acceptodds
Withdrawn from ICLR 2027

Where Refinement Pays: Measuring and Allocating Sampling Compute in Flow-Matching Manipulation Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow-matching manipulation policies typically use the same number of function evaluations (NFE) at every call, assigning the same sampling budget to free-space motion and precision insertion. Does each decision benefit from the same amount of computation? Whole-episode comparisons cannot isolate the value of one budget choice because changing the budget also changes the states visited afterward. We introduce counterfactual state branching, a testbed that restores exact simulator states and policy histories, matches sampler noise, and varies the budget at selected calls. All branches then use the same continuation budget. We compare terminal success and cumulative NFE from the intervention to termination. Across four flow-matching policies on nine VLA-Arena suites, budget responses vary strongly with the policy and decision state. Using more computation throughout an episode often fails to improve success and can reduce it. We use these paired success and cost outcomes to train FlowBudgetAllocator (FBA), which operates across suites while keeping the action policy frozen. In simulation, it achieves higher observed success with fewer NFE than the default high budget. On basic configurations, it improves task success over the best fixed budget with only a modest increase in NFE. FBA exhibits distinct budget-selection patterns across tasks and execution phases. On the real robot, it uses substantially lower sampling budgets than the default high-budget policy without reducing observed task success. By measuring the consequences of individual budget choices, we turn sampling computation into a resource to allocate during control.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Manipulation, Embodied Intelligence, Flow-Matching Policy, Closed-Loop Control, Counterfactual Evaluation.

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 3 reviews, above 27% of the others.
Reviewer_trbYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies sampling-budget allocation for flow-matching robot manipulation policies. Existing methods usually use the same number of function evaluations (NFE) at every policy call, but free-space motion, approach, contact, and precision manipulation may need different amounts of sampling computation. The authors introduce counterfactual state branching, which restores the same simulator state, policy history, and sampler noise, changes the NFE only at selected calls, and compares success and computation under a common continuation budget. They use these paired outcomes to train FlowBudgetAllocator (FBA), which selects the sampling budget from the current history while keeping the action policy frozen. Experiments show that budget responses vary strongly across policies and decision states, and that uniformly increasing NFE throughout an episode does not always improve closed-loop success and can even reduce it. FBA achieves good success and latency with lower sampling budgets in simulation and on a real robot, but strict counterfactual state branching is implemented only in simulation, while the real-robot results depend on scene-specific offline adaptation data.

Critical strengths
  1. The problem is meaningful. Keeping the action policy at a fixed NFE throughout an episode is not always optimal, so this problem is worth studying.
  2. The design of FlowBudgetAllocator (FBA) for adaptive sampling-budget allocation is generally reasonable, and the experiments show some effect.
Critical weaknesses

W1. A testbed that saves and restores the robot or simulator state at a given time is already supported by the underlying systems of many simulators. The technical contribution here is therefore mainly to package an existing capability into an experimental framework for budget interventions. Besides, counterfactual state branching seems difficult to deploy on the real robot. To address this limitation, the authors first train FBA on simulation tasks and then offline fine-tune it with real-robot data, but this raises concerns about the generalization ability and ease of use of the method.

W2. The authors find that “uniformly higher budgets often fail to improve closed-loop success or degrade it” and give some explanations, but the explanation is not fully convincing and the theory is not sufficiently rigorous. For example, “Expert-state probes characterize changes in mean-action agreement and sampling variance alongside numerical refinement” seems to explain this phenomenon. The authors roughly argue that there are two types of error: the model's action error relative to the expert action, and the error caused by using a small number of integration steps. If the latter happens to cancel the former, the total error is reduced. This seems accidental, because the two errors could also add together and make the result worse.

Questions for authors
  1. For weakness 1: When counterfactual states cannot be restored precisely on the real robot, how can the authors show that FBA trained in simulation and offline fine-tuned with real-robot data generalizes to real scenes with reasonable data and deployment costs?

  2. For weakness 2: Can the authors provide a more rigorous theory or controlled experiments showing that the relationship between finite-step integration error, model action error relative to the expert, action-mean bias, and sampling variance is sufficient to explain why a higher NFE can sometimes reduce closed-loop success?

Reviewer_odssOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how the number of inference steps (NFE) for a flow-matching policy affects the downstream performance. The author shows that increasing NFE does not necessarily lead to higher success rates. Then they proposed FlowBudgetAllocator (FBA) to predict the best NFE at each state. Experiments in VLA-Arena and Libero show that FBA improves performance over using a fixed-value NFE.

Critical strengths

The finding that increasing sampling steps can actually hurt task success is somewhat counterintuitive and provides interesting insights. The extensive analyses across four different flow-matching policies are also valuable. The proposed FBA is a reasonable application of these findings, and the simulation results demonstrate that adaptive sampling can somewhat improve task success.

Critical weaknesses

W1. While the paper provides analyses of sampling budgets across four different VLAs, the proposed FBA is only evaluated on π0.5. Given that the relationship between NFE and task success varies significantly across policies, it is unclear whether the proposed allocation strategy generalizes to other flow-matching architectures. Would it be possible to evaluate FBA on other VLAs as well?

W2. The proposed matched counterfactual branching approach relies heavily on simulation and cannot be easily applied in the real world. Instead, the real-world experiments require human-in-the-loop interventions for data collection. This somewhat limits the practical applicability of the proposed framework.

W3. The performance improvement of FBA over the fixed-budget baseline (NFE 1) seems to be modest both in simulation and in the real world.

Questions for authors

Q1. The fact that NFE 1 achieves the best performance in Figure 1 is still counterintuitive to me. Could the authors elaborate on why increasing NFE hurts performance? Also, would it be possible to conduct NFE ablations on other benchmarks, such as RoboLab and RoboDojo, to see whether this finding holds across different benchmarks?

Reviewer_9RgBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper seeks to investigate at what states additional denoising sampling for actions would have a marked effect on policy performance. By dynamically adjusting the sampling cost under different situations, the overall policy can improve on latency and solve tasks faster. The authors seek to analyze the effect of using lower sampling budgets using repeatable/state-restorable simulators, and use the resulting datasets to train a dynamic compute allocator.

Critical strengths

The paper tackles an interesting question, which has practical implications. It would be nice to understand what kind of situations require more careful action generation and under what conditions fast rough action sampling is sufficient.

Critical weaknesses

One weakness is that the findings do not seem to be consistent across different models (Figure 3), not just between model types (such as WAM vs VLA) but even across VLAs (π0.5,π0\pi_{0.5}, \pi_0). This means that each model has to have its own study on how to best allocate its compute.

It also seems a bit contrived in Figure 1(b) that timeout occurs after call 29 and the successful run terminates after call 23 - the only difference is saving 9 NFEs, so it appears that the authors specifically chose an arbitrary threshold for "timeout". In practice, such as the real world, the additional 9 NFEs would not generally add significant time - especially as VLAs by themselves can offer real-time reactive control by default.

It is not clear how the real-world FBA is trained, especially as we no longer have exact replayable states for counterfactual interventions to collect data. The performance listed in Table 2 also does not provide strong takeaways for me - and even has some strange results. For example, why is it that NFE 10 has such low success rate performance? As there is no timeout, should it not be the most expressive in terms of generating actions? Why is it that NFE 10 and NFE 4 seem to consistently underperform NFE 1? This seems rather strange and should be explained.

The limitation of requiring a simulator that can perfectly replay from states (e.g. "event-anchor sampling") is indeed a major weakness of this work. Such simulators are needed to perform per-model studies on compute allocation, as the authors have shown that different models have different compute-conditioned behaviors. But then there is no mechanism with which to study this on real-world deployments, where we are most curious about practical VLA performance, as we don't have access to such an intervention scheme in the real world. Thus, practical takeaways are hard to extract from this proposed method.

Questions for authors

Why does the graph in Figure 6(a) seem incomplete? Some curves (such as Dynamic Distractors) seem to just not continue to the full range of 30 policy calls.

Do the relationships between stages and transitions in Section 2.3 that are found in simulation for a particular model translate to the same stages and transition relationships in the real world? Whereas they may not be consistent across models, for a given model are the relationships consistent across simulation and real? Could this be studied or shown through a sim-to-real experiment? For example, comparing the FBA strategy found for a specific model in simulation, then deploying it in the real world for a corresponding task using the same FBA strategy, and showing it matches the improvement trends discovered in simulation.

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

Related papers

Discussion (0)

Sign in to comment.