acceptodds
Withdrawn from ICLR 2027

De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Cascade world-action models (WAMs) use predicted visual futures to condition robot actions, but inherit costly iterative inference from diffusion-pretrained experts. We ask what a single-step predicted future must preserve to support accurate single-step action prediction. In the evaluated WAM, resampling video noise has little effect on downstream actions, yet naively truncating both streams to one step substantially reduces task success. Video-space supervision only partially closes this gap, and a local sensitivity analysis shows that vision prediction errors have highly unequal consequences for action accuracy. These findings motivate DirectWAM, which converts pretrained video and action experts into single-step predictors without stochastic sampling through action-sufficient distillation, while retaining the explicit future-to-action cascade. After learning a one-step action model, we train the future predictor by decoding actions from its own predicted futures and backpropagating demonstration action error through the expert, with an optional future-latent reconstruction loss. The resulting objective evaluates the student-generated future through its downstream control consequences rather than relying on representation matching alone. DirectWAM achieves 97.1% average success across four LIBERO suites and 85.4% on RoboTwin 2.0, outperforming Flash-WAM at the same one-evaluation-per-stream budget. This advantage has also extend to three real-world tasks that we designed. Compared with the full iterative teacher, it reduces inference latency by .

Primary area:
generative models
Keywords:
World Action Model, Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_9J5AOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DirectWAM, which converts pretrained diffusion-based video and action experts into deterministic single-step predictors while retaining an explicit future-to-action cascade. The authors first train a single-step action model using demonstrated future representations, then freeze it and train the future predictor by backpropagating demonstration-action error through the action model. An optional reconstruction loss anchors the predicted future representations. Experiments on LIBERO, RoboTwin 2.0, and three real-robot tasks demonstrate substantial improvements over naive step reduction and favorable performance against Flash-WAM, with a reported 28× reduction in model-side inference latency relative to the full iterative teacher.

Critical strengths

1.The method addresses a practically important efficiency problem and achieves a useful performance–latency trade-off. Iterative inference in both video and action streams makes cascade WAMs expensive to deploy. DirectWAM reduces each stream to one evaluation while preserving explicit future conditioning. On RoboTwin, it achieves 85.4% average success, compared with 54.9% for naive single-step inference and 78.2% for Flash-WAM. Simulation and real-robot results support its ability to recover much of the control performance lost through step reduction while substantially reducing inference cost.

2.Controlled representation interventions provide meaningful evidence for action-oriented future learning. The selective-replacement experiments hold the action model fixed and modify only the future representation supplied to it. Transferring a small, locally action-sensitive component recovers a substantial portion of the action-error improvement and closed-loop performance gain. This provides stronger support than visualization alone for the claim that representation-error magnitude is an incomplete measure of control relevance. The visual-quality evaluation also indicates that action supervision preserves decodable visual structure in the evaluated configurations.

Critical weaknesses

1.The methodological contribution is relatively incremental, and the benefits of the specific training design and retained future pathway are insufficiently isolated. The central mechanism is demonstration-action regression through a differentiable cascade initialized from pretrained experts. The comparison with visual-only supervision supports the usefulness of action supervision, but does not establish the individual importance of the two-stage schedule, freezing the action expert, or the reconstruction anchor. Matched comparisons with joint single-step fine-tuning and anchor-strength ablations would better substantiate these choices. Furthermore, the paper lacks an action-only baseline trained with comparable pretrained components, data, and optimization resources. FastWAM and FasterWAM provide context, but their architectural differences prevent a controlled assessment of the intermediate future’s value. Consequently, the control benefit of retaining this additional prediction pathway remains insufficiently quantified.

2.The scope and uncertainty of the experiments limit the strength of the broader conclusions. The offline stochasticity analysis uses a learned single-step action model, so its attenuation of video variation may partly reflect that particular readout. Closed-loop interventions provide complementary evidence, but remain limited to one WAM instantiation and the evaluated task distributions. They do not establish when deterministic conversion remains appropriate for tasks with genuinely distinct valid strategies or substantial uncertainty. In addition, the main policy results do not report variability across independent training runs, real-robot evaluation uses only 20 trials per task, and the Flash-WAM LIBERO results are quoted rather than obtained through a matched reproduction. DirectWAM also remains below the full teacher on RoboTwin, at 85.4% versus 90.7%. The evidence therefore supports a strong efficiency–performance trade-off more clearly than broadly lossless preservation of teacher capabilities.

Questions for authors

1.How important are the two-stage schedule and the frozen action expert? Do you have matched comparisons with joint single-step fine-tuning and an anchor-strength ablation that isolate the benefits of these design choices?

2.How does DirectWAM compare with an action-only student trained using comparable pretrained components, data, and optimization resources? This would help quantify the control benefit and computational cost of retaining explicit future prediction.

3.Does the attenuation of video-sampling variation persist when the offline analysis uses the teacher’s native iterative action model? Do any evaluated tasks involve distinct valid action strategies for the same context, and how does the deterministic student behave in those cases?

4.How stable are the principal improvements across independent training runs? Additional uncertainty estimates would strengthen the simulation comparisons, particularly where performance is close to or slightly above that of the teacher.

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer 9J5A (1/2)

We thank the reviewer for the careful assessment and constructive questions.

Q1. Training design ablations

LIBERO models are trained jointly across all four suites with an effective batch size of 16 and 1,000 total steps, allocated as 600 Stage-1 + 400 Stage-2 steps for staged training. The 1,000-step joint run is the budget-matched comparison; additional joint runs at 400 and 1,600 steps examine the effect of training duration. LIBERO-Long is the evaluation suite for the reported ablations. RoboTwin uses 8,000 Stage-1 + 6,500 Stage-2 steps, with a budget-matched 14,500-step joint run. Staged versus joint training. Appendix B.2, Table 7 shows that staging gives more consistent performance across the tested LIBERO-Long checkpoints, while joint training reaches comparable performance on RoboTwin at the same total step budget. Action-expert freezing and reconstruction weight. Appendix B.3, Tables 8–9 support freezing as a practical choice in the evaluated settings and show that reconstruction strength should be selected per benchmark: anchoring helps on RoboTwin, while LIBERO-Long works well without it. These are empirical, potentially task-dependent implementation choices; our central argument concerns learning predicted futures through their downstream action consequences, rather than establishing a universally optimal training configuration.

Q2. Action-only comparison and the role of explicit futures

Our research scope is efficient conversion of pretrained cascade WAMs while preserving their future-to-action interface. Establishing whether explicit futures are generally preferable to action-only policies requires a broader architectural study. Fig. 4(a) provides comparisons with Fast-WAM and FasterWAM, with numerical Fast-WAM results in Appendix B.6, Table 11. In an additional LIBERO-Plus Long evaluation, DirectWAM achieves 69.4%, versus 40.2% for Fast-WAM and 67.3% for FasterWAM. We leave this additional benchmark outside the main paper to maintain its focus on cascade conversion. Architectural and training differences limit these results to contextual comparisons; we do not have an action-only student matched in pretrained components, data, and optimization resources that would isolate the control benefit of retaining the future pathway. The retained, decodable visual future also supports monitoring and debugging. Our LIBERO visualizations remain decodable even without reconstruction anchoring, and Appendix C directly tests the effects of changing future representations on downstream actions. Preserving this inspectable and intervenable interface is a practical objective of our cascade conversion.

Q3. Native-teacher attenuation and distinct strategies

Appendix A.1, Table 4 (lower panel) reports the offline attenuation analysis using the teacher's native 50-step action model, holding the future samples and contexts fixed and using common action noise across video samples. On RoboTwin (256 contexts × 32 samples), attenuation remains 70.4× pooled / 25.4× per-context RMS. On LIBERO (64 contexts × 16 samples), it is 5.5× / 2.9×. Thus, attenuation persists under the native iterative readout and is not solely an artifact of the learned single-step action model; its magnitude varies across benchmarks. Distinct-strategy diagnostics. On RoboTwin, Appendix A.4, Table 6 reports calibrated multimodality tests on 237 clean contexts across 47 tasks, with 32 teacher samples per context. The tests detected no contexts in the 8-dimensional PCA analysis and one in the 16-dimensional analysis; the latter's visual clusters did not induce distinct action clusters (Appendix A.2). Appendix A.3 and Fig. 2(b) report 8 teacher rollouts for each of 30 task–initial-state pairs (240 episodes), holding scene, instruction, and action noise fixed while varying video noise. We did not observe separate successful strategy branches: successful trajectories remained concentrated around the same path. These diagnostics characterize the evaluated teacher and task distribution. We have not established the deterministic student's behavior on tasks deliberately constructed to require choosing between distinct valid strategies, and will state this limitation explicitly.

AuthorsOpenReview ↗

Official Response to Reviewer 9J5A (2/2)

Q4. Variability across independent training runs

Appendix B.4, Table 10 reports three independent 6,500-step Stage-2 training runs from the same 8,000-step Stage-1 checkpoint. Each seed is evaluated on 50 tasks with 50 episodes per task per split (2500 episodes per split), the seed means therefore summarize a separate evaluation. The limited variation across seeds supports the stability of this Stage-2 recipe under the evaluated settings.

Scope of the evidence. The revised Table 2 replaces quoted Flash-WAM LIBERO scores with our own trained reproduction, with protocols described in Appendix B.5.

Reviewer_LKyBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how to turn a cascade world–action model (LingBot-VA: video expert → future latent → action expert) into a deterministic one-step predictor. An analysis first shows that resampling the video noise barely affects actions, while naive one-step truncation and video-space supervision leave a large control gap. A Jacobian probe also suggests that a small action-sensitive part of the representation error matters disproportionately. DirectWAM then trains a one-step action head on ground-truth futures. Next it trains the one-step video predictor by backpropagating the demonstration action error through that frozen head, with an optional latent anchor.

Critical strengths
  1. A simple, effective method. Two short supervised stages yield a deterministic one-step cascade that clearly beats consistency distillation at the same budget on RoboTwin (85.4 vs 78.2, all 50 tasks × 50 episodes). The gain is larger under domain randomization (+9.5).

  2. An informative analysis. Sec. 3 separates stochastic sampling from iterative refinement, using offline propagation statistics and closed-loop interventions with paired tests (App. A.3). It shows convincingly that video-noise diversity is not what the action model relies on, which is a useful finding for the WAM-acceleration literature.

  3. Careful controls. The selective-replacement experiments include equal-norm random controls (1–2% recovery vs 80%), and the closed-loop version (Tab. 8) shows that a 1.4%-energy component recovers the full success gain.

Critical weaknesses

W1 [Experiments] Key baselines are missing from the main comparison.

The vision-supervised one-step student (76.0/78.2 on RoboTwin, Sec. 3.2) is the direct control for the central claim (action loss vs representation loss), but it appears only in the text. It should be in Table 2, on both benchmarks. There is no action-only one-step baseline on the same backbone, i.e., end-to-end fine-tuning of the policy without the explicit future. With λ_v = 0 on LIBERO, the "future" is simply an intermediate representation trained by action loss. Whether keeping the cascade helps control at all is therefore untested. Fast-WAM appears only as points in Fig. 4, with no numbers.

W2 [Experiments] The comparison with Flash-WAM needs care. On LIBERO, the Flash-WAM numbers are quoted from its paper (†). On RoboTwin they are reproduced (78.2), whereas the Flash-WAM paper reports 85.5% RoboTwin and 96.1% LIBERO on the same LingBot-VA backbone (at a slightly larger step budget). Please clarify the reproduction protocol and training budget, and report Flash-WAM's own 1/1 number if available. DirectWAM also receives supervised fine-tuning, whereas the truncation baselines receive none, so budgets should be stated. All results come from a single backbone; Flash-WAM also evaluates on Motus.

W3 [Presentation/Integrity] The figures are unreliable.

Fig. 1 appears to be AI-generated illustration rather than a faithful schematic of the method and data.

It is built from generic, repeated robot thumbnails. The inset "overlapped action traj" in panel (a) is a top-down stove scene that matches none of the evaluated RoboTwin, LIBERO or real-robot setups, and its trajectories are not tied to any reported experiment. The success rates printed in panel (b) do not match Table 2 (teacher 91.6 vs 90.7; naive 1/1 57.6 vs 54.9). The AI-use statement mentions only writing assistance, not figure generation. Please state how Fig. 1 was produced, replace it with a schematic and real data, and make its numbers consistent with the tables.

Fig. 4(b) is uninformative. The four columns in each row (full teacher, one-step teacher, Flash-WAM, DirectWAM) are almost indistinguishable. There is no ground-truth future column, and the stage or time step of each example is not indicated. The figure therefore cannot support the claims drawn from it ("fewer ghosting artifacts", "no visual degradation relative to Flash-WAM").

Questions for authors
  1. Can you add the vision-supervised one-step student and an action-only one-step baseline (same backbone, no future) to Table 2?

  2. How was Flash-WAM reproduced on RoboTwin, and with what budget? How does this relate to the 85.5% reported in its paper?

  3. How was Fig. 1 produced (including the "overlapped action traj" inset), and why do its numbers differ from Table 2? Can Fig. 4(b) add the ground-truth future and indicate time steps?

2 comments
AuthorsOpenReview ↗

Official Response to Reviewer LKyB (1/2)

W1 / Q1. Baselines and explicit futures

We distinguish between learning an effective future-to-action cascade and establishing whether explicit future prediction is universally necessary. Our work addresses the former.

Vision-supervised baselines. Thank you for your advice, we agree that vision baseline should be an important component of our central claim and should be placed at the main table. Thus, in addition to the controlled teacher-future and GT-future supervision comparisons in Sec. 3.2, the revised Table 2 now includes the vision-supervised baseline results alongside DirectWAM, making this comparison explicit in the main results.

Action-only baselines. Our scope is efficient conversion of pretrained cascade WAMs while preserving their future-to-action interface. Establishing the general merits of cascade versus action-only architectures requires a broader study. Appendix B.6, Table 11 provides the numerical Fast-WAM comparison, complementing Fig. 4(a). In an additional LIBERO-Plus Long evaluation, DirectWAM achieves 69.4%, compared with 40.2% for Fast-WAM and 67.3% for FasterWAM. We leave this additional benchmark outside the main paper to maintain its focus on cascade conversion. These results provide broader context, but architectural and training differences prevent them from isolating the control benefit of retaining a future pathway. Importantly, retaining the future provides an inspectable and intervenable prediction interface. Even without reconstruction anchoring (λv=0\lambda_v=0), our additional LIBERO visualizations show clearly recognizable future frames. A decodable visual future supports monitoring and debugging of the model's predictions, while Appendix C demonstrates direct interventions on future representations and their downstream action consequences. These are practical benefits of preserving the cascade interface within our research scope.

W2 / Q2. Flash-WAM comparison and training budget

Inference budget. Flash-WAM's cited 85.54% RoboTwin / 96.1% LIBERO results use 1 video / 2 action steps. Our comparison consistently uses 1/1, for which Flash-WAM officially reports 81.41% / 95.1%.

RoboTwin reproduction protocol. We use the officially released Flash-WAM checkpoint without retraining or fine-tuning. Both methods are evaluated on all 50 RoboTwin tasks under demo_clean and demo_randomized, with 50 episodes per task per split (2,500 per split), seeds starting at 10000, and 1/1 inference. Flash-WAM achieves 78.2% average success under this common protocol. Since the exact seeds and settings underlying its reported 81.41% are not fully specified, we report our common-protocol evaluation rather than claim an exact reproduction of that number.

LIBERO reproduction protocol. All four LIBERO entries for Flash-WAM in the revised Table 2 now come from our own training using its released distillation implementation, replacing the previously quoted results. We evaluate Long, Goal, Object, and Spatial using 100 closed-loop episodes per task and 1/1 video/action inference, as for DirectWAM. The suite-wise scores and average are reported directly in Table 2; Appendix B.5 describes the reproduction protocol.

Training budget. For LIBERO, DirectWAM and the learned single-step baselines are trained jointly across the four suites with an effective batch size of 16 and 1,000 optimization steps. DirectWAM's staged schedule allocates 600 steps to Stage 1 and 400 steps to Stage 2; our Flash-WAM reproduction uses 1,000 distillation steps. The comparison thus matches optimizer-step counts and effective batch size, without assuming equal training compute per step. The RoboTwin main-table DirectWAM checkpoint uses 8,000 Stage-1 + 6,500 Stage-2 optimization steps. Flash-WAM's released RoboTwin configuration specifies a maximum of 10,000 training steps, but the actual training duration of its released checkpoint is not explicitly reported; our evaluation adds no training. Naive truncation also receives no additional training: it isolates the effect of reducing inference steps and is not a training-budget-matched baseline.

Architectural scope. Flash-WAM also validates its distillation method on Motus. Our experiments focus on converting the pretrained LingBot-VA cascade while preserving its future-to-action interface. Extending DirectWAM to additional architectures requires adapting and validating the corresponding training and inference interfaces. Our results establish the method's effectiveness in the evaluated cascade, without claiming architecture-independent generality.

AuthorsOpenReview ↗

Official Response to Reviewer LKyB (2/2)

W3 / Q3. Figure provenance, evaluation protocols, and visual evidence

We strongly disagree with the characterization that our figures are “unreliable.” This is a serious assessment that is not substantiated by the points raised. The review conflates photographs of our experimental setup with generic imagery, schematic trajectory overlays with experimental traces, and results obtained under explicitly disclosed evaluation protocols. It also overlooks the quantitative visual-quality evidence already provided in the appendix. We address each point below with its source in the manuscript.

Figure 1: provenance and purpose. The robot thumbnails and inset background originate from photographs of our own Galaxea A1 experimental setup, described in Sec. 4.2 and shown in Fig. 5, rather than unrelated stock scenes. We used Photoshop to remove background clutter for clarity. The accusation of AI-generated experimental imagery is not substantiated by the appearance or reuse of these photographs. The overlaid trajectories are schematic illustrations of trajectory overlap, not recorded rollout traces; the actual trajectory evidence is presented in Sec. 3.1, Fig. 2(b), and Appendix A.3.

Figure 1: numerical results. The cited numbers are not inconsistent measurements of the same evaluation. The 91.6% and 57.6% results appear explicitly in Sec. 3.2 and Fig. 3(a), using the 500-episode diagnostic protocol; 91.6% also appears in Appendix A.3, Table 5. In contrast, 90.7% and 54.9% in Table 2 are Clean/Randomized averages from the larger evaluation. Sec. 4.1 already states that Table 2 uses 50 episodes per task, whereas the other success-rate evaluations use 10. These values correspond to distinct, documented evaluation protocols and are therefore not expected to coincide.

Figure 4(b): qualitative and quantitative evidence. The reviewer's concern that Fig. 4(b) is uninformative is not established by the observation that its columns are “almost indistinguishable.” This preservation of overall visual appearance is precisely what we intend to convey: action supervision does not cause severe visual degradation, as explicitly stated in Sec. 4.1, “Action-supervised futures remain visually decodable.” The revised Fig. 4(b) now includes ground-truth futures and temporal labels, making the qualitative comparison easier to assess. Quantitative visual-quality results are reported in Appendix C.2, Table 14, supporting the conclusion that action-supervised futures retain decodable visual structure. The evidence therefore extends beyond the thumbnail comparison.

We have improved the presentation by adding ground-truth futures and temporal labels. These presentation improvements do not substantiate the review’s broader concern of unreliable experimental evidence.

Reviewer_L4cNOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DirectWAM, which trains a single-step future predictor through downstream action supervision while preserving the future-to-action cascade. Experiments in simulation and on real robots report improved control performance over evaluated single-step baselines and approximately 28× faster model-side inference than the iterative teacher.

Critical strengths
  1. The method is relatively simple and yields practically meaningful improvements under a constrained inference budget.

    DirectWAM retains the existing cascade interface and trains the future predictor through downstream action error, resulting in a straightforward optimization objective. In Table 2, average RoboTwin success increases from 78.2% for the evaluated Flash-WAM baseline to 85.4%, an improvement of 7.2 percentage points. Under randomized evaluation, success increases from 74.2% to 83.7%. These results indicate that the training approach has practical value in the evaluated settings and merits further validation.

Critical weaknesses
  1. The methodological novelty needs clearer justification. Flash-WAM already enables single-step inference in both streams, and related methods incorporate action supervision. Although DirectWAM differs from these approaches, matched ablations are needed to establish the benefits of its frozen action expert and two-stage training.

  2. The experimental evidence needs stronger controls. The unmatched LIBERO comparison, unreported training-seed variability, and missing component ablations limit attribution of the gains. The small real-world evaluation supports feasibility but does not establish equivalence to the teacher.

Questions for authors
  1. Novelty: How does DirectWAM differ from action-regularized distillation or joint policy learning, particularly in its gradient pathway?
  2. Ablations: Under matched training conditions, how do freezing the action expert and varying the reconstruction weight affect performance?
  3. Reliability: How many independent training runs were evaluated, and are the Appendix C evaluation contexts disjoint from training data?
2 comments
AuthorsOpenReview ↗

Official Response to Reviewer L4cN (1/2)

We thank the reviewer for the careful review and constructive questions.

Q1. Clarifying the contribution and gradient pathway

Our contribution goes beyond a naive change in training recipe: we systematically investigate what a predicted future must preserve for robot control, derive the principle of action sufficiency, and validate it through improved control performance. Specifically, we reveal that video stochasticity has limited downstream impact, iterative refinement remains critical, and representation errors of similar magnitude have highly unequal consequences for action accuracy. These findings motivate a fundamental shift in the learning criterion for WAMs: the visual stream should be optimized for the actions its predictions enable, rather than generative fidelity alone. Through experiments, DirectWAM outperforms Flash-WAM at the same single-step inference budget. The two-stage schedule, frozen action expert, and optional reconstruction anchor are implementation choices rather than our central claimed novelty.

Methodologically, DirectWAM backpropagates action error through the video predictor's own output. Joint training can sum video and action gradients on shared parameters while the action branch still uses demonstrated futures. We compute action loss from actions induced by the student's predicted future and backpropagate it into the video predictor, directly asking the question: what makes a predicted future useful for downstream control. The following schematic illustrates this conditioning difference:

\text{Flash-WAM:}\quad \mathcal L_{\mathrm{action}} &=\ell\\!\left(G_\psi(z_{\mathrm{GT}}),a^\star\right),\\\\ \text{DirectWAM:}\quad \mathcal L_{\mathrm{action}} &=\ell\\!\left(G_\psi(F_\theta(h_t)),a^\star\right). \end{aligned}$$ Here, $z_{\mathrm{GT}}$ is the demonstrated future latent, $h_t$ the conditioning context, $F_\theta$ the future predictor, $G_\psi$ the action model, and $a^\star$ the reference action target in this schematic; $\ell$ denotes the action loss. The distinction therefore lies in the cascade's supervision interface, rather than the particular form of its losses. DirectWAM explicitly places the student's generated future inside the downstream action-loss computation, so the video predictor is trained on the action consequences of its own predictions, rather than merely receiving gradients from jointly optimized modality-specific losses. ### Q2. Ablations We report separate ablations of the training schedule, action-expert configuration, and reconstruction weight in the revised Appendix B.2–B.3. LIBERO training jointly uses all four suites, with an **effective batch size of 16 and 1,000 total steps (600 Stage-1 + 400 Stage-2 steps for staged training)**; LIBERO-Long denotes the evaluation suite for the reported ablations. The **1,000-step joint run is the budget-matched comparison**; additional joint runs at 400 and 1,600 steps examine the effect of training duration. RoboTwin staged training uses **8,000 Stage-1 + 6,500 Stage-2 steps**, with a budget-matched 14,500-step joint run. **Freezing and co-training the action expert.** Appendix B.3, Table 8 reports the effects of freezing, unfreezing, and co-training the action expert. These comparisons support freezing as a practical choice in the evaluated settings, without establishing that it is universally preferable. **Reconstruction weight.** Appendix B.3, Table 9 shows different responses to reconstruction regularization across benchmarks: a nonzero weight helps on RoboTwin, while LIBERO-Long works well without an anchor. We therefore select $\lambda_v$ empirically per benchmark. **Two-stage training.** To address the related concern in W1, Appendix B.2, Table 7 shows that staged training gives more consistent performance across the tested LIBERO-Long checkpoints, while joint training reaches comparable performance on RoboTwin at the same total step budget. We regard these configurations as empirical, potentially task-dependent choices; our central argument concerns learning predicted futures through their downstream action consequences, rather than the optimality of a particular training recipe.
AuthorsOpenReview ↗

Official Response to Reviewer L4cN (2/2)

Q3. Training-seed variability and evaluation-data separation Training-seed stability. Appendix B.4, Table 10 reports three independent 6,500-step Stage-2 training runs from the same 8,000-step Stage-1 checkpoint. Mean success rates are 85.6 ± 1.2% on RoboTwin clean and 83.9 ± 0.8% on randomized evaluation, indicating limited variation across seeds for this Stage-2 recipe. Each seed is evaluated on 50 tasks with 10 episodes per task per split (500 episodes per split), whereas Table 2 uses 50 episodes per task (2,500 per split); the seed means therefore summarize a separate evaluation. These runs measure variability conditional on the shared Stage-1 initialization. Data separation. The clean50 contexts used for the learned-prediction analyses in Appendix C are disjoint from the Stage-2 training data. Stage-2 optimization uses aug500 and excludes clean50, while the analysis contexts are drawn from the separate clean50 demonstrations. We clarify this split at the beginning of Appendix C: these analyses evaluate contexts that were not used to optimize the Stage-2 future predictor.

Matched LIBERO reproduction. The revised Table 2 replaces quoted Flash-WAM LIBERO results with our own training using its released distillation implementation. Flash-WAM and DirectWAM are trained jointly across all four suites with an effective batch size of 16 and 1,000 total optimization steps, and evaluated with 100 closed-loop episodes per task at the same 1/1 inference budget. Appendix B.5 describes the reproduction protocol.

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

Related papers

Discussion (0)

Sign in to comment.