. The pH field uses a separately sampled latent interpolation target, while the contact term is a weak prior. We take elementwise means for tensor MSEs and average stepwise losses over the 24 valid transitions before applying fixed weights. With this procedure, three independent RoboCasa-GR1 training seeds achieve 67.50% mean success with a 1.4-point across-seed standard error. Perturbing each auxiliary weight by ±50% one at a time yielded a maximum observed drop of 1.2 percentage points. Section 3.3 and Appendix D.2 specify the training order and reductions. Sec. 4.4 and Appendix G report the sensitivity result.

Q2. What physical information enters pH flow matching? We condition the learned pH vector field on demonstration actions and supervise its velocity using interpolations from corrected latent beliefs to independently sampled Gaussian points. The field’s structural prior consists of a learned convex potential, skew-symmetric interconnection, positive-definite dissipation, and separate learned control and predicted-contact terms. For the isolated autonomous component under fixed context, the skew term contributes zero latent power and dissipation contributes non-positive latent power. The reported training uses neither measured mechanical-energy targets nor force labels. Section 3.2 and Appendix C state the parameterization and the scope of its energy identity.

We test the functional value of this structure in Appendix I. On the same eight candidate actions from each of 360 held-out simulator snapshots, pH reduces action-effect MSE from 0.00243 for a retrained capacity-matched unconstrained MLP to 0.00190, a 21.8% reduction (paired difference 0.00053, 95% CI [0.00035, 0.00072]). The mean Spearman correlation between the negative candidate score and simulator task progress is 0.64 for Full pH and 0.53 for the capacity-matched MLP. The paired difference is 0.11 (95% CI [0.06, 0.16]) across 324 snapshots with nonconstant candidate progress. These interventions test whether the structural prior improves the action-dependent latent effects and candidate ranking used by the planner.

I appreciate the author’s detailed rebuttal; they conducted numerous experiments that addressed my concerns, so I have decided to maintain my score.

Reviewer_C59XOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CL-WAM, a Closed-Loop World–Action Model designed to improve long-horizon and contact-rich humanoid manipulation by explicitly maintaining and correcting a persistent internal predictive state.

Building upon the DiT4DiT architecture, the proposed framework introduces four main components. First, a learned Kalman-like innovation mechanism updates a persistent belief state using new observations. Second, a writable external memory stores task-relevant information across interactions. Third, a port-Hamiltonian-inspired latent dynamics model introduces structured autonomous, control-conditioned, and contact-related components. Fourth, scheduled closed-loop rollout training exposes the model to a mixture of real and self-generated visual representations to reduce error accumulation during deployment.

The framework further employs a two-rate inference mechanism, where a slower planning pathway periodically evaluates candidate action sequences and produces a conditioning token for a faster action-generation pathway.

Experiments on LIBERO, RoboCasa-GR1, and seven real-world Unitree G1 manipulation tasks demonstrate improvements over DiT4DiT. In particular, CL-WAM achieves 67.50% success on RoboCasa-GR1, compared with 55.06% for a protocol-matched DiT4DiT baseline. Additional controlled experiments investigate model capacity, planning computation, individual architectural components, and long-horizon prediction errors.

Critical strengths

S1. Meaningful problem formulation and a coherent approach to closed-loop world-action modeling

The paper addresses an important limitation of existing world-action models: the lack of an explicit mechanism for maintaining and correcting a persistent predictive state as new observations become available.

While conventional WAMs can incorporate updated visual observations at successive policy calls, this does not necessarily provide a principled mechanism for reconciling previously predicted states with actual execution outcomes.

The proposed framework introduces an explicit observation-correction mechanism, together with persistent memory and structured latent dynamics, to address this issue.

I find the overall formulation well motivated, particularly for contact-rich and long-horizon manipulation tasks where prediction errors can accumulate and the outcome of physical interactions may differ from the model's expectations.

The two-rate planning and action-generation design is also a reasonable way to incorporate longer-horizon predictive information while maintaining a more frequently updated control pathway.

S2. Strong simulation results with carefully controlled baseline comparisons

The experimental results on RoboCasa-GR1 are particularly encouraging.

CL-WAM achieves 67.50% average success across 24 tasks, compared with 55.06% for the protocol-matched DiT4DiT baseline, corresponding to a 12.44-percentage-point improvement.

Importantly, the authors provide comparisons beyond the officially released checkpoint. The locally matched DiT4DiT baseline is trained and evaluated under the same protocol with three independent training seeds, making the comparison more informative.

The paper also includes a capacity-matched DiT4DiT variant with approximately 58 million additional trainable parameters. Its performance remains close to the original matched baseline, suggesting that the improvements cannot be explained simply by increased parameter capacity.

Furthermore, providing the matched baseline with an eight-candidate planning intervention increases its success to 60.6%, still below the 67.50% achieved by CL-WAM. This is an especially useful control because it partially isolates the proposed closed-loop mechanisms from the additional planning computation.

Overall, these comparisons provide relatively convincing evidence for the effectiveness of the complete framework.

S3. Extensive ablations and analyses of the proposed closed-loop mechanisms

A notable strength of this paper is the breadth of its experimental analysis.

The authors provide targeted ablations for closed-loop rollout training, belief correction, writable memory, and the port-Hamiltonian pathway. The controlled comparison between the capacity-matched baseline, the rollout-schedule-only variant, and the complete framework is particularly informative.

Beyond aggregate success rates, the paper also analyzes the magnitude of belief corrections during different manipulation stages, memory-slot specialization, and the accumulation of prediction errors over extended rollouts.

For example, the observation-correction magnitude is substantially larger during contact than during free motion, and the reported latent prediction error at rollout depth 24 decreases from 0.0029 for the matched DiT4DiT baseline to 0.0017 for CL-WAM.

These analyses provide useful evidence that the proposed mechanisms affect the internal predictive state in ways consistent with the intended closed-loop formulation.

The inclusion of real-world Unitree G1 experiments across seven manipulation tasks is also appreciated.

Critical weaknesses

W1. Substantial architectural complexity and limited methodological novelty of the individual components

My primary concern is whether the complexity of the proposed framework is sufficiently justified by its methodological contributions.

CL-WAM combines several established ideas, including recurrent belief-state estimation, Kalman-inspired observation correction, external memory, port-Hamiltonian-inspired dynamics, scheduled rollout training, and hierarchical candidate-based planning.

Although their integration into a world-action model is meaningful, many of these components have close precedents in model-based reinforcement learning, state-space modeling, structured dynamics learning, and closed-loop imitation learning.

The paper introduces approximately 58 million additional trainable parameters and multiple auxiliary objectives, resulting in a relatively complicated training and inference pipeline.

The controlled experiments are appreciated and help justify the complete system. However, the results also indicate that a substantial part of the improvement comes from the closed-loop rollout schedule alone. In Appendix N, introducing this schedule increases RoboCasa-GR1 success from 55.08% to 62.33%, while adding the remaining belief, memory, and port-Hamiltonian components further increases performance to 67.50%.

These findings demonstrate that the additional components are useful collectively, but they also raise questions about whether a simpler closed-loop architecture could capture most of the benefits.

A more explicit investigation of the interactions among these components, particularly whether each component provides complementary improvements beyond the rollout-training mechanism, would help clarify the principal methodological contribution and the necessity of the full design.

W2. Insufficient validation of the physical interpretation and action-conditioned predictive capability of the port-Hamiltonian latent dynamics

The proposed port-Hamiltonian-inspired latent dynamics model is an interesting component, but its relationship to physically meaningful action-conditioned state transitions remains insufficiently established.

As described in the implementation details, the port-Hamiltonian flow is trained using a flow-matching objective between the corrected latent belief and independently sampled Gaussian noise, rather than directly supervising action-conditioned transitions between observed physical states.

Consequently, the objective does not by itself establish whether the resulting latent dynamics accurately represent the consequences of different robot actions.

This distinction is important because the slow planning pathway propagates candidate action sequences through the learned latent field and uses the resulting states for candidate scoring.

Furthermore, the conditional energy-balance result only establishes a dissipation property for the isolated autonomous component under fixed conditioning. It does not guarantee passivity or stability of the complete learned dynamics or the executed robot.

The positive-autonomous-power penalty is also identically zero in exact arithmetic under the proposed skew-symmetric and positive-semidefinite parameterization. Therefore, the reported ablation that disables the port-Hamiltonian structural penalties should not be interpreted as directly isolating the benefit of these structural constraints.

The authors do provide an additional capacity-matched unconstrained-MLP comparison, which is valuable. Nevertheless, more direct evaluations of action-conditioned prediction accuracy, counterfactual state transitions, and the relationship between candidate scores and realized outcomes would substantially strengthen the claims associated with the structured latent dynamics.

Overall, I believe the paper provides evidence that the port-Hamiltonian-inspired representation is useful for control, but the extent to which it constitutes a reliable physical dynamics model remains unclear.

W3. Limited statistical evidence for improvements in real-world humanoid manipulation

Although the paper includes seven real-world Unitree G1 manipulation tasks, the observed improvements over the matched DiT4DiT baseline are relatively modest.

CL-WAM achieves 109 successful trials out of 140, compared with 101 out of 140 for DiT4DiT, corresponding to a 5.7-percentage-point improvement.

At the individual task level, most improvements correspond to only one or two additional successful trials. Similarly, the reported real-world distribution-shift evaluations include only 20 trials per condition, limiting the strength of conclusions regarding generalization and robustness.

These results are encouraging, but they do not provide strong statistical evidence that the observed improvements will consistently generalize across different physical environments, robot instances, or interaction conditions.

Furthermore, the real-world evaluations primarily focus on tabletop manipulation using the humanoid's arms and grippers. The proposed closed-loop formulation is motivated partly by contact dynamics and recovery from prediction errors, but the experiments provide limited direct evaluation of recovery from unexpected physical disturbances, failed grasps, or significant deviations between planned and executed actions.

While the simulation experiments provide substantially stronger quantitative evidence, additional physical evaluations focused on these failure modes would help establish the practical significance of the proposed feedback mechanisms.

Questions for authors

Q1. Contribution and necessity of the individual closed-loop components

The controlled experiments show that closed-loop rollout training alone already provides a substantial improvement over the capacity-matched baseline.

Could the authors further investigate whether belief correction, writable memory, and the port-Hamiltonian pathway provide complementary benefits when added individually to the rollout-trained baseline?

In particular, it would be valuable to understand whether a simpler recurrent belief-based world-action model, trained with the same scheduled rollout objective and planning budget, could achieve comparable performance.

Such comparisons would help clarify which mechanisms are essential and whether their interactions justify the complexity of the complete framework.

Q2. Physical validity and counterfactual accuracy of the learned latent dynamics

Could the authors provide more direct evidence that the port-Hamiltonian-inspired latent field captures meaningful action-conditioned dynamics?

For example, given the same initial observation and different candidate action sequences, how accurately does the model predict the differences in resulting states compared with the corresponding simulated or observed outcomes?

It would also be interesting to evaluate whether the candidate-scoring function correctly ranks action sequences according to their realized execution outcomes.

Since the deployed planner converts the selected candidate into a conditioning token rather than directly executing that candidate, additional analysis of the relationship between the selected plan and the subsequently generated executable actions would be valuable.

These experiments would help clarify the functional role of the structured latent dynamics and planning mechanism.

Q3. Real-world robustness and statistical significance

Could the authors provide additional real-world evaluations that more directly test the benefits of observation correction and persistent memory?

For instance, controlled experiments involving failed grasps, unexpected object displacement, contact disturbances, or interrupted manipulation sequences would be particularly relevant to the proposed closed-loop formulation.

Additionally, could the authors report uncertainty estimates or paired evaluations for the seven real-world tasks to better assess the reliability of the observed 5.7-percentage-point improvement over the matched DiT4DiT baseline?

Such evaluations would strengthen the connection between the proposed closed-loop mechanisms and practical recovery capabilities during physical execution.

1 comment
AuthorsOpenReview ↗

Direct evidence for component value, latent action effects, and G1 recovery

We thank the reviewer for the concrete suggestions. We add mechanism tests in Sec. 4.5/Table 4 and protocols in Appendices H, I, M.1, and N.1.

Q1. Contribution and necessity of the individual closed-loop components. CL-WAM corrects a persistent predictive belief with each observation before that state conditions memory writing, candidate evaluation, and action generation. In the submitted capacity control, adding only rollout scheduling raises RoboCasa-GR1 success from 55.08% to 62.33%. Full CL-WAM reaches 67.50% (Appendix N).

To test ordinary recurrence, we trained three pH-free variants with 527M trainable parameters each under the same schedule, training budget, eight-candidate/24-position planner, non-pH transition, and evaluation resets. Across 6,000 trials per variant, success is 61.83% without persistent belief, 63.93% with GRU belief, and 66.40% with innovation-corrected belief. Innovation exceeds GRU by 2.47 points (five-seed paired-t 95% CI [2.31, 2.62]) and reduces horizon-8 future-observation MSE from 0.0106 to 0.0082 on 1,138 matched events (Appendix H).

In a separate three-seed matched-scoring study, adding memory to innovation belief plus a non-pH transition raises success from 64.17% to 65.00%. Replacing that transition with pH while keeping memory off yields 65.83% from the same 64.17% belief baseline (Appendix N.1). Component effects are compared within each study. The pure schedule-only checkpoint is a separate control.

Q2. Physical validity and counterfactual accuracy of the learned latent dynamics. We tested action-dependent latent predictions against independently executed simulator outcomes (Appendix I). For 360 held-out RoboCasa-GR1 snapshots, each variant evaluates the same eight candidates from the same restored state. At planning position 24, Full pH’s predicted pairwise belief effects have MSE 0.00190 and cosine 0.71, versus 0.00243 and 0.59 for a retrained capacity-matched unconstrained MLP. The 21.8% MSE reduction has paired difference 0.00053 (95% CI [0.00035, 0.00072]). Shuffling actions raises MSE to 0.00410, and zeroing the control field gives the zero-effect reference 0.00465. Mean Spearman correlation between the negative deployed score and independently measured simulator progress is 0.64 for pH versus 0.53 for MLP (paired gain 0.11 [0.06, 0.16] across 324 snapshots with nonconstant candidate progress).

At the first G1 slow-plan refresh, with checkpoint, policy state, candidates, other tokens, and fast-head noise fixed, selected/shuffled/removed plan tokens yield fast-action MSE 0.084/0.142/0.169 relative to the selected candidate (Appendix I.1). In separate paired evaluations, the selected, shuffled, and removed conditions succeed in 109/140, 96/140, and 93/140 episodes, respectively. The selected-token advantages over the shuffled and removed conditions are 9.29 points (95% paired CI [1.62, 16.95]) and 11.43 points (95% paired CI [3.46, 19.40]), respectively. The selected candidate is converted into a token that conditions newly sampled executable actions.

A separate capacity-matched complete-field MLP replacement reduces RoboCasa-GR1 success from 67.50% to 65.30% (Appendix A). For fixed context and action, Eq. (6) gives the full latent energy balance. The dissipation guarantee applies to the isolated autonomous term, whose rectified power penalty is zero in exact arithmetic. Table 3 removes auxiliary penalties while retaining the field. The complete-field replacement and action interventions assess its functional contribution.

Q3. Real-world robustness and statistical significance. We added 210 matched blocks across seven G1 tasks and three disturbances (grasp displacement, pre-contact displacement, and transport slip). Rates are conditional on valid triggers and evaluable recovery records (Appendix M.1). All conditions use the same checkpoint and policy before the disturbance. At the first policy call processing a post-disturbance observation, Belief-off disables innovation correction while retaining belief prediction, and Memory-off disables memory reads and writes. Full, Belief-off, and Memory-off succeed in 164/210, 146/210, and 153/210 trials and recover in 176/210, 151/210, and 160/210. Full exceeds Belief-off by 8.57 points in task success (95% task-cluster interval [0.48, 17.14]) and 11.90 points in recovery ([2.86, 21.43]). The differences versus Memory-off are 5.24 points ([−3.81, 14.76]) and 7.62 points ([0.00, 15.71]). Both confidence intervals for the comparisons with Memory-off include zero. Recovery requires the physical criterion and correct subgoal for three consecutive offline frames. The policy receives no extra observations. The original gain over matched DiT4DiT is 5.71 points (109/140 vs. 101/140). Its approximate unpaired 95% score interval under the independent-trial approximation includes zero ([−4.49, 15.99] points, Appendix M).

Comments on the paper

2 comments
Area_Chair_SQ7UOpenReview ↗

Discussion

Dear reviewers,

Currently, we have the following review ratings

  • niF5: 3: Weak acceptance
  • 6H5E: 4: Clear acceptance
  • C59X: 3: Weak acceptance

While all reviewers agree for acceptance, still, we also have weaknesses pointed out and questions that demand answers. Could you take a look at each other's reviews and the author responses and initiate a discussion?

  • Were your concerns resolved and questions answered properly?
  • How do you weigh the strengths and weaknesses?
  • Are there any issues or agenda you want to highlight?

Please note that ICLR 2027 has Policy on LLM use.

Best, Your AC

The rebuttal addressed my main technical questions, which is why I raised my score. That said, my concern about the limited methodological novelty still remains. The paper has solid experiments and clearly involves substantial work, but overall I still see it as a borderline case. I would also not be surprised if the paper is ultimately rejected.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.