- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Motion-Guided Asymmetric Prediction (MoGAP) to improve visual prediction supervision in World Action Models (WAMs). The method identifies dynamic regions using temporal differences between video latents, applies stronger noise to dynamic regions and weaker noise to static regions, and trains the model to predict dynamic regions using relatively clean static context. MoGAP is incorporated into existing WAMs through a training-only auxiliary prediction path without introducing additional inference overhead. Experiments on RoboCasa GR-1 Tabletop, RoboCasa-Kitchen, and real-world robotic manipulation tasks demonstrate improved task success rates and better adaptation to target object relocation during execution.
Critical strengths
(1) Clear research motivation. Rather than simply introducing additional visual supervision, the paper investigates how the spatial allocation of supervision influences robotic control. The motivating study provides empirical support for this research direction.
(2) Meaningful performance improvements. MoGAP improves the average success rate of Cosmos3 from 52.3% to 62.3% on GR-1 Tabletop and also improves performance when applied to Fast-WAM. Comparisons with Geometry Forcing, AGRA, TSA, and VideoREPA further strengthen the empirical evaluation.
(3) Well-targeted ablation studies. The authors examine the auxiliary prediction path, noise asymmetry, and the direction of asymmetric noise allocation. A compute-matched comparison is also provided. These experiments directly support several key components of the proposed method.
Critical weaknesses
(1) Limited methodological novelty. The main technical contribution is concentrated in Eqs. (8)–(11), which adjust noise levels, distinguish dynamic and static regions through binary thresholding, apply asymmetric noise, and introduce a region-selective flow-matching objective. While this constitutes a meaningful training modification, region-wise noise allocation and flow matching have already been investigated in prior work. Therefore, the methodological novelty appears moderate rather than substantial.
(2) Simplistic dynamic-region selection with unclear reliability. Equation (9) uses the spatial mean motion score within each temporal slice as a threshold to distinguish dynamic and static regions. However, motion does not necessarily indicate task relevance, while subtle task-relevant movements may not be reliably identified. Although Table 4a provides evidence that the motion scores distinguish moving regions, the sensitivity of this thresholding strategy to local visual noise, camera perturbations, and threshold selection has not been sufficiently investigated.
(3) Insufficient disentanglement of asymmetric noise allocation and other training components. The dynamic-region loss reweighting approach in Section 2.2 already improves the Cosmos3 success rate from 52.3% to 57.7%, while MoGAP further improves it to 62.3%. Although Table 3 includes relevant ablations, the two approaches also differ in their use of an auxiliary prediction path. Consequently, the additional 4.6 percentage-point improvement cannot be directly attributed to asymmetric noise allocation alone.
(4) Incomplete explanation of the underlying mechanism. Figure 8 and Table 4 show that MoGAP encourages action tokens to attend more strongly to the robot arm and manipulated objects. However, these observations mainly provide correlational evidence. They do not fully explain why this change leads to improved control performance or why asymmetric noise allocation provides greater benefits than dynamic-region loss reweighting alone.
(5) Unclear generality under complex observation conditions. In RoboCasa-Kitchen, the authors introduce a special treatment for the wrist-camera view because camera motion can dominate the latent motion scores. This raises concerns about the generality of the proposed motion-guided region selection strategy under moving-camera conditions and more complex visual environments.
(6) Limited analysis of experimental variability. Although the paper includes extensive simulated and real-world evaluations, it does not sufficiently report performance variability across independent training seeds. Additional repeated training runs would help establish the consistency of the reported improvements, particularly in real-world generalization settings.
Questions for authors
(1) Methodological novelty. Could the authors clarify the technical differences between MoGAP and existing token-wise noise assignment methods? A direct comparison with alternative noise allocation strategies under the same WAM architecture and training conditions would help establish the independent contribution of motion-guided asymmetric prediction.
(2) Reliability of dynamic-region selection. Could the authors investigate the sensitivity of the thresholding strategy in Eq. (9)? Comparisons among mean thresholding, Top-K selection, and continuous soft weighting, together with evaluations under local visual noise and camera perturbations, would help assess the reliability of dynamic-region identification and its effect on downstream control performance.
(3) Sources of performance improvement. Could the authors further disentangle the contributions of dynamic-region loss reweighting, asymmetric noise allocation, and the auxiliary prediction path? For example, applying dynamic-region loss reweighting and asymmetric noise prediction within the same auxiliary architecture and under matched computational budgets would help clarify the source of the additional improvement over the motivating study in Section 2.2.
(4) Underlying mechanism. Beyond attention visualizations, could the authors provide more direct evidence explaining how improved dynamic-region prediction affects action generation? For example, analyzing the relationship between dynamic-region prediction errors and task success rates, or evaluating action sensitivity under controlled perturbations of task-relevant regions, could strengthen the mechanistic interpretation.
(5) Generalization and experimental reliability. What motivates assigning all wrist-camera tokens to the high-noise group? Could the authors provide an ablation of this design choice and investigate whether motion-guided region selection remains effective under substantial camera motion? Additionally, reporting results across multiple independent training seeds would help establish the consistency of the performance gains.
4 comments
[W4, Q4] Further explanation and analysis of the mechanism.
MoGAP introduces a more challenging prediction task for dynamic regions while preserving static regions as context. This encourages the video backbone to better model manipulation dynamics, which in turn benefits action prediction through WAM’s joint video-action learning. To further investigate this mechanism, we measure region-wise visual prediction errors over four episodes for each of the 24 GR-1 Tabletop tasks.
We observe that dynamic regions exhibit higher prediction errors than static regions, with particularly large errors in regions containing the robot. MoGAP reduces prediction errors in dynamic and robot regions while leaving static-region errors nearly unchanged. This observation is consistent with our hypothesis that asymmetric prediction encourages the video backbone to better model manipulation dynamics, which may contribute to improved action prediction. This suggests that MoGAP selectively improves the video modeling of challenging regions, which is consistent with improved task success through joint video-action learning.
\begin{array}{lcccc} \hline \text{Method} & \text{Static region error} & \text{Dynamic region error} & \text{Robot-region error} & \text{Success Rate} \newline \hline \text{Baseline} & 0.031 & 0.170 & 0.241 & 52.3\% \newline \textbf{MoGAP} & 0.031 & \mathbf{0.165} & \mathbf{0.233} & \mathbf{62.3\%} \newline \hline \end{array}
We additionally conducted perturbation experiments at task-relevant regions, especially at the robot’s arm, as the reviewer suggested to better understand the scope and reliability of MoGAP. For noise and lighting perturbation, we add a localized gaussian noise in the region around the robot arm, and darken the same region, respectively in GR-1 Tabletop. For camera perturbation, we include moving camera experiments in our response to [W5, Q5]. We find that MoGAP consistently outperforms the baseline model under such perturbations.
\begin{array}{lccc} \hline \text{Method} & \text{Noise perturbation} & \text{Lighting perturbation} & \text{No perturbation} \newline \hline \text{Baseline} & 43.7\% & 45.5\% & 52.3\% \newline \textbf{MoGAP} & \mathbf{58.2\%} & \mathbf{54.0\%} & \mathbf{62.3\%} \newline \hline \end{array}
[W5, Q5] Generality under moving camera conditions.
We thank the reviewer for raising this concern. In RoboCasa-Kitchen, strong wrist-camera motion can dominate the latent motion scores, making it difficult to identify robot movements in the other exterior camera views. We therefore assign all wrist-camera tokens to the high-noise group, preserving motion-guided region selection in the external views while retaining visual prediction supervision for the wrist view. This provides a practical solution for particularly strong camera motion, but does not imply that MoGAP requires special treatment under all moving-camera conditions.
To further investigate its applicability during training, we separately trained the baseline and MoGAP in two additional settings. (1) egocentric camera movement on our OpenArm humanoid, where the robot continuously rotates its neck through a total range of 30° during manipulation, and (2) a mobile manipulation scenario on the SteamInMicrowave task from RoboCasa365 [3], where a mobile manipulator must place an object into a bowl, pick up the bowl, navigate to a microwave, place the bowl inside, and close its door.
\begin{array}{lcc} \hline \text{Method} & \text{OpenArm moving camera SR} & \text{Mobile manipulation SR} \newline \hline \text{Baseline} & 79.2\% & 58.0\% \newline \textbf{MoGAP} & \mathbf{87.5\%} & \mathbf{66.0\%} \newline \hline \end{array}
The results demonstrate that MoGAP remains effective under moving-camera and mobile manipulation settings. By deriving motion scores from consecutive latent slices in the spatially compressed VAE space, manipulation-related dynamics can remain distinguishable under camera movement, as also illustrated in Figure 8. We have also incorporated these results into the revised manuscript (Table 7 in Appendix E.1). Thank you again for the meaningful suggestion.
[3] Nasiriany et al., RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots, ICLR 2026
[W6] Repeated training runs on real-world generalization settings.
We conducted additional training runs with independent training seeds for our real-world experiments, and evaluated on the Object Relocation II task illustrated in Figure 7. We evaluated the two additional training seeds (0 and 7) over 12 trials per task, while retaining the original results for seed 42. We find that MoGAP consistently achieves meaningful performance gains across all three training seeds.
\begin{array}{lc} \hline \text{Method (Training seeds)} & \text{Object Relocation II Task success} \newline \hline \text{Baseline (0 / 7 / 42)} & 33.3\% / 41.7\% / 37.5\% \newline \textbf{MoGAP (0 / 7 / 42)} & \mathbf{75.0\% / 66.7\% / 79.2\%} \newline \hline \end{array}
We sincerely appreciate the time and effort you took to engage with our work. If there are any remaining concerns, we would be grateful if you could share them.
Gentle Reminder
Dear Reviewer 48hC,
Thank you again for your time and thoughtful efforts in reviewing our paper.
As the discussion period is coming to a close, we would like to kindly check whether you have any remaining questions or concerns. If so, please feel free to let us know.
Many thanks,
Authors
Dear reviewer 48hC,
We sincerely appreciate your efforts and insightful comments in reviewing our manuscript. Below, we address each comment in detail. In the revised manuscript, we mark our major revisions as “blue”.
[W1, Q1] Comparison to alternative noise allocation strategies.
MoGAP introduces token-wise noise assignment for the spatial asymmetry of dynamic and static regions, motivated by our finding that emphasizing dynamic regions in visual supervision improves WAM manipulation performance. Existing approaches such as Diffusion Forcing [1] and Self-Flow [2] employ frame-wise sequential noise allocation and randomized noise assignment, respectively. We conducted additional comparisons using these noise allocation strategies in the same training framework proposed at MoGAP. While alternative noise allocation strategies also improve performance over the baseline, motion-guided regional allocation achieves the highest success rate, suggesting its effectiveness for WAM manipulation. We have also incorporated these results into the revised manuscript (Table 9 in Appendix E.3). Thank you again for the meaningful suggestion.
\begin{array}{lc} \hline \text{Method} & \text{GR-1 Tabletop} \newline \hline \text{Baseline} & 52.3\% \newline \text{Self-Flow (random allocation)} & 53.8\% \newline \text{Diffusion Forcing (framewise allocation)} & 56.1\% \newline \textbf{MoGAP (regional allocation)} & \mathbf{62.3\%} \newline \hline \end{array}
[1] Chen et al., Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion, NeurIPS 2024
[2] Chefer et al., Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis, ICML 2026
[W2, Q2] Regarding the dynamic-region selection.
We thank the reviewer for raising this concern. As the reviewer noted, we provide quantitative and qualitative analysis about the motion scores at Table 4a and Figure 8, effectively distinguishing moving regions. We used a simple spatial mean threshold consistently across our experiments, which gave a reliable downstream control performance on diverse simulation and real robot benchmarks. To further investigate this sensitivity, we conducted additional comparisons with alternative threshold strategies under the same training setup.
\begin{array}{lc} \hline \text{Method} & \text{GR-1 Tabletop} \newline \hline \text{Baseline} & 52.3\% \newline \text{Random mask} & 53.8\% \newline \text{Continuous soft weighting} & 54.8\% \newline \text{Threshold = 0} & 55.3\% \newline \text{Threshold = mean - }\sigma & 55.9\% \newline \textbf{Mean thresholding} & \mathbf{62.3\%} \newline \hline \end{array}
We further investigate the reliability to local visual noise, camera perturbations and movements at our response to [W4, Q4] and [W5, Q5].
[W3, Q3] Disentanglement of asymmetric noise allocation and dynamic-region loss reweighting.
To disentangle the contributions of asymmetric noise allocation and dynamic-region loss reweighting, we conducted an additional controlled experiment using the same auxiliary prediction architecture under matched computational budgets. Adding the auxiliary prediction path to loss reweighting alone does not improve performance, whereas MoGAP's asymmetric prediction objective achieves a substantially higher success rate.
\begin{array}{lc} \hline \text{Method} & \text{GR-1 Tabletop} \newline \hline \text{Loss reweighting} & 57.7\% \newline \text{Loss reweighting + auxiliary prediction path} & 57.1\% \newline \textbf{MoGAP} & \mathbf{62.3\%} \newline \hline \end{array}