- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Disagreement-Guided Replanning (DGR), an online adaptive replanning strategy for VLA policies. It addresses the trade-off between frequent replanning, which can cause inconsistency between consecutive action plans, and infrequent replanning, which leads to accumulated execution errors. DGR uses the disagreement between overlapping segments of consecutive action chunks to dynamically adjust the execution horizon, without additional training or policy evaluations. Experiments across four VLA architectures and multiple benchmarks show improved long-horizon task performance over existing baselines, with additional validation on RoboCasa and a real-world Push-T task.
Critical strengths
- DGR utilizes the differences in overlapping regions between consecutive action predictions to dynamically adjust the step size of action execution. Without the need for additional training or increasing the number of model inferences, it can achieve a balance between execution efficiency and re-planning frequency.
- The experimental paper has verified this method on 4 VLA architectures, multiple simulation benchmarks, visual perturbation scenarios, and the real robot Push-T task. The results show that this approach is particularly effective in long-term tasks.
Critical weaknesses
- DGR uses the differences between overlapping action segments as the basis for adjusting the step size of the re-planning process. However, the differences in actions do not necessarily indicate execution errors or uncertainties. On the contrary, if the model consistently predicts similar but incorrect actions, this method may not be able to detect the problem in a timely manner.
- Prior work such as AAC (https://arxiv.org/pdf/2604.04161) adapts action execution horizons based on action entropy, while MoH (https://arxiv.org/pdf/2511.19433) uses cross-horizon consensus to dynamically select executable action prefixes. Although DGR employs overlap disagreement and an EMA-based adjustment rule, the paper could more clearly establish its unique advantages over these existing approaches and demonstrate that its design offers distinct benefits beyond alternative signals for adaptive execution.
Questions for authors
- The paper assumes that the degree of action inconsistency brought about by re-planning is fixed and does not change with each execution of a few steps. Moreover, the longer the consecutive execution lasts, the greater the cumulative execution error will be. However, this may not be the case in reality. For instance, when a robot grabs an object, errors may be particularly prone to occur at the moment of initial contact with the object, and the error does not necessarily increase uniformly with the number of execution steps; re-planning may also cause different degrees of action changes due to different task stages. Are these two assumptions still valid in different task stages, especially in operations involving grasping, contact, and object movement?
- Sometimes, the difference in predictions suddenly becomes larger. This might just be a random fluctuation of the model and does not necessarily mean that more frequent re-planning is truly necessary. If DGR subsequently shortens the execution step size, it might cause unnecessary re-planning. How can this be avoided?
3 comments
Thank you for the detailed clarifications and additional experiments. The comparisons with AAC and MoH help better establish the methodological distinctions and computational advantages of DGR. The adversarial perturbation results also provide useful evidence of its robustness to occasional policy errors. However, some concerns remain regarding the validity of the fixed inconsistency assumption across different task stages and the sensitivity of horizon adaptation to prediction fluctuations. The proposed dead zone and bounded updates are reasonable design choices, but their effectiveness under varying execution conditions is not yet fully established. Overall, I appreciate the authors' efforts, but I maintain my original rating.
Thank you for following up on our response. Your comments helped us identify where the aggregate modeling assumptions needed clearer explanation and motivated the additional robustness experiment below.
Q2. The sensitivity of horizon adaptation to prediction fluctuations. To evaluate the sensitivity of horizon adaptation to prediction fluctuations, we performed an evaluation where we perturbed the predicted action chunks with zero-mean Gaussian noise. The noise had standard deviation along action coordinate , where is the empirical standard deviation along that coordinate measured from clean policy outputs. We can thus increase the variance of the action chunks by setting . As the included table shows, DGR as proposed has higher task success rate than its variants without the deadzone and with a high horizon update step, at all three noise levels. DGR is also competitive with the Static Optimal (). At noise level all methods perform poorly due to the high noise. These results support the effectiveness of the safeguards under the tested fluctuations.
| Method | |||
|---|---|---|---|
| DGR | |||
| No dead zone () | |||
| Larger updates () | |||
| Fixed |
Q1. Fixed inconsistency assumption across different task stages. We do not assume fixed inconsistency across different task stages. Assumption 2 concerns the expected value over the distribution of states in a task. Different task stages (e.g., grasping, contact, or transport) correspond to different subsets of the state space , but our cost model does not assume independence of the costs from the state.
To make it explicit what we average over, let denote the distribution of states encountered at a uniformly sampled replanning event under fixed horizon . Let and be the replanning inconsistency and the instantaneous drift cost incurred after steps if the system is in state upon replanning, respectively. A formulation, where state-dependence of the cost surrogate is explicit would be
\left[ \sigma(X)+\sum_{j=0}^{r-1}\phi(X,j) \right]dX.$$ We can extract the terms corresponding to the replanning inconsistency and the instantaneous drift from the above expression, showing their dependence on $r$, leading to $$\sigma_r=\int p_{\text{task},r}(X)\sigma(X)\,dX, \qquad \phi_r(j)=\int p_{\text{task},r}(X)\phi(X,j)\,dX.$$ Observe that Assumption 2 assumes independence of the **expected replanning inconsistency cost** ($\sigma_r$ above) from **the replanning horizon $r$**, rather than independence of $\sigma(X)$ from the state $X$. The motivation behind this assumption is given by the sources of inconsistency described on lines 189-194. The task and the choice of $r$ do influence $\sigma_r$ through $p_{\text{task},r}(X)$. For different tasks we find similar values of $\sigma_r$ for $r=1$ (lines 369-371), and considering that $p_{\text{task},r}(X)$ is strongly dependent on the task, but only weakly on $r$ (the overall task progression remains similar regardless of $r$), it is reasonable to consider this **average** to be an intrinsic property of the policy, justifying the modeling Assumption 2. Importantly, state (and thus stage)-dependent inconsistency costs are compatible with this aggregate model. We would like to emphasize that the cost model based on the two components is used in our paper for motivating horizon adaptation, as written on lines 112-113 and 162-164. The proposed controller DGR does itself not estimate $\sigma$ or $\phi$. Its effectiveness is assessed empirically across the reported policies and benchmarks. We have also extended the evaluation with controls matched to DGR’s replanning horizon distribution and its mean horizon, providing further evidence for the benefit of online adaptation. We hope these additions, together with the clarifications and robustness experiments above, address your remaining concerns underlying the rating.
Thank you for the careful review.
W1. Consecutive predictions can agree while both are wrong. We agree and we acknowledged this limitation in the manuscript (see Limitations). DGR aims to improve the execution of an already capable policy through better replanning. If the underlying policy consistently predicts incorrect actions, the policy itself needs improvement, for example through retraining; horizon adaptation alone cannot generally correct a consistently incorrect policy.
Nevertheless, our results in Appendix F show that when the input (and hence the policy output) is perturbed using an adversarial patch, DGR remains competitive and provides substantial gains on LIBERO-Spatial: 74% success, compared with 54% for the clean static optimal horizon and 69% for the attack-optimized static optimal horizon. This shows that DGR can compensate for occasional policy errors.
W2. Novelty. Both AAC and MoH involve generating multiple action chunks upon inference time. In addition, MoH involves retraining and architectural changes of the VLA policy prior to inference time. AAC estimates entropy from multiple candidate chunks generated based on the same observation. It would thus not work if chunks are generated using a deterministic regression policy that produces identical chunks under repeated queries. On the contrary, DGR compares chunks generated at subsequent observations, which can disagree even when action chunks are generated deterministically. Our OpenVLA-OFT results demonstrate applicability to this setting. Table 3 additionally shows approximately two to three times fewer total FLOPs than cached AAC on LIBERO-Long. This is particularly relevant for robotics, since onboard compute capacity can be limited.
MoH involves retraining a policy with multiple prediction horizons as well as weights to be used for fusing multiple action chunks. Upon inference time, it then selects an executable prefix, using consensus among horizon-specific predictions generated using the same observation, similar in spirit to AAC. While batching can reduce latency, actual inference time remains heavily dependent on the available hardware, and the overall compute, memory, and hardware requirements are higher than for DGR.
DGR is fundamentally different from these approaches in two ways. First, it relies on the comparison of two consecutively generated chunks. Thus, it can be used with any existing action-chunking policy without additional training, architectural modifications, or multiple evaluations. Thus, DGR preserves the base policy’s per-call inference cost, adding only negligible computation and memory overhead for the controller. Second, DGR's design is motivated by a surrogate analysis and the fixed-horizon problem studied offline. We empirically investigate weighted overlap disagreement as a bridge to task performance, then extend the same objective to variable horizons to derive the greedy online controller. This connection strengthens the justification for DGR’s controller: its online update is derived from an explicit objective, connected to the fixed-horizon analysis, and supported by empirical evidence. To our knowledge, this connection is also novel.
Q1. The static surrogate uses average costs across rollout states. We agree that drift and replanning inconsistency can vary across grasping, contact, and object movement. Our static surrogate uses state-averaged costs. Here, j denotes the number of steps since replanning, rather than absolute task time t. Assumption 1 concerns the expected drift profile ϕ(j): it permits nonlinear growth and sharp increases, and does not require error to increase uniformly along every trajectory. Similarly, σ represents an expected per-replan cost; Assumption 2 takes this average as independent of r, rather than requiring every replanning event to introduce identical inconsistency.
A stage-dependent interpretation is possible under a mean-field assumption that replanning encounters the same expected mixture of task stages across fixed horizons. Stage-specific costs then aggregate into the same surrogate form, and the proof applies provided the averaged profile satisfies the stated assumptions. We will clarify this interpretation and scope.
It is important to note that DGR does not estimate σ or ϕ at inference time. It uses the current disagreement and its smoothed average, allowing the execution horizon to change across stages. For example, the within-episode Push-T trace in Appendix G illustrates this behavior.
Q2. The dead zone and bounded updates limit responses to fluctuations. Directional calibration is conditioned on the score leaving the dead zone, whose width is controlled by α. Additionally, each horizon change is bounded by Δ. Therefore, α controls the controller sensitivity, while Δ controls the size of each horizon change. These two parameters can be jointly tuned to address this issue.