acceptodds
Under review as a conference paper at ICLR 2027

Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction

Abstract

Action-conditioned world models must predict the consequences of proposed controls quickly enough to support repeated queries. We introduce Beckmann World Models (BWM), which learn a direct map from noise, visual history, and actions to future RGB frames without a pretrained generative teacher. Beckmann transport supplies a conservation-based training rule for this terminal map using sampled noise–future pairs, without a population of generated negatives. However, learning the map on interpolated inputs does not ensure accurate prediction from the pure noise used at deployment. We rectify this source-endpoint gap by proposing endpoint-aligned training: direct supervision of the source query, motion and temporal losses, and spatial feature supervision, supported by a generated-history regularizer. These additions train the same RGB map and preserve one network call per future chunk. On PushT and Robomimic Can, BWM reduces 64-frame MSE by – relative to same-backbone regression at comparable inference cost, with measured latency of – ms per native chunk. It also improves prediction over locally trained DriftWorld and policy-performance estimation, while the planning comparison shows no detected improvement. Cumulative training costs are disclosed because these are comparisons of complete recipes. The results identify a practical route from autonomous transport to accurate one-call world prediction. Anonymous project page: https://beckmann-world.pages.dev/https://beckmann-world.pages.dev/.

Primary area:
generative models
Keywords:
Diffusion, World Models, Efficiency, Robotics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_izE4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It seems the paper is proposing a a new world model called Beckmann World Models based on Beckmann Transport Models (Lee et al., 2026).

Critical strengths

None

Critical weaknesses

This paper is very hard to read and full of non-standard terminology. The method section is extremely unclear, and the contribution is vague. Section 3 (Method) reads as though it was AI-generated from the implementation: the problem formulation does not actually define the problem; the language describing the method is non-standard; hyperparameters are hard-coded into the method; and several quantities and parameters are undefined, e.g., Lmove\mathcal{L}_\text{move}. The experiments are also unclear and very limited: only two tasks are considered, the baselines are unclear, there is no downstream task, and the goal of the experiments is unclear.

The authors do not appear to have spent sufficient time writing the paper, and the current submission is unusually unclear. Finally, I would encourage the authors to submit only once the manuscript has been carefully prepared.

Questions for authors

None

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

This paper introduces Beckmann World Models (BWM), an action-conditioned visual prediction method that generates future RGB frame chunks with a single network evaluation. It adapts the terminal-map learning framework of Beckmann Transport Models (BTM) to conditional video prediction.

The principal methodological addition is endpoint-aligned training, which addresses the discrepancy between mixed noise–data inputs during transport training and pure-noise inputs at inference. The approach combines direct source supervision with motion-weighted, temporal, spatial-feature, and generated-history objectives.

Experiments on PushT and Robomimic Can demonstrate 45–49% lower 64-frame MSE than a same-backbone regression baseline. BWM also outperforms locally trained DriftWorld on the reported prediction metrics and improves policy-performance estimation, although no significant improvement in planning performance is established.

Critical strengths
  1. Strong empirical improvements in prediction accuracy at low inference cost.

BWM achieves substantial reductions in autoregressive prediction error over the same-backbone regression baseline: 48.8% on PushT and 45.1% on Robomimic Can. The improvements extend to moving-region error and LPIPS, while inference requires only one network evaluation per native future chunk.

These results provide meaningful evidence that the complete training recipe improves the accuracy–latency tradeoff in the evaluated settings.

  1. A useful diagnosis of the deployment-endpoint mismatch.

The distinction between prediction quality on interpolated inputs and on the actual pure-noise inference inputs is an interesting practical observation. The reported source-endpoint error gap, together with the source-supervision experiments, provides useful insight into training direct prediction maps.

The paper also appropriately distinguishes prediction improvements from downstream planning performance and discloses important limitations concerning training budgets, checkpoint provenance, and qualitative sample selection.

Critical weaknesses

1. The independent contribution of the Beckmann objective is not established.

The terminal-map learning framework is inherited from BTM, and one-step action-conditioned prediction already exists in prior work. The contribution therefore depends substantially on whether Beckmann learning adds value beyond the endpoint supervision and auxiliary losses. Equation (11) combines these components, while Table 2 evaluates cumulative training stages rather than isolating the Beckmann term. Since source-endpoint MSE also penalizes output variation across noise samples, stronger supervised regression remains a plausible explanation for the gains. The decisive experiment would remove only LBTML_{\mathrm{BTM}}, retaining all other losses and matching initialization, optimization settings, and training budget.

  1. Unmatched training budgets confound the comparisons and ablations.

On PushT, the final BWM accumulates approximately 202,700 updates, compared with 39,350 for local DriftWorld; other training settings also differ. Update counts alone do not measure compute, but this disparity makes budget-controlled comparisons essential. Moreover, each cumulative ablation introduces additional objectives alongside further training. Comparisons matched by GPU-hours or data exposure, together with controls that continue the preceding recipe for the same additional budget, are needed to attribute improvements to the proposed method.

Questions for authors
  1. Requested controlled experiment: What happens when only LBTML_{\mathrm{BTM}} is removed, retaining source supervision and all auxiliary losses under the same initialization, optimization settings, and training budget? This result would materially affect my assessment of the central contribution.

  2. Budget clarification and requested controls: How do the complete recipes compare in cumulative GPU-hours and data exposure? Do the gains persist under matched budgets, and do the cumulative ablations outperform continuing the preceding recipe for the same additional training budget without adding new losses?

2 comments

Thank you for your constructive review. We conducted two additional experiments addressing the independent contribution of the Beckmann objective and the unequal cumulative training exposure.

1. Controlled Beckmann-objective ablation

To isolate the contribution of Beckmann training, we compare the full objective against a variant omitting the Beckmann objective while retaining identical auxiliary supervision, initialization, and training-data exposure.

LPIPS (lower is better):

ModelPushTCan
Full rolloutBWM0.1150.0447
w/o BTM0.1190.0578
60-frame rolloutBWM0.08500.0117
w/o BTM0.09230.0247

The Beckmann objective adds terminal-map consistency and boundary constraints beyond the retained endpoint-aligned supervision. With identical auxiliary losses, initialization, and training exposure, retaining the stabilized BTM objective reduces Can LPIPS by 52.6% on 60-frame rollouts and 22.7% on full rollouts, with corresponding improvements of 7.9% and 3.4% on PushT.

These results directly demonstrate that the Beckmann objective contributes beyond source regression and auxiliary regularization, rather than the improvements arising solely from stronger supervision or additional training.

2. Matched training budgets and efficiency

To address the cumulative training-budget confound, we additionally trained the complete RGB BWM recipe from scratch for five epochs, matching the local baselines' data exposure, training split, and evaluation protocol.

Five-epoch comparison:

MethodPushT MSEPushT LPIPSCan MSECan LPIPS
BWM0.02110.07800.004770.0153
MSE U-Net0.03030.08040.005100.0104
DriftWorld0.03140.07900.01040.0682
GPC0.06300.16090.02760.0589
AVDC0.06840.18600.04000.1190

The fresh BWM models use global batch sizes of 4 and 2, with training exposure of 157,400 and 40,000 primary windows for PushT and Robomimic Can, respectively. Different global batch sizes explain differences in update counts across methods.

Even with fixed data exposure and the same five-epoch training schedule, BWM reduces MSE relative to DriftWorld by 32.8% on PushT and 54.1% on Can, with 77.6% lower Can LPIPS. Relative to the same-backbone MSE regressor, BWM also achieves lower MSE on both tasks (30.4% and 6.5%, respectively), although regression retains better Can LPIPS in this short-budget comparison.

Training efficiency. Beyond matching data exposure, our measured complete-update profiling on Can demonstrates a substantial training-resource advantage over local DriftWorld:

  • 3.62 vs. 1.09 training windows/s — 3.3× higher throughput.
  • 7.77 vs. 36.07 GiB peak memory/GPU — 78% lower memory.

These measurements use equal primary-window batches on two A800 GPUs and include the JVP, generated-history recovery, backward pass, optimizer, and EMA. They profile the motion stage before final VGG supervision. Unlike DriftWorld's condition-specific generated-negative population, BWM uses sampled noise–future pairs and JVPs for its transport objective.

Our cumulative Can training records further show that the motion BWM checkpoint reaches LPIPS 0.0090 after 4.32 training-loop hours, versus 0.0682 after 10.19 hours for DriftWorld on the same two-A800 setup, including inherited training time. These are measured complete-recipe comparisons.

Together, these results address both central concerns: the objective-omission ablation demonstrates the independent contribution of the stabilized Beckmann objective under matched auxiliary supervision, while the fresh five-epoch comparison shows that BWM's predictive advantages do not depend on its original extended training lineage. The measured throughput and memory improvements further support the practical training efficiency of the approach.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my main concerns.

The controlled BTM-objective ablation provides meaningful evidence of its independent contribution, while the five-epoch comparison and training-efficiency measurements substantially address my concerns about unequal training budgets.

Some limitations remain, particularly the incremental methodological novelty, differences in computational costs, and the lack of multiple training seeds. Nevertheless, the additional experiments significantly strengthen the paper and address my primary concerns.

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

This paper presents Beckmann World Models (BWM) for one-step action-conditioned RGB prediction. It learns a terminal map using pair-based stopped-Jacobian updates, without a pretrained generative teacher or a conditional population of generated negatives. The paper identifies a gap between reconstruction at the data boundary and prediction from the pure-noise inputs used at deployment. The method adds direct supervision for predictions from noise, motion and temporal losses, training with generated histories, and a loss based on frozen visual features. On PushT and Robomimic Can, it improves prediction accuracy with similar inference time. It also improves policy evaluation, but shows no clear improvement in planning.

Critical strengths

1.The paper identifies a clear gap between training and prediction from pure noise. The experiments support the need for direct supervision at the noise input.

2.BWM improves prediction accuracy on both tasks with similar inference time.

3.The paper clearly reports training costs, limitations, and the lack of improvement in planning.

Critical weaknesses

1.The benefit of the Beckmann objective is unclear. The paper does not compare it with a regression model using the same auxiliary losses and a matched training budget. The ablations also add losses while continuing training, so the gains may partly come from extra training.

2.Better image quality does not necessarily lead to better action selection. The paper improves policy evaluation, but shows no clear improvement in the tested planning setting. Further analysis of task-relevant errors would help explain the value of the prediction gains.

3.The tests cover only two mostly deterministic tasks. The source loss reduces variation across noise inputs, but the paper does not test cases with multiple possible futures. Its performance in such settings remains unclear.

Questions for authors

1.Can the authors analyze which task-relevant errors affect action ranking? Why does better image prediction not lead to a clear gain in the tested planning setting?

2.Can the authors compare BWM with a regression model using the same auxiliary losses and a matched training budget?

1 comment

Thank you for your constructive review and for recognizing the source-endpoint mismatch and efficiency of our one-step prediction approach. We conducted additional experiments addressing your concerns.

1. Independent contribution of the Beckmann objective (W1/Q2)

We directly address the concern that our improvements may arise solely from auxiliary supervision or additional training. We compare full BWM against a variant removing the Beckmann objective while retaining all auxiliary losses, including source, motion, temporal, spatial-feature, and generated-history supervision. Both variants use identical initialization and training-data exposure.

LPIPS (↓):

SettingPushT BWMPushT w/o BTMCan BWMCan w/o BTM
Full rollout0.1150.1190.04470.0578
60-frame rollout0.08500.09230.01170.0247

Retaining the Beckmann objective reduces LPIPS by 52.6% on Can's 60-frame rollout and 22.7% on its full rollout, with corresponding improvements of 7.9% and 3.4% on PushT. Crucially, the supervised control retains the complete auxiliary recipe. These results therefore provide direct evidence that the Beckmann objective contributes beyond endpoint regression and auxiliary regularization, rather than the gains being attributable solely to additional supervision.

We also trained BWM from scratch for five epochs, matching the baselines' dataset exposure. Without the original extended training lineage, BWM achieves MSE 0.0211 versus 0.0314 for DriftWorld and 0.0303 for regression on PushT, and 0.00477 versus 0.0104 and 0.00510, respectively, on Can. This demonstrates that the prediction improvements do not require the original long continuation schedule. Beckman objective adds terminal-map consistency and boundary constraints beyond the retained source and spatial-feature supervision. We hypothesize that these additional constraints can improve perceptual structure.

2. Generalization to real-world robot videos (W3)

We further extended our evaluation beyond PushT and Can to RT-1 and Bridge-V2 real-world robot videos at 256×256 resolution. Using a hybrid latent-space terminal-map model, we compare generated-history recovery against an equal-update control initialized from the same parent checkpoint.

DatasetVariantLPIPS ↓FID ↓FVD ↓
RT-1Control0.093320.56172.89
RT-1+ Recovery0.06128.05104.91
Bridge-V2Control0.110630.98171.56
Bridge-V2+ Recovery0.084121.34117.20

Recovery supervision reduces LPIPS by 34.4% on RT-1 and 24.0% on Bridge-V2, alongside substantial FID and FVD improvements. These results demonstrate that our generated-history training strategy extends to more diverse, real-world robotic video prediction, beyond the original simulation benchmarks.

3. Task-relevant prediction and downstream utility (W2/Q1)

Our primary contribution is accurate, efficient action-conditioned world prediction, not a new action-ranking algorithm. We nevertheless evaluate downstream utility to determine whether prediction improvements extend beyond perceptual metrics.

On PushT, BWM reduces moving-region MSE by 52.8% relative to local DriftWorld (0.4093 → 0.1933), demonstrating improved prediction of changing robot/object regions rather than merely static backgrounds. Our action-swap diagnostics further confirm that predictions remain sensitive to the supplied controls.

More importantly, these improvements translate into more accurate policy-performance estimation: across seven policies with 300 trials each, BWM achieves policy-IoU MAE 0.0100, compared with 0.0306 for local DriftWorld and 0.0121 for MSE regression.

In our updated PushT GPC-RANK evaluation, BWM also achieves executed IoU 0.721 versus 0.709 for local DriftWorld. However, the paired difference (+0.012, 95% CI [−0.047, +0.071]) does not establish a statistically significant planning improvement.

These evaluations measure distinct properties: policy-performance estimation assesses predicted outcomes across policies, whereas planning requires correctly ordering candidate actions at individual decision states. Our demonstrated downstream contribution is improved policy-performance estimation, not a claimed improvement in action ranking. We report the planning result transparently without conflating these two objectives.

Reviewer_Z624OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Beckmann World Models (BWM), a one-call action-conditioned RGB world model. Starting from Beckmann terminal-map transport, BWM directly maps source noise, visual history, and future actions to a future video chunk, trained with a pair-based stopped-Jacobian terminal-map loss. The paper argues that this objective is misaligned with deployment because training emphasizes interpolated/data-boundary inputs while inference starts from pure noise. It therefore adds direct source anchoring, motion and temporal losses, frozen VGG spatial supervision, and a generated-history recovery regularizer. On PushT and Robomimic Can, the complete staged recipe improves autoregressive prediction metrics relative to local one-call and diffusion controls at comparable one-call latency, and improves policy-performance estimation, although it does not improve the reported planning benchmark.

Critical strengths

S1: The paper identifies a precise and practically relevant source-endpoint mismatch in terminal-map training. Demonstrating that good data-boundary reconstruction does not imply good pure-noise prediction is an important diagnostic, and source anchoring is a simple response targeted at the actual deployed query.

S2: Applying the Beckmann terminal-map formulation to action-conditioned RGB world prediction is a meaningful conceptual contribution. The pair-based, stopped-Jacobian objective provides a training route that does not explicitly learn a velocity field and then numerically solve it, require conditional negative populations, or require a pretrained generative teacher, while preserving one-call inference.

S3: The empirical presentation is unusually transparent in several respects. The paper separates same-split and released baselines, reports native latency, discloses inherited continuation costs, states that planning does not improve, and includes endpoint/action-swap probes rather than relying solely on selected videos.

S4: The full recipe produces substantial prediction improvements on both reported manipulation benchmarks at essentially the same inference cost as the MSE and DriftWorld U-Nets. The staged ablation gives useful evidence that source anchoring, motion/recovery, and spatial supervision each contribute to the final result.

Critical weaknesses

W1: The method's distinct mechanism and novelty relative to existing one-step flow-map / MeanFlow paradigms are not empirically established sufficiently. The related-work discussion explicitly positions BWM as a time-independent, autonomous terminal-map-conservation approach with no interpolant-time input to the network; this is a reasonable conceptual distinction. On the other hand, BWM's deployed map can be written as Tθ(x,c)=x+Uθ(x,c)T_\theta(x,c)=x+U_\theta(x,c); at the one-step source-to-endpoint interface, if one only models the average displacement/velocity to the endpoint, UθU_\theta differs only by a remaining-time scale. The paper provides no matched, same-recipe control against time-conditioned MeanFlow, endpoint-average-velocity, or consistency/flow-map methods. It is therefore unclear whether the gain comes from the Beckmann autonomous terminal-map/chord-JVP constraint or from a reparameterizable one-step endpoint predictor plus stronger supervision.

W2: The evidence also does not isolate the claimed Beckmann mechanism from the larger endpoint-supervision recipe. The final model combines the terminal-map objective with direct source MSE, motion and temporal losses, generated-history recovery, frozen VGG features, staged training, and changed optimization/exposure. Since source MSE against a nearly deterministic recorded future both encourages a conditional mean and suppresses source-induced variance, the results could substantially reflect supervised regression and rollout regularization rather than terminal-map conservation. The paper needs controls that keep the complete auxiliary recipe fixed while replacing or removing the Beckmann objective.

W3: The central comparisons are not matched in training budget or optimization. BWM inherits a long staged lineage (about 202.7k PushT updates), while scratch baselines run five configured epochs; the paper also reports different optimizers, batch sizes, negative-sample counts, and exposure. The authors properly disclose this, but the resulting comparisons establish a better complete recipe, not a clean quality/efficiency advantage of BWM training. Inference latency is matched for one-call models; training-efficiency and method-attribution claims are substantially less secure.

W4: The theory-to-algorithm connection is not closed. The ideal terminal-conservation identity JxTc(x)bc(x)=0J_xT_c(x)b_c(x)=0 relies on BTM's support and regularity assumptions. The paper states that, for the linear interpolant xs=(1−s)z+syx_s=(1-s)z+sy, the chord d=y−zd=y-z estimates a derivative along the interpolant's conditional field, and it uses a fixed 0.25JxTθd0.25J_xT_\theta d stopped target. This does not by itself give the additional identification needed to relate that interpolant-conditioned direction to the autonomous field bc(xs)b_c(x_s) in the identity. The paper does not prove this identification, a fixed-point/consistency property, or SGD convergence for the actual surrogate objective, and provides neither a theoretical origin nor a sensitivity/stability range for 0.250.25. Whether the terminal-conservation interpretation survives after adding the auxiliary supervision is likewise unclear. This does not negate the theoretical BTM foundation under its explicit assumptions.

W5: All recipes use one training trajectory. Bootstrap intervals over episodes for fixed checkpoints do not address seed variation, data order, hyperparameter/checkpoint selection, or instability of the staged continuation. This is especially consequential for the 30-episode Can evaluation and for claimed gains between adjacent ablation stages.

W6: The scope of the control evidence is limited. The action-swap probe and policy-performance evaluation across seven policies provide useful evidence of action sensitivity and some downstream value. However, the two benchmarks remain low-resolution, single-view, largely deterministic manipulation datasets evaluated using recorded action sequences. A fixed-window action swap does not establish calibrated counterfactual action prediction or uncertainty handling, and the existing fixed-proposal planning test shows no improvement. The stronger practical implication should therefore be prediction and policy-performance estimation, rather than validated planning or robust closed-loop utility.

Questions for authors

Q1: Can the authors provide rigorously matched comparisons with time-conditioned MeanFlow, endpoint-average-velocity, or consistency/flow-map baselines? These controls should use the same architecture, source/motion/temporal/VGG/generated-history recipe, optimizer, and total compute. In particular, when generating one step from the source, what non-reparameterizable benefit does BWM's residual Uθ=Tθ−xU_\theta=T_\theta-x provide over an average displacement to the endpoint?

Q2: Can the authors, while holding the full auxiliary recipe fixed, compare source-anchored regression, the final recipe without the Beckmann term, and a version that adds only the Beckmann loss? This would isolate the terminal-conservation constraint from standard supervision and rollout regularization.

Q3: The paper describes the chord d=y−zd=y-z as estimating a directional derivative along the interpolant's conditional field; what additional conditions connect this quantity to the unknown autonomous field bc(xs)b_c(x_s) in Eq. 2? Please state the identification conditions needed to pass from the BTM assumptions to this surrogate update, and provide ablations/stability analyses for the JVP coefficient (especially the fixed 0.250.25), interpolation distribution, and stopped-target design. Can the paper give a fixed-point, consistency, or convergence property for its actual surrogate objective?

Q4: Can the main MSE/LPIPS/moving-region comparisons be repeated under matched total compute, data exposure, architecture, optimizer tuning budget, and training duration for BWM, MSE regression, and DriftWorld? Please report full training cost, not only inference latency, for the matched runs.

Q5: Can the authors report multiple independent training seeds and confidence intervals for the final models and every major ablation? The existing fixed-checkpoint episode bootstraps, including the paired Can table, are useful but do not replace an estimate of training randomness; please add uncertainty across runs for both 64-frame and full-episode outcomes.

Q6: Beyond the existing action-swap and fixed-proposal downstream tests, can the authors evaluate counterfactual action fidelity and downstream use more directly: e.g., action-sequence intervention tests against simulator ground truth, calibrated uncertainty under ambiguous futures, and a planning benchmark with enough proposals/seeds to resolve whether improved prediction can improve decisions?

Q7: Beyond the major theoretical controls, please provide sensitivity analyses for the gradient-balancing clip and source/motion/recovery/VGG weights and ramps, and clarify whether the method remains effective on a more stochastic or multi-modal prediction setting where source MSE's variance suppression may be undesirable.

1 comment

Thank you for your thoughtful and detailed review. We appreciate your recognition of the source-endpoint mismatch, the conceptual contribution of conditional terminal-map learning, and our empirical evaluation. We conducted additional experiments and would like to clarify three points.

1. Independent contribution of the Beckmann objective (W2/Q2)

To distinguish the Beckmann objective from the auxiliary supervision, we conducted a controlled ablation while retaining the complete source, motion, temporal, spatial-feature, and generated-history supervision**, with identical initialization and matched training exposure.

LPIPS (↓):

SettingPushT BWMPushT w/o BTMCan BWMCan w/o BTM
Full rollout0.1150.1190.04470.0578
60-frame rollout0.08500.09230.01170.0247

Retaining the Beckmann objective reduces LPIPS by 52.6% on Can's 60-frame rollout and 22.7% on its full rollout, with corresponding improvements of 7.9% and 3.4% on PushT.

Because both variants retain the same endpoint-aligned auxiliary supervision, this directly demonstrates an incremental benefit from the stabilized Beckmann objective beyond supervised regression and auxiliary regularization. The ablation isolates the complete BTM objective, including its transport and boundary terms. Beckman objective adds terminal-map consistency and boundary constraints beyond the retained source and spatial-feature supervision. We hypothesize that these additional constraints can improve perceptual structure.

2. Training-exposure-controlled comparison (W3/Q4)

We additionally trained BWM from scratch for five epochs, removing the advantage of its original extended continuation. All compared methods use the same training split, validation protocol, and primary-window exposure: 157,400 windows for PushT and 40,000 for Can.

MethodPushT MSEPushT LPIPSCan MSECan LPIPS
Fresh BWM0.02110.07800.004770.0153
MSE U-Net0.03030.08040.005100.0104
DriftWorld0.03140.07900.01040.0682

Even without inherited training, BWM reduces MSE relative to DriftWorld by 32.8% on PushT and 54.1% on Can, and relative to regression by 30.4% and 6.5%, respectively.

These results show that the predictive advantages do not depend exclusively on the original longer training lineage. This comparison controls dataset exposure, although optimizer choices and total training FLOPs are not identical.

3. Autonomous-field identification (W4/Q3)

We clarify how the sampled interpolant chord relates to the autonomous field in Eq. (2).

For a fixed condition c=(h,a)c=(h,a), let ZZ be source noise, YY the recorded future, and S∼U[0,1]S\sim U[0,1], independent of the source-target pair. Define

X=(1−S)Z+SY,D=Y−Z.X=(1-S)Z+SY,\qquad D=Y-Z.

The time-independent interpolant field underlying BTM is

bc(x)=E[D∣X=x,c].b_c(x)=\mathbb E[D\mid X=x,c].

Consequently,

=J_xT_\theta(x,c)b_c(x).$$ Thus, the sampled chord provides a conditional estimator of the autonomous direction used in the terminal-map conservation equation. This establishes the population-level identification under the corresponding BTM assumptions, rather than requiring a separately learned velocity field. We distinguish this identification from a convergence guarantee for the complete implemented objective, whose stabilized weighting and auxiliary losses extend beyond the original ideal conservation analysis. **Additional real-world evaluation (W6).** We also evaluated a hybrid latent-space extension on RT-1 and Bridge-V2 real-world robot videos at 256×256 resolution. Against equal-update controls, generated-history recovery reduces FVD from **172.89 to 104.91 on RT-1** and **171.56 to 117.20 on Bridge-V2**, with LPIPS improvements of 34.4% and 24.0%, respectively. These results extend evidence for the recovery mechanism to real-world video.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.