- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies why world–action models (WAMs) trained on static demonstrations struggle to manipulate moving targets. The authors identify a failure mode called target-response collapse, where the policy gradually loses its ability to respond to target relocation as execution proceeds. To address this, they propose Dynamic Predictive Planning (DPP), which estimates the future interaction time and target position, constructs counterfactual observations using canonical robot configurations, and aligns the resulting plans with ongoing execution. The framework aims to reuse pretrained manipulation skills without additional training on dynamic tasks. Experiments on RoboTwin2.0-Dyn, more complex motion scenarios, and real robots show substantial improvements over several reactive, future-conditioned, and dynamically trained baselines.
Critical strengths
-
The paper identifies and carefully analyzes an interesting failure mode of existing WAMs. I found the matched-prefix experiments in Section 3 and Appendix B particularly informative. For FastWAM, the mean directional response gain drops from 0.897 to 0.078 after 48 executed actions, while the response-hull area decreases to only 1.47% of its initial value (Table 6). A similar trend is observed for ImageWAM in Table 7. These controlled experiments provide fairly convincing evidence of target-response collapse. More importantly, they offer a useful perspective on why policies trained on static demonstrations may fail in dynamic environments, rather than simply reporting low success rates on moving-target tasks.
-
The proposed framework provides an interesting way to reuse pretrained manipulation skills without additional training. DPP predicts where the target will be at the expected interaction time and combines this prediction with a familiar robot configuration to generate a new plan. The plan is then aligned with the robot's current execution state. Compared with SIDO (Shin et al., 2026), which relies on counterfactual action augmentation during training, and ReOI (Chen et al., 2026b), which modifies visual distractors at inference time, DPP uses canonical observation conditioning to reuse the existing WAM policy. The ablation in Table 2 also supports this design choice: removing canonical-state conditioning reduces Task 1 success from 83.33% to 51.67%. I think this training-free formulation is one of the more appealing aspects of the work.
-
The experimental results are strong across different tasks and model backbones. In Table 1, FastWAM-DPP and ImageWAM-DPP achieve 83.33% and 78.67% success on bimanual grasping, respectively, compared with 52.00% for the dynamically trained DynamicWAM baseline. Improvements are also reported for pick-and-place and additional task types in Appendix A. The real-robot results are encouraging as well, with success rates improving from 0% to 76.7% on Task A and from 40.0% to 86.7% on Task B (Figure 4). The relatively consistent gains across these settings suggest that the proposed approach has practical value beyond a single task or backbone.
-
The paper pays attention to practical execution issues, which is important for this type of planning framework. In particular, the authors address the mismatch between a trajectory planned from a canonical state and the robot's actual state through trajectory bridging and asynchronous plan activation. Table 11 reports that generating a 32-action chunk takes approximately 0.495 seconds, while executing it takes 3.2 seconds at 10 Hz. This suggests that planning can be performed asynchronously without necessarily interrupting execution. The real-robot deployment using an RTX 4090, together with the controller details in Appendix C, also makes the implementation more convincing and easier to reproduce.
Critical weaknesses
-
The benchmark setup raises some concerns about the fairness of comparisons with dynamically trained baselines. According to Appendix E.2, the evaluation regions are selected based on FastWAM's performance in static validation. This may favor configurations that are already well supported by FastWAM, while other independently trained policies may have different regions of competence. Another concern is the change in the motion-to-physics handoff protocol between RoboTwin2.0-Dyn and DOMINO. The difference appears to matter: PUMA achieves 86.67% success with DOMINO-style contact stopping in Table 10, compared with 50.33% in Table 1. DPP also uses a timing adjustment specific to this protocol. Although DPP still performs better in the reported experiments, I think the comparison would be more convincing with static-support controls for each baseline and evaluations under a more consistent handoff protocol.
-
It is unclear how much of the improvement comes specifically from WAM-derived interaction-time prediction. A key part of DPP is using the interaction event predicted by the initial WAM rollout to estimate the future target position (Eq. 8). However, this component is not directly isolated in the experiments. The fixed-horizon Future-Conditioned baselines use different controllers, and the fixed-time VLA-style alternative in Appendix F.7 also changes the trajectory generation and execution procedure. As a result, these comparisons do not really show whether the WAM-predicted interaction time is better than a fixed or calibrated time estimate when the rest of DPP remains unchanged. Since interaction timing is central to the method, I would like to see a more controlled ablation of this component.
-
The explanation of target-response collapse is plausible, but the underlying cause is not fully established. Section 3 suggests that the reduced responsiveness comes from robot–target configurations that are poorly covered by static training demonstrations. The matched-prefix experiments clearly demonstrate the behavioral effect, but they do not directly measure whether these configurations are outside the training distribution. There may also be other explanations. For example, the robot could become less able to recover from target changes due to kinematic constraints or execution progress, or the model's predictions may become less reliable at later states. Appendix H.7 acknowledges this limitation, which is good. Still, I think the paper should distinguish more carefully between establishing target-response collapse as an observed phenomenon and explaining its primary cause.
-
The generalization results for more complex target motion rely on assumptions that may not hold in less structured environments. The base DPP controller uses a constant-velocity estimate that is frozen after bootstrap (Appendix F.3). For the complex-motion experiments in Figure 5, the method uses predefined motion-family information, late canonical planning, and conditional spatial warping. Several motion checks in the base controller are also relaxed for this evaluation. These adaptations are reasonable, and the results are useful, but they make it difficult to assess how robust DPP would be when the target motion is not known in advance. The real-robot experiments also use controlled motor-driven object trajectories. I would therefore be cautious about generalizing the conclusions to unpredictable motion, noisy trajectory changes, or more unconstrained interactions.
Questions for authors
-
Could the authors report the static-task success rates of each backbone and dynamically trained baseline in the selected evaluation regions? I would also be interested in whether the relative rankings change when using regions not selected based on FastWAM, or when applying the same motion-to-physics handoff protocol to all methods.
-
How does WAM-derived interaction timing compare with fixed, empirically calibrated, or oracle interaction times when all other DPP components are kept unchanged? It would also be helpful to report the timing prediction error and how sensitive task success is to errors in the estimated interaction time.
-
Could the authors provide further evidence that target-response collapse is mainly caused by insufficient coverage of joint robot–target states in the training data? How can this be distinguished from reduced recoverability or increasing prediction errors at later execution stages?
-
For the complex-motion experiments in Figure 5, how much do the improvements depend on knowing the motion family in advance and relaxing the controller's motion checks? Would similar results be expected with automatic motion-model selection and the same controller constraints as the base setting?
2 comments
We sincerely thank the reviewer for their thoughtful feedback and for taking the time to carefully consider our responses. We are pleased that our clarifications have fully addressed your concerns, and we greatly appreciate your continued positive assessment of our work.
Best regards,
The Authors
Official Comment to Authors
Thanks for the responses, which fully address my concerns. Thus, I retain my positive rating of 3.