acceptodds
Under review as a conference paper at ICLR 2027

FastOPD: On-Policy Distillation for Lightweight VLA Deployment

15 views

Abstract

Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies. In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combine it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model. We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, on-policy distillation, few-step distillation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_zM3FOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FastOPD, an efficient on-policy distillation method for compressing large flow-based VLA models into lightweight, few-step policies. Instead of evaluating the teacher at every denoising step, FastOPD uses supervision at a single sampled on-policy state, together with a self-consistency objective. The paper provides theoretical analysis and evaluates the method across several VLAs and a WAM in simulation, as well as one real-world robot task.

Critical strengths
  1. Effective and efficient. The method substantially reduces training and inference costs while retaining much of the teacher's performance. The results across different VLAs and a WAM are promising and suggest that the approach is fairly general.

  2. Interesting theoretical insights. The theoretical analysis helps explain why sparse on-policy supervision, combined with self-consistency, can work for few-step distillation.

Critical weaknesses
  1. Performance still falls short of the teacher. While the method improves over the student baselines, there is still a noticeable gap to the teacher, especially on RoboTwin and in the real-world experiment. It is unclear how much of this gap comes from the smaller student architecture versus the distillation method itself.

  2. Limited real-world evaluation. The real-world experiment only considers one relatively simple pick-and-place task (that presumably doesn't need VLA). It would be more convincing to see results on more challenging manipulation tasks.

Questions for authors
  1. What do the authors think is the main reason for the remaining gap between the student and teacher, particularly in the MolmoAct2 real-world experiment? Is there a way to further close this gap?

  2. Does distillation hurt the teacher's generalization capabilities? For example, how does the student perform on unseen objects, environments, or task variations compared to the teacher?

  3. By supervising only one sampled on-policy state per trajectory, could FastOPD be losing useful supervision from the other denoising steps? Have the authors tried supervising multiple sampled states to see whether this improves performance?

3 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

We sincerely thank reviewer zM3F for the encouraging comments and constructive suggestions. Below we provide point-by-point responses on your questions and concerns.

W1 & Q1. Performance still falls short of the teacher on RoboTwin and in the real-world experiment. Is there a way to further close this gap?

Performance still falls short of the teacher. While the method improves over the student baselines, there is still a noticeable gap to the teacher, especially on RoboTwin and in the real-world experiment. It is unclear how much of this gap comes from the smaller student architecture versus the distillation method itself.

What do the authors think is the main reason for the remaining gap between the student and teacher, particularly in the MolmoAct2 real-world experiment? Is there a way to further close this gap?

We appreciate the reviewer for the constructive comment. We acknowledge that a performance gap remains relative to the teacher and the 10-step base student.

To further investigate this gap, we conducted an additional real-robot experiment on a challenging long-horizon, multi-stage task (“open-close”). Since MolmoAct2 achieved only 52.0% success on this task, we adopted π0.5\pi_{0.5} as a more reliable teacher. All policies were evaluated over 50 trials. Further experimental details are provided in Appendix B.5 of the revised manuscript.

As shown in Table 1, our method achieves performance close to the teacher and outperforms the base student at both 4 and 10 steps. Thus, using a stronger teacher improves the performance and narrows the gap between the teacher and the student. This suggests that more reliable and less noisy teacher supervision may facilitate more effective knowledge transfer.

Further closing this gap will require addressing remaining challenges, including limited student capacity. Nevertheless, FastOPD provides a practical framework for efficiently distilling large foundation teacher policies into lightweight students, reducing the computational cost of on-policy distillation while enabling few-step inference.

Table 1. Real world results on the open-close task.

MethodSuccess rate (%)
π0.5\pi_{0.5} (teacher, NFE=10)82.0
Base (NFE=10)48.0
Base (NFE=4)20.0
Ours (NFE=4)72.0

W2. Limited real-world evaluation

The real-world experiment only considers one relatively simple pick-and-place task (that presumably doesn't need VLA). It would be more convincing to see results on more challenging manipulation tasks.

We agree that a more challenging task would better demonstrate the effectiveness of our approach. Following the reviewer’s suggestion, we conducted an additional real-world experiment on a long-horizon, multi-stage task: “open the drawer, pick up the black ball, put it in the drawer, and close the drawer.” (“open-close” task in Appendix B.5). As shown in Table 1 above, our method outperforms the base student at both 4 and 10 steps and even achieves performance close to the teacher.

Unlike the original pick-and-place task, this task requires the policy to execute a sequence of distinct manipulation stages, including articulated-object interaction, object grasping and placement, and drawer closure. Successful execution thus requires maintaining coherent behavior across multiple stages rather than completing a single pick-and-place operation. The strong performance of FastOPD on this task provides additional evidence that our distillation approach can preserve useful teacher capabilities in more complex manipulation settings.

AuthorsOpenReview ↗

Response by Authors (2/2)

Q2. Does distillation hurt the teacher’s generalization capabilities? For example, how does the student perform on unseen objects, environments, or task variations compared to the teacher?

As shown in Table 2 of the revised manuscript, our evaluation on the RoboTwin benchmark covers unseen environments under both clean and randomized settings. Although an absolute performance gap remains between the teacher and student, FastOPD demonstrates robustness to environmental variations in these unseen environments. Specifically, the distilled student achieves higher success rates under randomization at NFE=2, 4, and 10, while maintaining comparable performance at NFE=1, as detailed in Table 19 and 20 of the revised manuscript. These results suggest that the distilled student maintains comparable performance across the evaluated clean and randomized environments.

Q3. By supervising only one sampled on-policy state per trajectory, could FastOPD be losing useful supervision from the other denoising steps? Have the authors tried supervising multiple sampled states to see whether this improves performance?

We appreciate the reviewer’s valuable suggestion. To investigate this, we evaluated an augmented supervision variant that queries the teacher at two sequential student rollout states on LIBERO.

As shown in Table 2, supervising two states yields comparable performance to our single-state approach, suggesting that a single on-policy state already provides effective teacher guidance without requiring additional teacher evaluations. This supports our design choice of using single-state supervision for efficient distillation.

Table 2. Ablation study on the number of supervised on-policy states on LIBERO. Success rates (%) are reported across different NFEs.

# of supervised statesNFE=12410
Base69.171.472.871.6
1 (default)77.381.880.379.0
277.580.681.479.5

We thank the reviewer again for the valuable feedback, and hope these revisions address the reviewer's concerns.

AuthorsOpenReview ↗

Follow-up on Author Response

Dear reviewer zM3F,

Thank you again for your time and constructive comments. As the discussion period is ended within a couple of days, we would like to kindly follow up and ask whether our responses have addressed your concerns. If any concerns remain, we would be glad to address them within the remaining time.

Thank you very much for your effort in reviewing our work.

Reviewer_Z7pSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FastOPD distills large VLA and world-action teachers into a 451M-parameter student. It combines teacher velocity supervision at a sampled student state with self-consistency across denoising intervals, allowing the student to act with fewer denoising steps. Experiments on LIBERO, RoboTwin, and a real robot compare success rates, inference latency, and distillation cost.

Critical strengths
  1. The student achieves a useful speed–success trade-off. On LIBERO, two-step success rises from 71.4% to 81.8%, with model inference taking 66 ms on an RTX 3090. The one-step improvements on RoboTwin with different teachers suggest that the benefit is not specific to one teacher or benchmark.

  2. The evaluation includes training cost and a physical deployment. On the real-robot pnp-plate task, success improves from 42% to 50% at the same four-step inference setting (Figure 5). This tests the compressed policy on hardware, although the gain is modest.

Critical weaknesses
  1. The ablation does not fully isolate the value of external teacher supervision. Table 5 compares the combined method with self-consistency alone and teacher supervision alone. Self-consistency without an anchor collapses, but this does not establish that the anchor must come from the teacher. Appendix C discusses a flow-matching anchor, as in SnapFlow, without reporting a numerical comparison. The gains could partly reflect adding an anchored shortcut objective to the base student.

  2. The simulation comparisons use one training seed per configuration (Appendix B.4), leaving the stability of the method ranking unclear. Some adapted baselines perform substantially below the initial student, particularly CTM on RoboTwin (Table 2); this alone does not establish that they were poorly tuned. The real-robot gain is a net difference of four successes over 50 trials per policy, and the paired outcomes are not reported (Appendix B.5).

Questions for authors
  1. Can the authors compare teacher anchoring with flow matching plus the same self-consistency objective, using the same initialization, action data, and update budget? This would help clarify how much of the gain comes from the external teacher.

  2. For a representative benchmark, how much do the main result and the closest competing baseline vary across training seeds, and how were the adapted baselines’ loss weights selected? For the existing paired robot trials, please report the number of outcomes that change in each direction.

3 comments
AuthorsOpenReview ↗

Response by Authors (1/2)

We appreciate the reviewer Z7pS for your detailed comments and valuable suggestions. Below we provide point-by-point responses on your questions and concerns.

W1&Q1. Can the authors compare teacher anchoring with flow matching plus the same self-consistency objective?

The ablation does not fully isolate the value of external teacher supervision. Table 5 compares the combined method with self-consistency alone and teacher supervision alone. Self-consistency without an anchor collapses, but this does not establish that the anchor must come from the teacher. Appendix C discusses a flow-matching anchor, as in SnapFlow, without reporting a numerical comparison. The gains could partly reflect adding an anchored shortcut objective to the base student.

Can the authors compare teacher anchoring with flow matching plus the same self-consistency objective, using the same initialization, action data, and update budget? This would help clarify how much of the gain comes from the external teacher.

We thank the reviewer for the thoughtful suggestion. We would like to clarify that the primary goal of FastOPD is to efficiently distill large foundation VLAs into lightweight students, rather than to establish the superiority of teacher anchoring over data anchoring.

Since our focus is teacher distillation, data-anchored methods such as SnapFlow [1] address a different objective of learning flow maps directly from demonstration data. Thus, comparing these approaches evaluates the choice of supervision anchor rather than isolating the effect of on-policy sampling alone.

That said, we appreciate the valuable suggestion and conducted a head-to-head comparison on LIBERO using data-anchored baselines as well as a combined variant under identical settings. For SnapFlow, we applied the self-consistency objective at fixed endpoints, following the original paper [1]. As shown in Table 1, adding the data-anchored flow matching loss to our method does not yield consistent improvements over ours alone. These results validate the effectiveness of the teacher anchor, demonstrating that our on-policy teacher distillation already provides sufficient supervision without requiring additional data-level objectives. All variants used the same student initialization, action data, and update budget.

Table 1. Ablation study on LIBERO. Success rates (%) are reported across different denoising steps.

MethodAnchorNFE=12410
Base-69.171.472.871.6
SnapFlowdata77.274.770.771.2
SC + FMdata77.176.873.774.0
Oursteacher77.381.880.379.0
Ours + FMteacher + data77.380.580.079.3

[1] Luan, Wuyang, et al. "Snapflow: One-step action generation for flow-matching vlas via progressive self-distillation." arXiv preprint arXiv:2604.05656 (2026).

W2-1 & Q2-1. How much do the main result and the closest baseline vary across training seeds for a representative benchmark?

To evaluate statistical robustness, we trained FastOPD and the closest baseline, CTM, on LIBERO across three random training seeds. As shown in Table 2, FastOPD consistently outperforms CTM across all denoising steps. These results demonstrate the robustness of our improvements.

Table 2. Success rates (%) on LIBERO across three training seeds, reported as mean ± standard deviation.

MethodNFESpatialObjectGoalLongAvg.
CTM180.9 ± 2.886.3 ± 3.174.0 ± 3.141.1 ± 4.470.6 ± 1.7
276.5 ± 1.080.1 ± 4.274.0 ± 3.136.7 ± 8.266.8 ± 2.2
472.3 ± 4.176.1 ± 3.072.0 ± 0.741.2 ± 5.665.4 ± 1.2
1072.4 ± 0.477.7 ± 3.472.5 ± 0.341.3 ± 2.966.0 ± 1.4
Ours184.6 ± 2.494.7 ± 0.584.9 ± 2.153.3 ± 4.679.3 ± 2.0
285.3 ± 1.597.2 ± 0.488.9 ± 1.459.8 ± 1.982.8 ± 1.0
484.4 ± 1.295.7 ± 0.688.6 ± 2.657.6 ± 0.581.5 ± 1.2
1082.3 ± 2.394.0 ± 0.685.8 ± 1.457.3 ± 1.079.8 ± 1.2
AuthorsOpenReview ↗

Response by Authors (2/2)

W2-2 & Q2-2. How were the adapted baselines’ loss weights selected?

We primarily followed the loss configurations from prior works, with adjustments by experiments in our distillation setting. We kept the selected settings fixed across LIBERO and RoboTwin 2.0.

  • CTM [1]: We used L=LCTM+LDSM\mathcal{L} = \mathcal{L}_{\mathrm{CTM}} + \mathcal{L}_{\mathrm{DSM}}, with equal outer loss coefficients of 1.0. This follows the relative weighting used in the released Consistency Policy configuration [2]. In our adaptation, the DSM term supervises the student's instantaneous velocity using the frozen teacher's velocity field.

  • iMF [3]: We used a single teacher-supervised Pseudo-Huber objective with δ=1\delta = 1, without additional loss weighting. We also set P(t=s)=0.75P(t=s) = 0.75 for the time-sampling strategy.

  • DMD [4,5]: Following OneDP [4], which adapts DMD to robot action generation, we optimized the generator and critic separately using the distribution-matching and denoising objectives, respectively. For stable generator updates, we applied per-sample normalization by the mean absolute magnitude of the teacher's correction, following the normalization principle of DMD, Eq.(8) [5].

[1] Kim, Dongjun, et al. "Consistency trajectory models: Learning probability flow ode trajectory of diffusion." ICLR 2024.

[2] Prasad, Aaditya, et al. "Consistency policy: Accelerated visuomotor policies via consistency distillation." RSS 2024.

[3] Geng, Zhengyang, et al. "Improved mean flows: On the challenges of fastforward generative models." CVPR 2026.

[4] Wang, Zhendong, et al. "One-step diffusion policy: Fast visuomotor policies via diffusion distillation." ICML 2025.

[5] Yin, Tianwei, et al. "One-step diffusion with distribution matching distillation." CVPR 2024.

W2-3 & Q2-3. For the real-robot experiment, report the number of outcomes that change in each direction.

Thank you for the valuable suggestion. We now report paired outcomes for both the original pnp-plate task and the additional open-close task, with 50 trials per comparison. As shown in Table 3, FastOPD succeeds when the 4-step base student fails in 8 and 27 trials, respectively, while the reverse occurs in only 4 and 1 trials. These results have also been added to the revised manuscript.

Table 3. Paired outcomes on the real-world pnp-plate and open-close tasks. Each comparison includes 50 trials.

TaskCompared toBoth succeedOurs onlyBase onlyBoth fail
pnp-plateBase (NFE=4)178421
Base (NFE=10)169619
open-closeBase (NFE=4)927113
Base (NFE=10)2115311

We sincerely appreciate your thoughtful feedback, which has helped us improve and strengthen our work.

AuthorsOpenReview ↗

Follow-up on Author Response

Dear reviewer Z7pS,

Thank you again for your time and constructive comments. As the discussion period is ended within a couple of days, we would like to kindly follow up and ask whether our responses have addressed your concerns. If any concerns remain, we would be glad to address them within the remaining time.

Thank you very much for your effort in reviewing our work.

Reviewer_AucVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces FastOPD, an on-policy distillation method that distills larger VLA policies more efficiently. The resulting policies require less denoising steps while distillation itself is faster. FastODP achieves this through matching the velocity state at a single policy state paired with a self-consistency loss. The paper shows that the two losses are complementary to each other and result in faster inference and better policy distillation across different benchmarks with different teachers.

Critical strengths
  • Relevant and strong motivation
  • Simple and cheap to train
  • Evaluated with diverse teachers (π0.5, LingBot-VLA, Fast-WAM, MolmoAct2) across simulation and a real robot.
  • Throughout ablations of design choices.
Critical weaknesses
  • Limited transfer from teacher:
    • On Libero, the student gains from the teacher but the base model barely degrades with fewer steps, making step reduction nearly free
    • RoboTwin: FastODP gains in few-step performance but barely learns from the teacher, staying far below teacher and below 10-step base, similar on real robot
    • Libero gains mostly on simple tasks and negative or within noise on harder long-horizon suite
    • The base student already performs well with few steps (LIBERO: 71.4% at 2 vs. 71.6% at 10 steps. RoboTwin: 55.6% vs. 60.4%), so step reduction only matters at 1 step on RoboTwin, where no teacher transfer is observed. The claimed joint compression (model size and sampling steps) is therefore never demonstrated in a setting where both axes matter.
  • Missing ablations and baselines:
    • The claim that on-policy distillation is required is never verified. An off policy ablation would support the claim
    • The SC-only ablation collapses by construction, as it regresses onto the students own unanchored velocity field. A teacher-free counterpart is needed to isalotate the gain (i.e. with normal flow matching objective on data)
    • A continued flow-matching baselines for same number of steps would show its not just additional training, which is especially important given the next point:
    • The teacher’s success rate at NFE = 1/2/4 is not reported. Since the theory targets a teacher discretized with two Euler steps per interval, this is the natural reference point for whether the few-step student is actually better than simply running the teacher with fewer steps.
  • Weak base student: SmolVLA reaches 71.6% on LIBERO, while the SmolVLA paper reports 87.3%. Improvements over the student may partly reflect additional training.
  • Statistical Robustness: All experiments are single-seeded. A lot of experiments might be within seed noise, e.g., in Fig. 4b FastOPD and the baselines are nearly identical at several checkpoints.
  • Real Robot:
    • Small gains, unexplained. Other distillation methods not evaluated. The number of trials are not in the main text.
    • Training efficiency claim: The ODP curve is only evaluated at a few checkpoints between 20 and 57h. ODP might reach the reported level earlier.
  • The acronym NFE is never defined, and the text switches between “inference steps,” “sampling steps,” and “denoising steps.”
  • The meaning of bold and underlined numbers in Table 2 (and Table 1) is not stated.
Questions for authors
  1. Why are the RoboTwin gains smaller than on LIBERO, and why does FastOPD fall below the base student at 4 and 10 steps?
  2. What are the teacher success rates at NFE = 1/2/4?
  3. Why is MolmoAct2 used as the real-robot teacher rather than π0.5?
  4. Why distill the WAM into SmolVLA rather than a smaller WAM student?
  5. Can you report results over multiple seeds and explain how checkpoints and number of steps are selected?
7 comments
AuthorsOpenReview ↗

Response by Authors (1/4)

We sincerely appreciate reviewer AuCV for the encouraging comments and constructive suggestions. Below, we provide point-by-point responses on your questions and concerns.

W1-1. Step reduction and Joint Compression

On Libero, the student gains from the teacher but the base model barely degrades with fewer steps, making step reduction nearly free

The base student already performs well with few steps, so step reduction only matters at 1 step on RoboTwin, where no teacher transfer is observed. The claimed joint compression (model size and sampling steps) is therefore never demonstrated in a setting where both axes matter.

We first emphasize that FastOPD focuses on efficient teacher distillation, rather than step reduction alone. Conventional on-policy distillation requires costly repeated teacher evaluations along student trajectories. FastOPD addresses this by combining single-state on-policy teacher supervision with self-consistency, using a flow map to efficiently propagate teacher guidance while naturally supporting few-step inference.

Table 1 quantitatively illustrates the benefits of a lightweight policy from an inference-latency perspective. For the 3.3B teacher, the VLM prefix costs 136.6 ms, so reducing the steps alone still leaves a latency of 155 ms at NFE=1, motivating model-size reduction. Once the model is small, denoising accounts for a larger share of latency: reducing the base student from 10 to 2 steps decreases latency from 193 ms to 65 ms. These results illustrate the complementary benefits of reducing model size and denoising steps.

Table 1. VLM prefix and E2E inference latency (ms) on LIBERO.

ModelVLM prefixNFE=10NFE=4NFE=2NFE=1
π0.5\pi_{0.5} (teacher)~137301203171155
SmolVLA~ 31193976549
Ours~ 31193986650

W1-2. LIBERO gains mostly on simple tasks and negative or within noise on harder long-horizon suite.

To assess the consistency of gains on LIBERO-Long, we trained FastOPD with three random seeds and compared it with the base SmolVLA. As shown in Table 2, FastOPD achieves 59.8 ± 1.9% success at NFE=2, compared with 54.4% for the base student. At NFE=1 and 10, its mean performance is also numerically higher than the base student, while at NFE=4 it is slightly lower.

Table 2. Success rates (%) on LIBERO. FastOPD results are reported as mean ± standard deviation across three training seeds.

MethodNFESpatialObjectGoalLongAvg.
Base SmolVLA174.673.277.451.269.1
276.874.479.854.471.4
479.074.879.258.072.8
1077.275.079.454.671.6
Ours184.6 ± 2.494.7 ± 0.584.9 ± 2.153.3 ± 4.679.3 ± 2.0
285.3 ± 1.597.2 ± 0.488.9 ± 1.459.8 ± 1.982.8 ± 1.0
484.4 ± 1.295.7 ± 0.688.6 ± 2.657.6 ± 0.581.5 ± 1.2
1082.3 ± 2.394.0 ± 0.685.8 ± 1.457.3 ± 1.079.8 ± 1.2
AuthorsOpenReview ↗

Response by Authors (2/4)

W1-3 & Q1. Remaining gap with teacher and multi-step base student

RoboTwin: FastODP gains in few-step performance but barely learns from the teacher, staying far below teacher and below 10-step base, similar on real robot

Why are the RoboTwin gains smaller than on LIBERO, and why does FastOPD fall below the base student at 4 and 10 steps?

We appreciate the reviewer for the constructive comment. We agree that there is a remaining performance gap with a teacher and a 10-step base student.

To further investigate this gap, we conducted an additional real-robot experiment on a challenging long-horizon, multi-stage task (“open-close”). Since MolmoAct2 achieved only 52.0% success on this task, we adopted π0.5\pi_{0.5} as a more reliable teacher. All policies were evaluated over 50 trials. The experimental setup and implementation details are provided in Appendix B.5 of the revised manuscript.

As shown in Table 3, FastOPD achieves performance close to the teacher and outperforms the base student at both 4 and 10 steps. Using a stronger teacher improves the performance and narrows the gap between the teacher and the student highlighted by the reviewer. This result suggests that effective transfer is possible with a strong teacher, although it does not isolate the effect of teacher choice or explain the original MolmoAct2 gap, and addressing this gap further requires considering additional factors such as student capacity limits.

Nevertheless, we would like to emphasize that our main contribution lies in proposing a flow map-based framework designed to efficiently distill large-scale foundation teacher policies into lightweight policies.

Table 3. Real world results on the open-close task.

MethodSuccess rate (%)
π0.5\pi_{0.5} (teacher, NFE=10)82.0
Base (NFE=10)48.0
Base (NFE=4)20.0
Ours (NFE=4)72.0

W2-1. Missing ablations and baselines

The claim that on-policy distillation is required is never verified. An off policy ablation would support the claim

The SC-only ablation collapses by construction, as it regresses onto the students own unanchored velocity field. A teacher-free counterpart is needed to isolate the gain (i.e. with normal flow matching objective on data)

A continued flow-matching baselines for same number of steps would show its not just additional training, which is especially important given the next point

We would like to clarify that the primary goal of FastOPD is to efficiently distill large foundation VLAs into lightweight students, rather than merely accelerating inference through flow map construction. To this end, FastOPD combines single-state on-policy teacher supervision with self-consistency to efficiently propagate teacher guidance while supporting few-step inference.

Unlike data-anchored shortcut methods that rely on flow matching over demonstrations, FastOPD explicitly anchors the student to the pretrained teacher's dynamics. Thus, incorporating a teacher-free flow-matching baseline introduces a different supervision anchor, making it an anchor-to-anchor comparison rather than isolating the effect of on-policy sampling alone.

That said, we additionally conducted ablation studies with all requested variants on LIBERO, as summarized below. These results confirm that our performance gains do not simply stem from extra training, but rather from effectively leveraging the teacher anchor alongside on-policy supervision and self-consistency to boost few-step performance. Furthermore, we also evaluated a variant combining both teacher and data anchors. The results demonstrate that adding the data anchor (i.e., standard flow matching) does not yield consistent improvements over our method, confirming that our on-policy teacher distillation already provides sufficient and effective supervision on its own.

Table 4. Ablation study on LIBERO.

MethodAnchorNFE=12410
Base SmolVLA-69.171.472.871.6
Continued FMdata57.071.671.571.0
SC + FMdata77.176.873.774.0
Off-policy distillation + SCteacher67.462.260.957.3
On-policy distillation + SC (Ours)teacher77.381.880.379.0
On-policy distillation + SC + FMteacher + data77.380.580.079.3
AuthorsOpenReview ↗

Response by Authors (3/4)

W2-2&Q2. Does the few-step student outperform the teacher at the same NFEs (1/2/4)?

The teacher’s success rate at NFE = 1/2/4 is not reported. Since the theory targets a teacher discretized with two Euler steps per interval, this is the natural reference point for whether the few-step student is actually better than simply running the teacher with fewer steps.

Thank you for your valuable suggestion. We would like to clarify that our primary motivation is to develop compact VLA models suitable for deployment on resource-constrained, consumer-grade hardware, rather than to reach large teacher’s performance. In this context, Theorem 1 in the manuscript provides a theoretical justification that our training objective is designed to approximate the teacher's action distribution. It does not guarantee identical teacher performance under imperfect optimization.

Indeed, the teachers already achieve strong performance with few Euler steps. However, reducing sampling steps alone does not eliminate the computational burden of foundation teacher models, and high success rates have limited practical value if deployment remains infeasible. Even at 1-step, the teachers require 155 ms, compared with only 66 ms for our 2-step, as shown in Table 1 above. While our students do not surpass the teachers in absolute success rate, FastOPD improves the 2-step base SmolVLA on both simulation benchmarks.

Table 5. Success rates (%) on LIBERO and RoboTwin2.0. For RoboTwin, averaged over clean and randomized settings.

BenchmarkModelNFE=12410
LIBEROπ0.5\pi_{0.5} (teacher)97.397.297.597.5
Ours77.381.880.379.0
RoboTwin 2.0LingBot-VLA (teacher)87.187.088.485.8
Ours51.257.759.958.3

W3. Weak base student on LIBERO

SmolVLA reaches 71.6% on LIBERO, while the SmolVLA paper reports 87.3%. Improvements over the student may partly reflect additional training.

We would like to assure the reviewer that we use the publicly released SmolVLA [1] LIBERO checkpoint without modification (Appendix B.1), so that the base student is not affected by our own training choices. We observed that evaluations of the same checkpoint report 72.5% [2] and 64.0% [3] on LIBERO, respectively, under different evaluation settings. Our measurement falls within the range of these reported results.

[1] M. Shukor et al., "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics," arXiv:2506.01844, 2025.

[2] "Cannot reproduce SmolVLA results on LIBERO benchmark," GitHub issue #2354, huggingface/lerobot. https://github.com/huggingface/lerobot/issues/2354

[3] J. Lee et al., "Training-Free Inference-Budget Reduction in Flow-Matching Vision–Language–Action Models for Robot Manipulation: An Evaluation Against a Prespecified Task-Success Margin," Robotics 2026.

W4 & Q5. Statistical Robustness

All experiments are single-seeded. A lot of experiments might be within seed noise, e.g., in Fig. 4b FastOPD and the baselines are nearly identical at several checkpoints.

Can you report results over multiple seeds and explain how checkpoints and number of steps are selected?

To assess statistical robustness, we trained FastOPD and the closest baseline, CTM, on LIBERO across three random training seeds. As shown in Table 6, our method consistently outperforms CTM.

For checkpoint selection, we trained each model for 10,000 iterations on LIBERO and 50,000 iterations on RoboTwin, and evaluated the final checkpoint at 1, 2, 4, and 10 denoising steps (Appendix B.2).

Table 6. Success rates (%) on LIBERO across three training seeds, reported as mean ± standard deviation.

MethodNFESpatialObjectGoalLongAvg.
CTM180.9 ± 2.886.3 ± 3.174.0 ± 3.141.1 ± 4.470.6 ± 1.7
276.5 ± 1.080.1 ± 4.274.0 ± 3.136.7 ± 8.266.8 ± 2.2
472.3 ± 4.176.1 ± 3.072.0 ± 0.741.2 ± 5.665.4 ± 1.2
1072.4 ± 0.477.7 ± 3.472.5 ± 0.341.3 ± 2.966.0 ± 1.4
Ours184.6 ± 2.494.7 ± 0.584.9 ± 2.153.3 ± 4.679.3 ± 2.0
285.3 ± 1.597.2 ± 0.488.9 ± 1.459.8 ± 1.982.8 ± 1.0
484.4 ± 1.295.7 ± 0.688.6 ± 2.657.6 ± 0.581.5 ± 1.2
1082.3 ± 2.394.0 ± 0.685.8 ± 1.457.3 ± 1.079.8 ± 1.2

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comments by Authors

We sincerely thank all reviewers for their thoughtful and constructive feedback.

We are encouraged that the reviewers recognize the efficiency and practical benefits of our distillation framework (AucV, Z7pS, zM3F), its applicability across diverse teacher models and evaluation settings (AucV, zM3F), and its interesting theoretical insights (zM3F). Here, we would like to clarify the primary motivation and contribution of FastOPD and address the major concerns shared across multiple reviews.

1. Motivation and Contribution of FastOPD

We clarify that FastOPD aims to efficiently distill large foundation VLAs into lightweight student policies, rather than merely accelerating inference via flow map construction. Conventional on-policy distillation requires costly repeated teacher evaluations along student trajectories. FastOPD addresses this by combining single-state on-policy teacher supervision with self-consistency, which propagates local teacher guidance to finite-interval transitions for efficient teacher distillation while naturally supporting few-step inference.

Unlike data-anchored shortcut methods that rely on flow matching over demonstration data, FastOPD explicitly anchors the student to the pretrained teacher's dynamics. Here, the flow map serves not as the primary objective, but as a mechanism to efficiently propagate teacher supervision.

2. Statistical robustness

In response to the reviewers’ suggestions, we evaluated FastOPD and CTM, a closest baseline, across three training seeds on LIBERO. FastOPD consistently outperforms CTM across these seeds, achieving 82.8 ± 1.0% success at NFE=2.

Table 1. Success rates (%) on LIBERO.

MethodNFESpatialObjectGoalLongAvg.
CTM180.9 ± 2.886.3 ± 3.174.0 ± 3.141.1 ± 4.470.6 ± 1.7
276.5 ± 1.080.1 ± 4.274.0 ± 3.136.7 ± 8.266.8 ± 2.2
472.3 ± 4.176.1 ± 3.072.0 ± 0.741.2 ± 5.665.4 ± 1.2
1072.4 ± 0.477.7 ± 3.472.5 ± 0.341.3 ± 2.966.0 ± 1.4
Ours184.6 ± 2.494.7 ± 0.584.9 ± 2.153.3 ± 4.679.3 ± 2.0
285.3 ± 1.597.2 ± 0.488.9 ± 1.459.8 ± 1.982.8 ± 1.0
484.4 ± 1.295.7 ± 0.688.6 ± 2.657.6 ± 0.581.5 ± 1.2
1082.3 ± 2.394.0 ± 0.685.8 ± 1.457.3 ± 1.079.8 ± 1.2

3. Additional real-world experiment

To further evaluate our method on a more challenging task, we conducted an additional long-horizon, multi-stage real-world experiment (“open-close” task in Appendix B.5). Since MolmoAct2 achieved only 52.0% success rate on this task, we adopted π0.5\pi_{0.5} as a more reliable teacher. All policies were evaluated over 50 trials. The results confirm that FastOPD successfully scales to more complex manipulation scenarios through effective teacher distillation. Further experimental details are provided in Appendix B.5 of the revised manuscript.

Table 2. Real world results on the open-close task.

MethodSuccess rate (%)
π0.5\pi_{0.5} (teacher, NFE=10)82.0
Base (NFE=10)48.0
Base (NFE=4)20.0
Ours (NFE=4)72.0

Dear Reviewers,

Thank you for the review. The authors have responded to the points you raised, and the response period closes on 18 November (AoE). I would be grateful if you could post a brief reply on whether your concerns are resolved, partly resolved, or still open, so that the authors have a fair chance to address them. If anything still needs clarifying, please ask now while they can still respond.

Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.