acceptodds
Under review as a conference paper at ICLR 2027

FlatVLA: Robust Vision-Language-Action Model with Sharpness-Aware Minimization

Abstract

Vision-Language-Action (VLA) models integrate visual observations and language instructions to generate actions, yet remain vulnerable to multimodal distribution shifts and errors arising during sequential execution. Existing robustness methods target predefined variations, making it difficult to cover a broad range of variations across VLA inputs and actions. To address this challenge, we propose FlatVLA, a flatness-aware training framework that adapts Sharpness-Aware Minimization (SAM) to the sequential structure of VLA action prediction. FlatVLA applies larger weights to positions near the beginning of each action chunk, as errors at these positions can affect a larger portion of the remaining prediction. Our theoretical analysis connects local smoothness to policy sensitivity and compounding error, motivating this sequential weighting. Across recent models, FlatVLA shows improvements on performance and robustness across diverse evaluation factors and settings, with further analysis of model behaviors. These gains also extend from simulation to real-world tasks, as demonstrated via hardware experiments. Overall, FlatVLA optimizes parameter space smoothness while accounting for sequential error accumulation in action prediction, improving robustness without explicitly tailoring training to specific deployment conditions.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Robust Vision-Language-Action, Sharpness-Aware Minimization, Multimodal Robustness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_DH2qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlatVLA, which combines Sharpness-Aware Minimization (SAM) with sequential weighting of action prediction losses to improve VLA robustness. The authors theoretically connect parameter-space flatness to an upper bound on compounding error. Experiments on four VLA models, LIBERO, LIBERO-Plus, and real-world robotic tasks demonstrate improvements in performance and robustness.

Critical strengths

The paper attempts to explain closed-loop robustness in VLAs from the perspective of parameter-space flatness while considering the sequential structure of action prediction. This is an interesting and potentially significant research direction. The method requires no changes to the original model architecture or inference procedure.

Critical weaknesses
  1. Prior work cited in the paper has already applied SAM to VLA training [1]. The main algorithmic addition in FlatVLA is Sequential Weighting, and the two methods share much of their core optimization approach. The paper provides neither a direct experimental comparison with [1] nor ablations of standard SAM and Sequential Weighting alone. The independent contribution of the proposed weighting strategy remains unclear.

  2. Proposition 1 relies on additional sensitivity assumptions, while Theorem 1 only establishes a comparison of upper bounds on compounding error. The connection between episode-level error propagation and action-chunk-level weighting, as well as the choice of a fixed γ\gamma, is not sufficiently justified.

  3. Although the success rates improve, the loss landscape, predictive uncertainty, and action smoothness analyses do not directly demonstrate reduced policy sensitivity or compounding error. The suggested connection between increased uncertainty in failed executions and recovery potential also lacks direct evidence.

Questions for authors
  1. What is the main methodological contribution of FlatVLA compared with the cited work [1]? Why is there no direct experimental comparison? Could the authors provide ablations of standard SAM, Sequential Weighting alone, and the full FlatVLA method to clarify the benefits of the proposed weighting strategy?

  2. Could the authors further explain the applicability of Assumptions 4 and 5, and how Eq. (4) supports the weighting design in Eq. (9)? Is there experimental evidence supporting the choice of γ\gamma?

  3. Could the authors directly measure policy sensitivity and error propagation under perturbations to validate the theoretical analysis and the claimed robustness improvement?

[1] Flatness Preserves Instruction Following in Vision-Language-Action Models.

5 comments
AuthorsOpenReview ↗

Response to Reviewer DH2q (Part 1/2)

We thank the reviewer for the constructive feedback and address the novelty concerns below. We have incorporated additional analyses and theoretical clarifications into the revised manuscript. For convenience, all changes are highlighted in orange.

 

W1, Q1: Contribution beyond SAM and Ablations

FlatVLA differs from prior work [1] in both its contribution and robotics-specific motivation. Unlike prior work [1], FlatVLA goes beyond applying SAM by incorporating the sequential structure of action chunks to account for error propagation during execution. We also provide an intuitive theoretical analysis linking loss sharpness, policy sensitivity, and sequential error propagation under stated conditions.

We additionally compare the baseline as follows: (1) the original baseline, (2) SAM only (γ=0\gamma=0; partial reproduce of [1]), (3) sequential weighting without SAM (SW only), and (4) FlatVLA The result shows that FlatVLA consistently improves in all cases, with gains across all backbone and methods.

LIBERO

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA75.8----78.5
OpenVLA-OFT96.896.996.897.4
π0\pi_093.393.593.594.3
π0.5\pi_{0.5}95.896.395.997.4

LIBERO-PLUS

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA16.5----19.5
OpenVLA-OFT67.468.368.170.1
π0\pi_045.745.746.047.3
π0.5\pi_{0.5}80.281.180.682.6

We provided full LIBERO and LIBERO-PLUS results and the corresponding revisions are included in the revised manuscript.

[1] Zhang and Bisk, Flatness Preserves Instruction Following in Vision-Language-Action Models Arxiv'26

AuthorsOpenReview ↗

Response to Reviewer DH2q (Part 2/2)

W2, Q2 Part 1: Theoretical Framework and Sequential Weighting

Our analysis links local loss sharpness, policy sensitivity, and sequential error propagation. Assumptions 4–5 specify the local conditions under which Proposition 1 relates sharpness to policy sensitivity, while Theorem 1 conditionally compares compounding-error upper bounds without guaranteeing lower realized errors. Sequential weighting in Eq.(9) is motivated by, rather than directly derived from, temporal error propagation in Eq.(4): during open-loop execution, earlier action errors can affect subsequent physical states while later actions are not recomputed.

We acknowledge that γ\gamma is a fixed design hyperparameter, not an estimated dynamics constant. We set γ=0.7\gamma=0.7 for OpenVLA-OFT (H=8H=8) and π0.5\pi_{0.5} (H=10H=10), assigning 61.6% and 60.4% of the normalized weight to the first half of each chunk, respectively. For the longer chunks of π0\pi_0 (H=50H=50), we use a more conservative γ=0.5\gamma=0.5, assigning 51.0% to the first half. These settings adjust the emphasis on earlier actions according to the predicted action horizon.

To test the novelty of FlatVLA, we evaluated replanning intervals (E∈(E\in { 10,20,30,5010,20,30,50 } )) with π0\pi_0 (H=50)(H=50), which has the longest predicted action horizon. Here, FlatVLA consistently shows improvement across all approach, providing empirical support of our apporach.

Replan (EE)BaselineSAM onlySW onlyFlatVLA
E=10E=1093.393.593.594.3
E=20E=2088.189.189.291.0
E=30E=3083.484.383.687.3
E=50E=5068.770.169.975.6

We provided these explnataions and complete results for above experiments in the revised manuscript.

 

W3, Q2 Part 2, Q3: Policy Senstivity and Error Propagation

  • Policy Sensitivity: We distinguish training-loss sharpness from execution-level policy robustness. Since VLA models are trained with action-supervised objectives, training-loss sharpness provides an optimization-level perspective on the learned policy. To quantify this, we measure local loss sharpness as Sρ(θ)=max⁡∥ϵ∥2≤ρ[L(θ+ϵ)−L(θ)]S_\rho(\theta)=\max_{\|\epsilon\|_2\leq\rho}[\mathcal L(\theta+\epsilon)-\mathcal L(\theta)]. As shown below, FlatVLA achieves the lowest measured sharpness within each backbone.

    BackboneBaselineSAM onlySW onlyFlatVLAReduction vs. Baseline ↓\downarrow
    OpenVLA0.2143——0.070767.0%
    OpenVLA-OFT0.01020.00810.01310.003862.7%
    π0\pi_00.16310.05520.15410.018788.5%
    π0.5\pi_{0.5}0.13250.03220.15390.010891.8%

    Both Fig. 5 and our quantitative sharpness results point to reduced loss sensitivity (Assumption 4), whereas the robustness gains under multimodal perturbations on LIBERO-PLUS substantiate policy sensitivity (Assumption 5). This empirical evidence further validates our theoretical connection between loss sharpness and policy sensitivity.

  • Error Propagation: To examine the effect of early action errors, we conduct additional experiments by injecting action noise during the first 20 rollout steps (1–20) or the next 20 steps (21–40) on π0.5\pi_{0.5}. As shown below, early-stage perturbations cause substantially greater performance degradation than later-stage perturbations, consistent with the greater downstream influence of earlier action errors. As results, FlatVLA exhibits the smallest performance drops under both settings, providing further evidence of its robustness to temporally localized action perturbations.

    Noise Injection (π0.5\pi_{0.5})BaselineSAM onlySW onlyFlatVLA
    No Noise92.292.992.594.3
    First 20 steps (1–20)59.9 (−32.3 pp)63.3 (−29.6 pp)60.5 (−32.0 pp)67.5 (−26.8 pp)
    Next 20 steps (21–40)87.4 (−4.8 pp)88.8 (−4.1 pp)87.1 (-5.4pp)90.7 (−3.6 pp)

    Here, () indicates performance drop pp relative to No Noise.

 

In last, we revised the interpretation of predictive uncertainty (Table 5), clarifying that the observed success–failure uncertainty gap does not directly demonstrate recovery capability.

AuthorsOpenReview ↗

Brief Reminder for Reviewer DH2q

Dear Reviewer DH2q,

Thank you again for your time and feedback. We hope our responses and the updated manuscript (e.g., the additional experiments and theoretical clarifications) have adequately addressed the questions and concerns raised. For the convenience,, we have also posted an overall summary of where each revision was made in our general response at the top. As the end of the discussion period approaches, we would be glad to respond if any further questions or concerns remain.

Best regard,

The Authors

Reviewer_H4uuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

FlatVLA combines early-action loss weighting with SAMSharpness-Aware Minimization (SAM) for VLA fine-tuning, motivated by a conditional error-propagation analysis. Experiments cover four VLA backbones, LIBERO, LIBERO-Plus, and real-world manipulation.

Critical strengths
  • Simple implementation, compatibility with multiple action formulations, and no inference overhead.
  • Improvements across several backbones, with preliminary hardware validation.
Critical weaknesses
  1. Very limited novelty and no substantive new insight. Applying the well-established SAM regularizer to VLA fine-tuning, with heuristic temporal weighting, offers little robotics-specific insight or demonstrated value at scale. Prior work [1] further limits the novelty.

  2. Limited theoretical contribution. The analysis combines standard sharpness arguments with conventional imitation-learning error propagation. The crucial curvature–input-sensitivity connection relies on Assumptions 4–5. Comparing conditional upper bounds establishes neither lower actual compounding error nor the superiority of temporal weighting over uniform SAM.

  3. Missing compute-matched comparisons and essential controls. SAM’s two backward passes nearly double training cost. The paper should compare against spending the same compute on additional standard updates or stronger augmentation, and include SAM-only and weighting-only ablations.

  4. Unclear relevance to large-scale VLA pretraining. SAM’s generalization benefits in limited-data fine-tuning may not translate to large-data pretraining. Watts et al. [2] report similar or worse pretraining loss than AdamW in controlled language-model experiments, despite improved resistance to subsequent forgetting. This suggests benefits in parameter stability rather than pretraining efficiency. Whether SAM benefits large-scale VLA pretraining therefore remains unestablished.

  5. Limited practical significance of simulation results. LIBERO is nearly saturated, with Being-H0.7 reporting 99.2%. Small gains provide limited evidence of substantive progress. Stronger baselines and richer evaluations such as results in RoboTwin are needed. LIBERO-Plus also remains far removed from real-world robotic deployment, making improvements on this benchmark of limited practical significance.

  6. Real-world robotics-specific robustness evaluations are needed. The authors should test physically realized changes that occur during deployment. For example, after an accidental displacement of the robot base changes the camera–robot geometry, does the policy trained on the original demonstrations still work without retraining? Such experiments would be more informative than adding synthetic noise to images, whose relevance to actual deployment disturbances is not established.

[1] Zhang and Bisk (2026), Flatness Preserves Instruction Following in Vision-Language-Action Models.
[2] Watts et al. (2026), Sharpness-Aware Pretraining Mitigates Catastrophic Forgetting.

Questions for authors
  1. What robotics-specific insight does FlatVLA provide beyond standard SAM, heuristic temporal weighting, and prior work [1]?

  2. See weakness 2

  3. See weakness 3

  4. Do the benefits extend to large-scale VLA pretraining, or primarily reflect limited-data regularization?

  5. Do gains persist on stronger baselines and richer benchmarks such as RoboTwin?

  6. Can you provide real-world experiments with robotics-specific robustness scenarios that reflect actual deployment disturbances?

6 comments
AuthorsOpenReview ↗

Response to Reviewer H4uu (Part 1/3)

We thank the reviewer for the feedback, and want to clarify and resolve the raised concerns. We have incorporated these additional analyses and theoretical clarifications into the revised manuscript. Lastly, we revised our manuscript with all changes highlighted in orange.

 

W1, Q1: Robotics-Specific Contribution and Novelty

We would like to clarify the robotics-specific contribution of FlatVLA compared with [1].

Robotic manipulation requires handling diverse tasks and complex physical interactions, and VLA models aim to address these challenges from a deep learning perspective. However, existing VLA models still face several limitations: (1) generalization issues under out-of-distribution (OOD) conditions and (2) compounding errors during long-horizon task execution. To address these challenges, our approach combines SAM and sequential weighting (SW), focusing on sequential error propagation in robotic action execution. Specifically, we hypothesize that SAM or SW alone may provide limited benefits, while combining them can yield better performance by encouraging local loss smoothness and emphasizing earlier actions to mitigate temporal error propagation. Unlike [1], which applies SAM to VLA models, FlatVLA leverages both loss smoothness and the sequential structure of robotic actions to improve given limitations.

To examine this, we provide additional ablation results comparing SAM only, SW only, and FlatVLA in the tables below. As shown, FlatVLA consistently achieves higher performance than either component alone.

LIBERO

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA75.8----78.5
OpenVLA-OFT96.896.996.897.4
π0\pi_093.393.593.594.3
π0.5\pi_{0.5}95.896.395.997.4

LIBERO-PLUS

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA16.5----19.5
OpenVLA-OFT67.468.368.170.1
π0\pi_045.745.746.047.3
π0.5\pi_{0.5}80.281.180.682.6

Additionally, we evaluate different replanning intervals EE. Since chunk-based VLA policies replan after every EE execution steps, discarding the remaining predicted actions, and it delays corrective feedback when EE is increased. Thus, to further examine this effect, we select π0\pi_0, which has the longest predicted action horizon (H=50H=50), and evaluate execution intervals E∈10,20,30,50E\in{10,20,30,50}. As shown in the table below, FlatVLA exhibits the smallest performance degradation compared with the other methods, further supporting the benefits of sequential weighting.

Replan (EE)BaselineSAM onlySW onlyFlatVLA
E=10E=1093.393.593.594.3
E=20E=2088.189.189.291.0
E=30E=3083.484.383.687.3
E=50E=5068.770.169.975.6

For brevity, we report only average success rates above. We have incorporated these additional ablations with all cases into the revised manuscript.

[1] Zhang and Bisk, Flatness Preserves Instruction Following in Vision-Language-Action Models Arxiv'26

AuthorsOpenReview ↗

Response to Reviewer H4uu (Part 2/3)

W2, Q2: Theoretical Motivation and Empirical Sharpness

We would like to emphasize that our theoretical contribution establishes a connection between loss-landscape sharpness, policy sensitivity, and sequential error propagation. Specifically,

  • Lemma 1 characterizes sequential action error propagation, while Proposition 1 connects local loss sharpness to policy sensitivity and error amplification through curvature analysis under Assumptions 4–5.
  • Theorem 1 establishes that a smoother loss landscape yields a no-larger compounding-error upper bound under shared local conditions and the same one-step imitation-error budget.

This analysis provides a principled understanding of the relationship between loss-landscape smoothness and sequential error propagation from a general deep learning perspective. To support this perspective, we conducted additional quantitative loss sharpness as Sρ(θ)=max⁡∥ϵ∥2≤ρ[L(θ+ϵ)−L(θ)]S_\rho(\theta)=\max_{\|\epsilon\|_2\leq\rho}\left[\mathcal L(\theta+\epsilon)-\mathcal L(\theta)\right].

BackboneBaselineSAM onlySW onlyFlatVLAReduction vs. Baseline ↓\downarrow
OpenVLA0.2143——0.070767.0%
OpenVLA-OFT0.01020.00810.01310.003862.7%
π0\pi_00.16310.05520.15410.018788.5%
π0.5\pi_{0.5}0.13250.03220.15390.010891.8%

As shown in the table above, FlatVLA consistently achieves the lowest measured sharpness within each backbone. With our qualitative loss-landscape analysis (Fig. 5), these results provide empirical support for reduced loss sharpness (Assumption 4). Additionally, our robustness improvements under multimodal input perturbations on LIBERO-PLUS provide further empirical support for our input-output sensitivity motivation (Assumption 5).

Although Eq. (9) is not directly derived from Eq. (4), both reflect the principle of temporal error propagation. Since the actual error amplification factors are unavailable during offline training, we use a fixed γ\gamma as a surrogate for propagation strength rather than an exact dynamics constant. The benefits of sequential weighting beyond SAM only (γ=0\gamma=0) are further supported by our ablations in W1/Q1.

 

W3, Q3: Computational Cost and Ablation Controls

As already discussed in Appendix D, FlatVLA requires approximately 2× the baseline GPU-hours due to SAM's additional gradient computation, without additional inference overhead. While extending training to match this computational budget provides a comparison of resource efficiency, it also changes the number of optimization steps and data exposures, making it less suitable for the contribution and the novelty of FlatVLA. Thus, we again emphasize the results of our ablation studies in W1/Q1 that FlatVLA consistently outperforms SAM only under the same computational budget, demonstrating the additional benefits of sequential weighting without further computational overhead.

 

W4, Q4: Scope of Large-Scale Pretraining

We would like to clarify that FlatVLA primarily targets robust fine-tuning of pretrained VLA models rather than large-scale pretraining. Our experiments demonstrate consistent improvements in downstream performance and robustness, supporting the benefits of sharpness-aware optimization in this setting. While [2] investigates SAM under a different training regime, our work focuses on improving policy robustness rather than pretraining loss or efficiency. In this work, we did not claim that these benefits necessarily extend to large-scale VLA pretraining, which remains an interesting direction for future work.

[2] Watts et al. Sharpness-Aware Pretraining Mitigates Catastrophic Forgetting. ICLR'26 Workshop GRaM

AuthorsOpenReview ↗

Response to Reviewer H4uu (Part 3/3)

W5, Q5: Benchmark Saturation and Robustness Evaluation

We acknowledge the concerns regarding saturated clean LIBERO performance and agree that additional benchmarks such as RoboTwin could strengthen our evaluation. Yet, we would like to emphasize the applicability of FlatVLA to real-world robotic manipulation beyond simulation benchmarks. To support this, we have uploaded real-world robot manipulation videos as supplementary material, where FlatVLA exhibits more stable action execution than the baseline. We believe these results provide additional practical evidence supporting our approach.

(Due to upload limitations, we selected two videos for each model under the clean task setting.)

Additionally, we would like to highlight the robustness results already reported in Table 14 of our manuscript. Our method shows improved robustness under Gaussian action noise across all evaluated noise levels (σ∈0.1,0.2,0.3\sigma \in {0.1, 0.2, 0.3}), demonstrating benefits beyond near-saturated clean benchmark performance.

Table 14 Recap

BackboneBaseline Spatial (σ=0.2\sigma=0.2)FlatVLABaseline Object (σ=0.3\sigma=0.3)FlatVLA
OpenVLA44.1052.2713.7815.60
OpenVLA-OFT63.8070.1025.7027.70
π0\pi_070.7072.8726.8034.00
π0.5\pi_{0.5}83.2086.9345.9347.13

These findings with our LIBERO-PLUS results and additional real-world evaluations (discussed further in W6/Q6) provide further evidence of FlatVLA's robustness beyond clean simulation benchmarks.

 

W6, Q6: Real-World Robotics-Specific Robustness

We evaluated OpenVLA-OFT and π0.5\pi_{0.5} without retraining under camera yaw rotations of ±10∘\pm10^\circ and robot-arm initial-position shifts of ±15 cm\pm15\,\mathrm{cm} along the Y-axis. The yaw rotations deliberately move part of the scene outside the camera's field of view. The position shifts place the arm closer to the cup (+15 cm+15\,\mathrm{cm}) or plate (−15 cm-15\,\mathrm{cm}). Each setting uses 10 trials per method.

DisturbanceBackboneBaselineFlatVLAImprovement
Camera yaw (+10∘+10^\circ, right)OpenVLA-OFT2/105/10+3
π0.5\pi_{0.5}3/107/10+4
Camera yaw (−10∘-10^\circ, left)OpenVLA-OFT1/105/10+4
π0.5\pi_{0.5}1/105/10+4
Initial shift (Y-axis, +15 cm+15\,\mathrm{cm})OpenVLA-OFT0/106/10+6
π0.5\pi_{0.5}1/106/10+5
Initial shift (Y-axis, −15 cm-15\,\mathrm{cm})OpenVLA-OFT2/106/10+4
π0.5\pi_{0.5}2/105/10+3

FlatVLA achieves 3–6 more successful trials out of 10 than the baseline across all settings.

We also want to highlight our existing action smoothness analysis (Table 13 of our manuscript), summarized again below, which shows that FlatVLA consistently achieves lower translational and rotational Action Total Variation (ATV). Both the lower ATV and the higher task success rates support improved real-world execution behavior.

Table 13 Recap

BackboneMethodTask successTranslation ATV ↓\downarrowRotation ATV ↓\downarrow
OpenVLA-OFTBaseline5/100.92 ±\pm 0.050.72 ±\pm 0.01
FlatVLA10/100.66 ±\pm 0.070.69 ±\pm 0.02
π0.5\pi_{0.5}Baseline5/102.85 ±\pm 0.071.25 ±\pm 0.02
FlatVLA8/102.64 ±\pm 0.071.21 ±\pm 0.02
Reviewer_ujtQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlatVLA, a fine-tuning method that combines sharpness-aware minimization (SAM) with normalized, horizon-dependent weighting of imitation losses. Earlier positions within an action chunk receive greater weight, and the same weighted objective is used in both SAM passes. The paper develops a conditional local analysis linking parameter-space sharpness to an upper bound on compounding distributional error through additional sensitivity assumptions. Experiments cover OpenVLA, OpenVLA-OFT, π0, and π0.5 on LIBERO and LIBERO-Plus, action-noise evaluations, and real-world manipulation. The reported results show positive average trends. However, for a single-action horizon the method reduces to ordinary SAM, and the incremental contribution of sequential weighting remains a central assessment question.

Critical strengths
  1. Evaluation across multiple architectures and robustness settings. The paper evaluates four VLA backbones and reports positive average improvements on both LIBERO and LIBERO-Plus. It also studies executed-action noise and includes real-world manipulation with two backbones. This breadth provides encouraging evidence that the complete training recipe can be useful, rather than relying on one architecture or only clean benchmark performance (Tables 1–2 and 8; Section 5.3).
  2. A simple training-time modification without architectural or inference changes. The method retains native per-position imitation losses and changes their aggregation and optimization. The unit-mean weighting rule and two-pass SAM procedure are explicitly described, making the approach relatively straightforward to integrate into existing fine-tuning pipelines. Connecting training geometry with sequential policy behavior is also a relevant motivation, even though the explanatory connection requires stronger support (Section 4; Algorithm 1).
Critical weaknesses
  1. The benefit of the new sequential weighting is not isolated, and the reliability of the smaller gains is unclear. The central comparisons are between the original training baseline and full FlatVLA. For the chunked models, there is no same-backbone comparison with ordinary SAM using uniform weights (γ = 0), or with sequential weighting without SAM. Therefore, the results do not establish whether the new weighting adds a stable benefit beyond SAM itself. The single-action OpenVLA result cannot substitute for this ablation because FlatVLA already equals ordinary SAM when H = 1. This distinction is consequential given that prior SAM-based VLA work is acknowledged in Section 2.2. The early-action emphasis is plausible but remains insufficiently validated: affecting more subsequent time steps does not by itself establish that earlier positions should always receive more training weight. The outcome depends on dynamics, task phase, execution length, and opportunities for correction. Position-wise sensitivity measurements or alternative weight-order controls are needed to support this specific design (Section 4, Eq. 9). Several clean LIBERO gains are small relative to the reported variation. For example, OpenVLA-OFT on SPATIAL improves from 97.9 ± 0.4 to 98.3 ± 0.5. Table 1 does not specify what the ± values represent or the training/evaluation seed structure. This does not establish that the improvement is statistically insignificant, but it prevents a clear assessment of whether the smaller differences exceed training and evaluation variability. Aggregate gains alone also do not validate the proposed weighting mechanism.

  2. The theoretical bridges and their correspondence to chunk execution are not sufficiently justified. Low sensitivity of the training loss to parameter perturbations does not automatically imply low sensitivity of executed actions to visual-input perturbations. Assumption 4 connects loss curvature to parameter-output sensitivity with a residual, while Assumption 5 introduces an input–parameter sensitivity decomposition with another residual. The paper does not provide sufficient conditions or measurements showing that these residuals are non-dominant and the common comparison constants are informative for the evaluated policies. The conditional theorem can be valid without establishing that these assumptions explain the observed improvements (Appendix A.3, Eqs. 28–35; Appendix A.5). The mapping to the actual losses also needs clarification: OpenVLA-OFT uses L1 regression, OpenVLA uses token prediction, and the flow models optimize velocity-matching objectives rather than directly optimizing the final sampled action map. The paper does not establish the required correspondence between these training formulations and the differentiable policy sensitivity used in the analysis (Appendix B.2). Finally, the episode-level recurrence describes feedback through newly observed inputs, whereas actions within a parallel or flow-generated chunk can be produced jointly before execution. An early action error can change subsequent physical states, but it does not automatically trigger recomputation of later actions in the same open-loop chunk. The closed-loop recurrence therefore does not directly justify the within-chunk weighting rule. The executed horizon, replanning schedule, and position-dependent propagation should be made explicit before interpreting the weighting as a consequence of the analysis (Eqs. 4 and 9).

Questions for authors
  1. Isolating the new contribution. On a representative chunked backbone, how do the original optimizer, ordinary SAM with uniform weights, weighting without SAM, and full FlatVLA compare under the same data and hyperparameter-selection protocol? Is there a stable incremental benefit from the proposed weighting over ordinary SAM across independent runs?
  2. Testing the early-position premise. What evidence shows that earlier chunk positions have greater downstream impact under the actual execution schedule? Please specify generated and executed horizons and the replanning frequency. Position-wise perturbation measurements or reversed/permuted weight controls would help distinguish the claimed temporal mechanism from generic loss reweighting.
  3. Connecting the assumptions to the implemented policies. What sufficient conditions or empirical measurements support Assumptions 4–5 for the actual token, L1, and flow-matching formulations? How are the residual terms controlled across the compared policies, and which part of the closed-loop argument remains valid during open-loop chunk execution? Please also clarify the relation between sharpness of the analyzed imitation objective and that of the weighted objective used by FlatVLA.
  4. Quantifying uncertainty in the reported gains. What do the ± values in Table 1 denote, and how many independent training runs and evaluation trials were used? Please report uncertainty for the performance differences, with matched evaluation conditions where appropriate, so that the smaller gains can be assessed against training and rollout variability rather than by overlap of error bars alone.
3 comments
AuthorsOpenReview ↗

Response to Reviewer ujtQ (Part 1/2)

We thank the reviewer for the valuable feedback. We want to clarify and resolve the raised concerns regarding the contribution of this paper, the evaluation protocol, and the theoretical analysis. Below we provide detailed responses to each question. Lastly, we revised our manuscript with all changes highlighted in orange.

 

W1 part 1, Q1, Q4: Contribution of Sequential Weighting

To isolate the contribution of sequential weighting beyond SAM, we conducted additional ablations on LIBERO and LIBERO-PLUS, comparing four configurations: (1) the original baseline, (2) SAM only (γ=0\gamma=0), (3) sequential weighting without SAM (SW only), and (4) FlatVLA.

As shown below, FlatVLA shows improvements across all models and apporaches, and these results demonstrate the complementary benefits of combining SAM with sequential weighting. For OpenVLA (H=1H=1), FlatVLA is equivalent to SAM only, and SW only equivalent to the baseline.

In last, regarding evaluation variability (Q4), the reported ±\pm represent standard deviations across three independently trained models.

LIBERO

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA75.8----78.5
OpenVLA-OFT96.896.996.897.4
π0\pi_093.393.593.594.3
π0.5\pi_{0.5}95.896.395.997.4

LIBERO-PLUS

BackboneBaselineSAM onlySW onlyFlatVLA
OpenVLA16.5----19.5
OpenVLA-OFT67.468.368.170.1
π0\pi_045.745.746.047.3
π0.5\pi_{0.5}80.281.180.682.6

For brevity, we provide the full results on LIBERO and LIBERO-PLUS in the revised manuscript, along with all corresponding revisions.

 

W1 part 2, Q2: Extended Execution

We would like to clarify the general evaluation protocol for chunk-based VLA policies and provide a more detailed analysis of the benefits of sequential weighting under extended open-loop execution. In chunk-based VLA policies, the first EE actions from a predicted chunk of length HH are executed, after which the policy predicts a new action chunk based on the updated observation (E≤HE \leq H). This indicates that if EE increases, delayed corrective feedback allows early action errors to affect more subsequent physical states.

To examine this effect, we evaluate π0\pi_0 (H=50H=50) with E∈E\in{10,20,30,5010,20,30,50}, as its longer action horizon allows a wider range of open-loop execution intervals than OpenVLA-OFT (H=8H=8) and π0.5\pi_{0.5} (H=10H=10). As shown below, FlatVLA achieves the highest success rates across all four LIBERO suites at every interval. These results support the benefit of sequential weighting under extended open-loop execution.

Replan (EE)BaselineSAM onlySW onlyFlatVLA
E=10E=1093.393.593.594.3
E=20E=2088.189.189.291.0
E=30E=3083.484.383.687.3
E=50E=5068.770.169.975.6

As shown below, FlatVLA achieves the highest average success rates across all evaluated replanning intervals, with larger improvements over SAM only as EE increases. These results support the benefits of sequential weighting under extended open-loop execution.

AuthorsOpenReview ↗

Response to Reviewer ujtQ (Part 2/2)

W2 Part 1, Q3: Local Sensitivity

We acknowledge that our theoretical assumptions are not directly verified across different VLA training objectives. However, Assumptions 4–5 establish local curvature and sensitivity conditions linking loss sharpness to policy sensitivity in Proposition 1, with residual terms included in the assumed bounds rather than explicitly controlled during training. Prior work [1] further supports the connection between local smoothness and input robustness.

To support our theoretical motivation, we conducted additional analysis of parameter-space loss sharpness. We evaluated local loss sharpness on the full LIBERO training datasets, defined as Sρ(θ)=max⁡∣ϵ∣2≤ρ[L(θ+ϵ)−L(θ)]S_\rho(\theta)=\max_{|\epsilon|_2\leq\rho}[\mathcal L(\theta+\epsilon)-\mathcal L(\theta)]. As shown below, FlatVLA achieves the lowest measured sharpness across all model.

BackboneBaselineSAM onlySW onlyFlatVLAReduction vs. Baseline
OpenVLA0.2143----0.070767.0%
OpenVLA-OFT0.01020.00810.01310.003862.7%
π0\pi_00.16310.05520.15410.018788.5%
π0.5\pi_{0.5}0.13250.03220.15390.010891.8%

Together, the parameter-space sharpness analysis (Assumption 4) and LIBERO-PLUS robustness results under diverse multimodal input variations (Assumption 5) provide complementary empirical support for our theoretical motivation.

Regarding error propagation, Eq.(4) describes closed-loop error propagation through updated action predictions, whereas Eq.(9) addresses open-loop execution without replanning. Eq.(9) is motivated by Eq.(4), as further supported by our replanning experiments in W1/Q2.

[1] Zhang et al., On the Duality Between Sharpness-Aware Minimization and Adversarial Training, ICML'24

 

W2 Part 2: Empirical Analysis of Error Propagation

To see the effects of temporal error propagation and position-dependent weighting, we conducted two action-noise analyses at the episode and action-chunk levels on LIBERO-10 using π0.5\pi_{0.5} (σ=0.2\sigma=0.2).

  • Temporal Action Noise Injection: We injected action noise during the first 20 execution steps (1–20) or the subsequent 20 steps (21–40). As shown below, early disturbances cause greater performance degradation than later disturbances, with a -32.3 pp drop under early noise compared with only -4.8 pp under later noise. Notably, FlatVLA shows the smallest performance drops under both perturbation settings. These results support our motivation regarding the greater impact of early action disturbance on subsequent task performance and the improved robustness of FlatVLA.

    Here, () indicates performance drop pp relative to No Noise.

    Noise Injection (π0.5\pi_{0.5})BaselineSAM onlySW onlyFlatVLA
    No Noise92.292.992.594.3
    First 20 steps (1–20)59.9 (−32.3 pp)63.3 (−29.6 pp)60.5 (−32.0 pp)67.5 (−26.8 pp)
    Next 20 steps (21–40)87.4 (−4.8 pp)88.8 (−4.1 pp)87.1 (-5.4pp)90.7 (−3.6 pp)
  • Action Chunk Noise Injection: To further examine position-dependent error propagation within action chunks, we injected action noise into either the first 5 or last 5 actions of each chunk (E=10E=10). As shown below, the baseline shows that disturbances at earlier positions within a given action chunk lead to greater performance degradation. And then, FlatVLA achieves higher success rates across all settings, further supporting our motivation for emphasizing earlier actions in Eq.(9).

    Noise Position (π0.5\pi_{0.5})BaselineSAM onlySW onlyFlatVLA
    First 5 actions (1–5)75.478.176.080.4
    Last 5 actions (6–10)78.580.679.881.4

With our replanning experiments in W1 Part 2/Q2, these analyses provide complementary empirical support for the temporal and position-dependent motivation behind FlatVLA. We provided complete results accross the benchmark suite on the revised manuscript.

AuthorsOpenReview ↗

Brief Reminder for Reviewer ujtQ

Dear Reviewer ujtQ,

We appreciate your time and constructive feedback. We hope our rebuttal and revised manuscript (e.g., the additional experiments and sharpness analyses) have sufficiently addressed the points and concerns raised. For quick reference, we have included an overview of where each revision appears in our top-level general response.

Lastly, with the discussion period nearing its end, we would be glad to clarify any remaining questions or concerns.

Best regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Responses and Manuscript Revisions for All Reviewers

Dear All Reviewers,

We sincerely thank the reviewers for their thoughtful and constructive feedback. Below, we summarize the key revisions and clarifications included in the updated manuscript, with all changes highlighted in orange\textbf{\color{orange}{orange}}.

For Reviewer ujtQ

  • Added SAM only and SW only results on LIBERO and LIBERO-PLUS in Sec.5.2 and 6.1 (Table 1–2). (W1 Part 1, Q1)
  • Clarified that the ±\pm are standard deviations computed across three independently trained models in the Table 1 caption. (Q4)
  • Clarified how many actions are executed before replanning and added extended open-loop execution experiments in Sec.6.3 (Table 3), with suite-level results in Appendix C.1 (Table 9). (W1 Part 2, Q2)
  • Explained the local conditions connecting loss sharpness to policy sensitivity and the motivation for sequential weighting in our response, referring to Sec.3.2 and Sec.4. Added parameter-space loss-sharpness measurements on the full LIBERO training datasets in Appendix C.5 (Table 15). (W2 Part 1, Q3)
  • Added action-noise experiments at early versus later execution steps and at early versus later positions within each action chunk in Sec.6.3 (Table 4), with suite-level results in Appendix C.2 (Table 10–11). (W2 Part 2)

For Reviewer H4uu

  • Added SAM only and SW only comparisons on LIBERO and LIBERO-PLUS in Sec.5.2 and 6.1 (Table 1–2), and extended open-loop execution experiments in Sec.6.3 (Table 3) and Appendix C.1 (Table 9). (W1, Q1)
  • Explained the connection between loss sharpness, policy sensitivity, and sequential error propagation in our response, referring to Sec.3–4. Added parameter-space loss-sharpness measurements in Appendix C.5 (Table 15). (W2, Q2)
  • Expanded Appendix D to explain the equal-compute comparison between SAM only and FlatVLA and the remaining limitation of a resource-matched baseline comparison. (W3, Q3)
  • Clarified in our response that the experiments evaluate fine-tuning of pretrained VLA models, and benefits for large-scale pretraining remain untested. (W4, Q4)
  • Provided supplementary real-world manipulation videos and highlighted the existing Gaussian action-noise results in Sec.6.2 and Appendix C.4 (Table 14). (W5, Q5)
  • Added real-world robustness evaluations under camera yaw rotations and robot-arm initial-position shifts in Appendix C.3 (Table 12). The yaw rotations deliberately move part of the scene outside the camera's field of view. (W6, Q6)

For Reviewer DH2q

  • Added SAM only and SW only results on LIBERO and LIBERO-PLUS to assess their combined benefits in Sec.5.2 and 6.1 (Table 1–2). (W1, Q1)
  • Clarified in our response how loss sharpness relates to policy sensitivity and compounding-error upper bounds in Sec.3.2–3.3, and how this motivates sequential weighting in Sec.4. Expanded the explanation of fixed γ\gamma settings in Appendix B.3. (W2, Q2 Part 1)
  • Added experiments varying the replanning interval to evaluate robustness under extended open-loop execution in Sec.6.3 (Table 3), with suite-level results in Appendix C.1 (Table 9). (W2, Q2 Part 1)
  • Added parameter-space loss-sharpness measurements in Appendix C.5 (Table 15) and experiments comparing early and later action noise in Sec.6.3 (Table 4) and Appendix C.2 (Table 10–11). (W3, Q2 Part 2, Q3)
  • Revised the predictive-uncertainty analysis to describe the relative gap between successful and failed executions in Sec.6.4 (Table 5). (W3, Q2–3)

Dear Reviewer ujtQ,

As the rebuttal discussion period is drawing to a close, we kindly encourage you to review the authors' responses to your previous concerns and provide any further feedback or clarification as appropriate.

Thank you for your time and cooperation.

Best regards,

Your Area Chairs

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.