acceptodds
Under review as a conference paper at ICLR 2027

Ground Your Steps: Optimizing Sampling Schedules in Flow Matching Robot Policies

Abstract

Flow matching robot policies generate action chunks by integrating a learned velocity field at inference. However, under a fixed budget of neural function evaluations (NFEs), the integration result depends on the sampling schedule. This motivates optimizing the schedule to bring the generated action distribution closer to the demonstrations and improve task success. We propose Ground Your Steps (GYS), a training-free method that optimizes the sampling schedule by matching the generated and demonstrated chunk distributions at each observation. GYS applies truncated fractional differencing along each arm joint's action sequence and compares two types of pairwise differences: generated–demonstrated and generated–generated, using independent policy samples at each observation. After fractional differencing, the distributions of generated–demonstrated and generated–generated pairwise differences coincide exactly when the distributions of generated and demonstrated action chunks match. GYS minimizes the energy distance between these distributions to obtain an improved schedule for each task and NFE budget. Compared with the uniform grid, GYS brings the generated chunk distribution closer to the demonstrated distribution and raises the success rate through the sampling schedule alone. On RoboTwin 2.0 benchmarks, GYS improves success rates at nearly every budget, with gains of percentage points for and percentage points for GR00T N1.7. The gains also transfer to three real-world manipulation tasks, reaching up to percentage points.

Primary area:
generative models
Keywords:
Flow Matching, Efficient Sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_pHTKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free sampling schedule adaptation method using truncated fractional differencing, which is applied to each arm joint's action sequence. The method aims to minimize the energy distance between the generated action distributions and the demonstration distributions to obtain an improved sampling schedule for each task and NFE budget. The goal is to optimize the sampling schedule to better align the generated action distribution with the demonstrations and improve task success under each fixed NFE budget. The authors validate the proposed method in both simulation and real-world robot experiments.

Critical strengths

The paper provides a comprehensive introduction and related work section. The writing is clear and easy to understand and follow.

The paper provides detailed theoretical support for the proposed method.

The proposed method is interesting. Instead of using gradient descent, it searches for an improved sampling schedule by proposing several candidate schedules and calculating their energy-distance losses to find the one that best aligns the generated actions with the demonstration data.

The method is validated on three real-world bimanual manipulation tasks, as well as simulation benchmarks.

Critical weaknesses

Limited success rate improvement: Compared with the baseline methods, the proposed method achieves relatively small success-rate improvements, generally around 2–5 percentage points. The results in Tables 1 and 2 do not demonstrate an obvious advantage of the proposed method over the baselines in either simulation or real-world experiments.

Ablation study: According to Table 3, removing observation conditioning by setting λ = 0 only slightly affects performance. In addition, the overall improvements of the proposed method over the baseline are relatively small. I am concerned about whether these minor improvements reflect a meaningful contribution from the observation-conditioning component or are simply due to random variation. Additional statistical analysis or ablation experiments would help clarify this.

Questions for authors

Clarity of the abstract: Could the authors provide a more explicit explanation of the term "generated–generated" in the abstract? This is the first time the term appears in the paper, and its meaning may not be immediately clear to readers.

Relationship between energy distance and task success: The experimental results do not provide sufficiently strong evidence of the proposed method's superiority over the baselines. Furthermore, the relationship between energy distance and task success is unclear. Does a lower energy distance consistently lead to a higher task success rate, including in out-of-distribution scenarios? An additional ablation study examining this relationship would be useful.

Overall, the proposed method is interesting and insightful. However, based on the experimental results, I do not see a clear or substantial advantage of the proposed method over the baseline methods. Additional experiments and analyses would help demonstrate the effectiveness and significance of the proposed approach.

1 comment
AuthorsOpenReview ↗

Response to Reviewer pHTK

We appreciate your valuable feedback and support of our work.

To W1:

We re-ran Uniform and GYS for three additional rounds under identical settings and report the mean ± std over four rounds in Table A. GYS consistently outperforms Uniform across all 12 settings with small standard deviations (≤ 0.7%), confirming that the gains are stable rather than random variation. It improves the average success rate by 2.82.8 percentage points on both π0.5\pi_{0.5} and GR00T.

Moreover, in RoboTwin 2.0 Clean2Random (Table C in To Q2), schedules optimized offline from limited clean demonstrations generalize across visual and spatial variations to consistently improve task success rates over Uniform in randomized scenes. As a training-free method with zero inference overhead, this consistent improvement provides clear practical value.

Table A. Success rate (%), mean ± std over four rounds. Avg. is the mean ± std across the six budgets (from Table 1 of the revised PDF).

Policy / ScheduleB=3B=3B=4B=4B=5B=5B=6B=6B=8B=8B=10B=10Avg.
π0.5\pi_{0.5} Uniform67.7 ± 0.668.2 ± 0.468.0 ± 0.567.7 ± 0.466.9 ± 0.466.6 ± 0.267.5 ± 0.6
π0.5\pi_{0.5} GYS70.9 ± 0.371.7 ± 0.370.8 ± 0.269.6 ± 0.469.3 ± 0.269.3 ± 0.370.3 ± 1.0
GR00T Uniform47.8 ± 0.748.6 ± 0.650.7 ± 0.752.1 ± 0.548.0 ± 0.647.7 ± 0.549.2 ± 1.8
GR00T GYS51.0 ± 0.551.9 ± 0.451.7 ± 0.452.5 ± 0.653.0 ± 0.551.8 ± 0.552.0 ± 0.7

To W2:

Retaining λ>0\lambda > 0 provides theoretical grounding for conditional distribution matching. Setting λ=0\lambda = 0 compares only pooled marginal distributions across observations, which theoretically allows a schedule to score well despite per-observation mismatches. Incorporating the observation summary (λ>0\lambda > 0) ensures a zero objective implies conditional action distribution agreement (Proposition 1).

In practice, because GYS searches for a single task-level schedule, the task-pooled signal already gives strong guidance, yielding similar schedules. Since λ\lambda is set automatically via the median heuristic with negligible overhead, we retain λ>0\lambda > 0 as a principled safeguard.

To Q1:

A "generated–generated" difference is the difference between two action chunks independently generated at the same observation under different noise draws. We updated the abstract in the revised PDF to clarify both pairwise differences:

"...compares two types of pairwise differences between action chunks at each observation: differences of a generated chunk against a demonstration, and differences of two independent generated chunks against each other."

To Q2:

  1. Relationship between energy distance and task success. GYS minimizes the energy score to bring generated action distributions into agreement with demonstrations. On blocks ranking rgb with GR00T N1.7 at B=4B=4, we evaluated GYS alongside Uniform and three representative manual schedules ranging from step allocations concentrated near the noise endpoint (τ=0\tau=0) to those near the action endpoint (τ=1\tau=1), all scored under the same calibrated λ\lambda and ν\nu. As shown in Table B, GYS achieves the lowest energy score (−64.20%), and lower energy scores consistently correspond to higher success rates (rising from 29% to 55%), confirming that closing the distribution gap directly improves closed-loop performance.

Table B. Energy score and closed-loop success rate across schedules (from Figure 4 of the revised PDF).

ScheduleEnergy score change relative to uniformSuccess rate
[0,0.583,0.595,0.744,1][0, 0.583, 0.595, 0.744, 1] (GYS)−64.20%55%
[0,0.200,0.450,0.680,1][0, 0.200, 0.450, 0.680, 1]−51.70%48%
[0,0.150,0.300,0.500,1][0, 0.150, 0.300, 0.500, 1]−40.91%39%
[0,0.250,0.500,0.750,1][0, 0.250, 0.500, 0.750, 1] (uniform)0.00%36%
[0,0.550,0.800,0.930,1][0, 0.550, 0.800, 0.930, 1]+49.61%29%
  1. Generalization to OOD scenarios. Using VLAs fine-tuned on clean demonstrations and clean-optimized per-task schedules without re-optimization, we evaluated all 50 RoboTwin 2.0 tasks in unseen domain-randomized scenes (Clean2Random). In Table C, across B=3B=3 and B=4B=4, GYS improves mean success over Uniform by 3.33.3 pp on π0.5\pi_{0.5} and 4.84.8 pp on GR00T, confirming that lowering energy distance on limited clean demonstrations generalizes across visual and spatial variations.

Table C. Generalization success rate (%) averaged across 50 RoboTwin 2.0 tasks in unseen domain-randomized scenes (from Table 3 of the revised PDF).

PolicyNFEsUniformGYSGain
π0.5\pi_{0.5}B=3B=341.244.5+3.3
π0.5\pi_{0.5}B=4B=442.745.9+3.2
GR00TB=3B=315.019.6+4.6
GR00TB=4B=416.221.2+5.0

We welcome further discussion if any concerns remain.

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

This paper proposes Ground Your Steps (GYS), a training-free method for offline optimization of flow-time sampling schedules in flow-matching robot policies. GYS uses fractional differencing to better measure the distribution mismatch between policy-generated actions and expert demonstrations. Experiments show that GYS outperforms alternative sampling strategies under a fixed number of flow-matching inference steps.

Critical strengths

GYS is a training-free approach that improves policy performance under a limited flow-matching inference budget without introducing additional inference-time computation. The effectiveness of the proposed method is supported by extensive experiments in both simulation and real-world robotic manipulation tasks.

Critical weaknesses
  1. Limited generalization across tasks. GYS optimizes a separate sampling schedule for each task, requiring task-specific demonstrations. This raises concerns about its generalization to unseen tasks, especially in the context of general-purpose VLA policies.
  2. Insufficient ablation studies. The paper lacks several important ablation studies to justify the necessity of its proposed components. First, the method compares two distributions constructed from generated–demonstrated and generated–generated action differences. However, the paper does not investigate whether simpler objectives based on only one type of difference could achieve comparable performance. Such comparisons would help clarify the contribution of each component to the overall improvement. Second, fractional differencing appears to be an essential component, but the paper lacks a comprehensive analysis of different fractional orders. Third, the calibration dataset size could significantly affect the quality of the optimized schedule, but the paper does not investigate its sensitivity to the number of calibration samples.
  3. Lack of adaptation during task execution. The optimized schedule remains fixed throughout task execution and does not adapt to the current observation or execution state. This may limit its effectiveness in tasks involving diverse states or dynamically changing conditions.
  4. Limited understanding of the relationship between distribution matching and task success, as listed in the limitation section.
  5. In the introduction, the authors identify trajectory curvature and the Lipschitz window as two important factors affecting finite-step flow integration. However, the paper does not directly analyze how GYS mitigates these issues. It would strengthen the paper to provide related analysis.
Questions for authors

Please refer to the weaknesses listed above.

2 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer AZFa (To W1–W4)

We appreciate your valuable feedback and support of our work.

To W1:

We added a Generalization Study in Section 5 of the revised PDF. Using the VLAs fine-tuned on clean demonstrations and the per-task schedules optimized on the clean calibration set without re-optimization, we evaluated performance across the 50 RoboTwin 2.0 tasks in unseen domain-randomized scenes (Clean2Random). As shown in Table A, GYS improves mean success rate over Uniform by 3.33.3 percentage points on π0.5\pi_{0.5} and 4.84.8 percentage points on GR00T, showing that schedules optimized on clean demonstrations generalize to more randomized scenes.

Table A. Generalization success rate (%) averaged across 50 RoboTwin 2.0 tasks in unseen domain-randomized scenes (from Table 3 of the revised PDF).

PolicyNFEsUniformGYSGain
π0.5\pi_{0.5}B=3B=341.244.5+3.3
π0.5\pi_{0.5}B=4B=442.745.9+3.2
GR00TB=3B=315.019.6+4.6
GR00TB=4B=416.221.2+5.0

To W2:

We added all three ablations in Section 5 of the revised PDF. As shown in Table B:

  1. Single-difference objective (w/o self). With the same calibrated ν⋆\nu^\star, it minimizes only generated–demonstrated differences via Lw/o self(h)=1MN∑m=1M∑i=1N∥Φν⋆(Xm,i−ym)∥L_{\mathrm{w/o\ self}}(h) = \frac{1}{MN}\sum_{m=1}^{M}\sum_{i=1}^{N}\lVert\Phi_{\nu^\star}(X_{m,i}-y_m)\rVert. This objective pulls each generated chunk toward its paired demonstration, often crowding integration steps near the action endpoint. Such crowding can collapse a four-step schedule to nearly two effective steps while placing multiple steps where the Lipschitz window can arise. For instance, with π0.5\pi_{0.5} at B=4B=4, w/o self yields [0,0.107,0.964,0.999,1][0, 0.107, 0.964, 0.999, 1] on blocks ranking rgb and [0,0.395,0.987,0.999,1][0, 0.395, 0.987, 0.999, 1] on beat block hammer. Consequently, it drops success rate by 6.16.1–9.19.1 percentage points below Full and falls below Uniform, confirming the necessity of including generated–generated differences (X−X′X-X').
  2. Fractional orders (ν=0,1\nu=0, 1). Both raw positions (ν=0\nu=0) and fixed first-order differences (ν=1\nu=1) underperform the calibrated order ν⋆\nu^\star (Full) across policies and budgets, confirming the benefit of calibrating a fractional differencing order.
  3. Calibration size (M=400M=400). Doubling MM from 200200 to 400400 changes success rate by at most 0.40.4 percentage points, showing that M=200M=200 is sufficient.

Table B. Ablation study on RoboTwin 2.0 (%) at B=3B=3 and B=4B=4 (from Table 3 of the revised PDF).

PolicyNFEsUniformFullw/o selfν=0\nu=0ν=1\nu=1M=400M=400
π0.5\pi_{0.5}B=3B=367.770.963.068.168.070.8
π0.5\pi_{0.5}B=4B=468.271.762.666.167.971.5
GR00TB=3B=347.851.044.938.847.350.6
GR00TB=4B=448.651.944.246.246.352.0

To W3:

We agree that online adaptive scheduling during execution is a promising future direction. Nevertheless, our fixed task-level design provides strong practical value. Even without per-step adaptation, offline-optimized schedules generalize across visual and physical variations to unseen scenes (as shown in To W1), while avoiding dynamic adaptation overhead and keeping zero inference latency.

To W4:

GYS minimizes the energy score to bring generated action distributions into agreement with demonstrations. On blocks ranking rgb with GR00T N1.7 at B=4B=4, we evaluated GYS alongside Uniform and three representative manual schedules ranging from step allocations concentrated near the noise endpoint (τ=0\tau=0) to those near the action endpoint (τ=1\tau=1), all scored with the same calibrated λ\lambda and ν\nu from offline optimization. As shown in Table C, GYS achieves the lowest energy score (−64.20%), and lower energy scores consistently correspond to higher success rates (rising from 29% to 55%), confirming that closing the distribution gap directly improves closed-loop performance.

Table C. Energy score and closed-loop success rate across schedules (from Figure 4 of the revised PDF).

ScheduleEnergy score change relative to uniformSuccess rate
[0,0.583,0.595,0.744,1][0, 0.583, 0.595, 0.744, 1] (GYS)−64.20%55%
[0,0.200,0.450,0.680,1][0, 0.200, 0.450, 0.680, 1]−51.70%48%
[0,0.150,0.300,0.500,1][0, 0.150, 0.300, 0.500, 1]−40.91%39%
[0,0.250,0.500,0.750,1][0, 0.250, 0.500, 0.750, 1] (uniform)0.00%36%
[0,0.550,0.800,0.930,1][0, 0.550, 0.800, 0.930, 1]+49.61%29%
AuthorsOpenReview ↗

[2/2] Response to Reviewer AZFa (To W5)

To W5:

Directly quantifying trajectory curvature and the Lipschitz window is challenging because the ideal and continuous marginal flow is unavailable. Instead, we examine how both factors affect performance (SharpEuler and Dense-Jump are two baseline scheduling methods).

  1. Targeting curvature alone is insufficient. SharpEuler allocates finer steps where the learned flow bends fastest, targeting curvature only. It concentrates steps near the action endpoint (Table 8, Appendix C.3), indicating higher curvature there. However, it still falls 1.21.2 and 4.14.1 percentage points below Uniform on π0.5\pi_{0.5} and GR00T (Table 1, Section 5).
  2. Crowding the Lipschitz window hurts. Schedules concentrating steps near τ→1\tau \to 1 consistently degrade success rate. As shown in To W2, w/o self crowds multiple steps near τ→1\tau \to 1 (e.g., [0,0.107,0.964,0.999,1][0, 0.107, 0.964, 0.999, 1]) and falls 2.92.9–5.65.6 percentage points below Uniform, while the late-concentrated schedule in Table C falls 77 percentage points below Uniform.
  3. Skipping the late region also hurts. Dense-Jump crosses [0.5,1][0.5, 1] in one step, leaving late curvature unresolved, and falls 7.97.9 and 4.64.6 percentage points below Uniform on π0.5\pi_{0.5} and GR00T (Table 1, Section 5).

GYS accounts for their combined effects by optimizing distributional agreement against demonstrations. Unlike Dense-Jump, GYS allocates interior steps into the mid-to-late flow (typically up to τ≈0.65\tau \approx 0.65–0.800.80 in Table 8, Appendix C.3) to resolve trajectory curvature. Unlike SharpEuler, it places fewer interior nodes near τ→1\tau \to 1 (Figure 1(b)), crossing the late-time Lipschitz window in a single terminal step to avoid amplifying deviations and achieving the highest success rate across budgets.

We welcome further discussion if any concerns remain.

Reviewer_n3w1OpenReview ↗
Rating
3
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes GYS, a training-free method that optimizes the sampling schedule of flow-matching robot policies under a fixed function evaluation budget. It minimizes an energy distance between generated–demonstrated and generated–generated action differences, which match exactly when the two distributions do. The evaluations are experimented on RobotTwin 2.0 benchmarks. The results show GYS improves both success rates on sim and real-world tasks.

Critical strengths
  • Principled method to improve the sampling scheduler from demonstrations. It is based on the agreement between transformed generated and demonstrated action distributions in order to improve sample qualities given a fixed budget.

  • Experiments show consistent improved results on both simulation benchmarks and real-world robots, using different base VLA models. Experiments are well described.

Critical weaknesses
  • The improvements are not significant, thought they are consistent over experiments. In particular, the reported results are not averaged over no multiple runs, and with std.

  • The methodology section is a bit difficult to follow, as there is no discussion how such mathematical equations are related as an overall approach, and can be technically implemented.

  • The overhead computation of internal optimization steps are not well discussed.

Questions for authors
  • Q1: How much is over-head computation needed for each internal optimization steps.

  • Q2: How the demonstrations can be provided ? How to understand the recorded action chunks to become demonstrations in this proposed framework?

3 comments
AuthorsOpenReview ↗

Response to Reviewer n3w1

We appreciate your valuable feedback and support of our work.

To W1:

We re-ran Uniform and GYS for three additional rounds under identical settings and report the mean ± std over four rounds in Table A.

GYS consistently outperforms Uniform with small standard deviations (≤ 0.7%), confirming stable gains rather than evaluation noise. It improves the average success rate by 2.82.8 percentage points on both π0.5\pi_{0.5} and GR00T, offering clear practical value with zero inference overhead.

Table A. Success rate (%), mean ± std over four rounds. Avg. is mean ± std across six budgets (from Table 1 of the revised PDF).

Policy / ScheduleB=3B=3B=4B=4B=5B=5B=6B=6B=8B=8B=10B=10Avg.
π0.5\pi_{0.5} Uniform67.7 ± 0.668.2 ± 0.468.0 ± 0.567.7 ± 0.466.9 ± 0.466.6 ± 0.267.5 ± 0.6
π0.5\pi_{0.5} GYS70.9 ± 0.371.7 ± 0.370.8 ± 0.269.6 ± 0.469.3 ± 0.269.3 ± 0.370.3 ± 1.0
GR00T Uniform47.8 ± 0.748.6 ± 0.650.7 ± 0.752.1 ± 0.548.0 ± 0.647.7 ± 0.549.2 ± 1.8
GR00T GYS51.0 ± 0.551.9 ± 0.451.7 ± 0.452.5 ± 0.653.0 ± 0.551.8 ± 0.552.0 ± 0.7

To W2:

GYS is a training-free offline method optimizing flow-matching sampling schedules under BB neural function evaluations (NFEs). In Section 4 Method, Sec. 4.1 formulates the distribution-matching problem. To turn it into a computable objective, Sec. 4.2 constructs an energy distance between two pairwise-difference distributions. Sec. 4.3 implements this objective offline.

  1. Overall Approach (Secs. 4.1–4.2, Eqs. 5–10).
  • Problem Formulation (Sec. 4.1). Our goal is to match policy-generated and demonstrated action distributions under budget BB (Eq. 5). We apply a differencing operator Φν\Phi_\nu along horizon HH (Eq. 6) to remove schedule-invariant biases.
  • Optimization Objective (Sec. 4.2). Since a single demonstration per observation cannot yield the full target distribution, we compare two types of pairwise differences at each observation: the difference X−YX - Y between a generated chunk and a demonstration, and the difference X−X′X - X' between two independent generated chunks (Eq. 7). Pairing each difference with the observation summary (Eqs. 8–9), their energy distance (Eq. 10) is minimized, which vanishes if and only if the generated and demonstrated action distributions coincide (Proposition 1).
  1. Technical Implementation (Sec. 4.3, Eqs. 11–14).

Section 4.3 parameterizes each candidate schedule by step sizes h∈HBh \in \mathcal{H}_B (Eq. 11). VLAs generate action chunks under candidates to form joint samples for the two types of pairwise differences (Eq. 12). Candidates are scored by their energy distance (Eq. 13) and minimized via particle swarm optimization (PSO) to obtain τ⋆\boldsymbol{\tau}^\star (Eq. 14). At deployment, τ⋆\boldsymbol{\tau}^\star replaces the uniform time grid with no re-training, no architectural changes, and zero extra inference cost.

We have clarified this pipeline in the revised Section 4 overview.

To Q1 and W3:

Offline optimization introduces zero inference latency, and its single-GPU optimization overhead is far below policy fine-tuning (17 h for π0.5\pi_{0.5} and 6 h for GR00T N1.7 on 8×8\times A100-80GB, 50 RoboTwin tasks, 40k iters). On M=200M=200 calibration observations with N=8N=8 noise draws (MN=1,600MN=1,600 chunks), the VLM backbone encodes and caches conditioning features once, so internal steps evaluate only the action expert in micro-batches (chunks per GPU forward pass), as reported in Table B:

  • Per Euler step (1,6001,600 chunks): 1.51.5 s on an A100-80GB (micro-batch 512512) or 5.65.6 s on an L4-23GB (micro-batch 3232).
  • Per candidate schedule (BB steps): 1.5B1.5B s on an A100 or 5.6B5.6B s on an L4.
  • Full search (P=24P=24 particles, R=3R=3 PSO iterations): Evaluates P(1+R)=96P(1+R)=96 candidates.

Table B. Single-GPU offline optimization time for π0.5\pi_{0.5} per task (from Table 5 in Appendix B.3).

GPUMicro-batchB=3B=3B=4B=4B=5B=5B=6B=6B=8B=8B=10B=10
A10051215 min18 min20 min23 min28 min32 min
L43238 min47 min56 min65 min83 min101 min

To Q2:

  1. GYS requires no online interaction or extra data collection. It reuses offline VLA fine-tuning datasets (RoboTwin 2.0 clean dataset in sim; teleoperation on the real robot), sampling M=200M=200 observation–chunk pairs.

  2. In this calibration set, each demonstrated chunk is the ground-truth (GT) action sequence of horizon HH following a given observation in the recorded trajectories. Conditioned on these same observations, GYS optimizes the sampling schedule to minimize the distributional discrepancy between policy-predicted chunks and demonstrated GT chunks under a fixed budget.

We welcome further discussion if any concerns remain.

Thanks the authors for the clarification and additional results. They answered my questions. I like to keep my current rating.

Dear Reviewer n3w1,

Thank you very much for your valuable feedback and for recognizing our efforts in addressing your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.