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); at the one-step source-to-endpoint interface, if one only models the average displacement/velocity to the endpoint, Uθ 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)=0 relies on BTM's support and regularity assumptions. The paper states that, for the linear interpolant xs=(1−s)z+sy, the chord d=y−z estimates a derivative along the interpolant's conditional field, and it uses a fixed 0.25JxTθd stopped target. This does not by itself give the additional identification needed to relate that interpolant-conditioned direction to the autonomous field bc(xs) 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 0.25. 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.
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):
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:
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:
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.