acceptodds
Under review as a conference paper at ICLR 2027

HRAE: One-Step Action Generation for Flow-Matching VLAs via Hierarchical Recurrent Action Expert

Abstract

Diffusion and flow-based vision-language-action (VLA) policies achieve strong robotic control through repeated model evaluations to progressively refine actions, but such multi-step inference introduce substantial latency. Existing one-step approaches reduce this cost by compressing the external generative process, often through distillation, consistency objectives, or specialized training targets, which can require pretrained teachers or additional trajectory-level training machinery. However, reducing external sampling steps also removes the iterative computation performed during generation. Recent evidence suggests that such iterative computation, when appropriately supervised, is an important inductive bias for generative robotic control. This raises a key question: can we retain iterative computation while avoiding repeated external flow evaluations? We address this by moving iterative computation inside each action-expert evaluation. We introduce the **Hierarchical Recurrent Action Expert (HRAE)**, which couples a fast low-level latent state for recurrent action refinement with a persistent high-level latent state that integrates information across refinements and guides subsequent computation. By reusing the same low-level and high-level modules across hierarchical recurrent cycles, HRAE achieves substantial effective computational depth without proportionally increasing model size. We further pair HRAE with clean-action -prediction, which aligns the prediction target with its internally refined action representation while remaining compatible with multi-step flow integration. Across tasks in LIBERO, LIBERO Plus, RoboTwin, and real-world manipulation, HRAE achieves strong in-distribution performance and robustness under distribution shifts with only one external flow evaluations. In real-world tasks, one-step HRAE outperforms a 10-step GR00T baseline while reducing action-expert latency from 44.73 ms to 6.85 ms. These results suggest that iterative computation need not be tied to repeated external flow integration and can instead be effectively realized within the action expert for efficient generative robotic control.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision Language Action, Hierarchical Recurrent Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_G2TdOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Hierarchical Recurrent Action Expert (HRAE) for efficient action generation in flow-matching VLA models. Instead of running the flow model many times, HRAE moves the iterative computation inside the action expert. A low-level recurrent module with shared parameters repeatedly refines the action representation, and a high-level recurrent module combines information across these refinement cycles. The method is evaluated on LIBERO, LIBERO-Plus, RoboTwin Clean-to-Random and two real-world tasks. In real-world experiments, it matches or exceeds multi-step baselines while cutting the action expert’s latency from 44.73 ms (GR00T) to 6.85 ms.

Critical strengths
  1. I think the motivation is quite good. It addresses a very practical limitation that diffusion- and flow-based VLA policies have, which is the computational cost of repeated action-expert evaluation. The method is also reasonable and straightforward. Rather than relying on teacher distillation or additional trajectory-level objectives, HRAE explores hierarchical recurrence as an alternative way to provide iterative computation.
  2. The results are promising across several evaluation settings, especially the action-expert latency reduction.
Critical weaknesses
  1. I think the contribution of hierarchical recurrence is not fully demonstrated. Although Table 4 includes a DiT x-prediction baseline, the inference settings are not fully specified. And there are no comparisons with baselines matched for model capacity or inference latency (default HRAE is 32 transformer block evaluations). Removing high-level recurrence may also reduce the computational cost. So it is hard to say whether the improvements come from the hierarchical design or simply from additional computation.
  2. The method novelty is a bit limited. The hierarchical recurrent architecture is largely inspired by existing hierarchical reasoning models, while x-prediction and truncated gradient computation are also established techniques. The main contribution appears to be adapting and combining these ideas for VLA action generation.
  3. I feel confused about the benefit of the x-prediction. The comparison between x-prediction and v-prediction changes not only the prediction target but also the loss weighting. So it is unclear whether the improvement comes from the prediction target itself or the different loss weighting.
Questions for authors
  1. Is hierarchical recurrence beneficial under matched computational budgets? Can the authors compare HRAE with a single-level recurrent expert and a direct-regression baseline, with matched parameter counts or inference latency?
  2. Can the authors compare clean-action MSE with equivalently time-weighted velocity MSE under the same training settings? Also, could the authors clarify the flow-time sampling distribution and the number of inference steps used for each model in Table 4?
  3. In the real-world experiments, each method is evaluated on only 10 trials per task, which seems insufficient to make good conclusions about the performance differences. Can the authors conduct more trials or report uncertainty estimates?
6 comments
AuthorsOpenReview ↗

Response to Reviewer G2Td Part 1/2

Thank you for your in-depth review and your recognition of the promising results of our work. We address the questions you raised as follows:

Q1: Comparisons with baselines matched for model capacity and computational budgets.

As shown in Table T5, we have added comparisons with baselines with approximately matched FLOPs per flow evaluation, including a DiT x-prediction baseline (Deeper StarVLA-GR00T) and a single-level recurrent expert (HRAE-Lonly-Deeper).

  1. With one flow step, HRAE (59.4) still outperforms HRAE-Lonly-Deeper (53.3) and Deeper StarVLA-GR00T-1s (52.3). With 10 flow steps, Deeper StarVLA-GR00T (54.1) outperforms StarVLA-GR00T (52.5) but still performs worse than one-step HRAE (59.4), despite requiring approximately 9.7 times the total inference computation (470.22 vs. 48.51 GFLOPs).

  2. For model capacity, HRAE-Lonly-Deeper matches HRAE’s parameter count (108.47M), but Deeper StarVLA-GR00T has substantially more parameters (202.83M).

  3. Increasing the parameter count and computational budget of the single-level recurrent expert in HRAE-Lonly-Deeper does not close the performance gap with HRAE, supporting the benefit of incorporating the high-level recurrent module.

These results support the effectiveness of the hierarchical design. We have updated Table 7 in the revised manuscript to include additional experiments and more detailed inference settings.

Table T5: More ablation studies on model capacity and computational budgets.

ModelBlock NumPredictionFlow StepGFLOPsParamsRobotwin
StarVLA-GR00T16x-pred10378.39164.2652.5
Deeper StarVLA-GR00T20x-pred10470.22202.8354.1
Deeper StarVLA-GR00T20x-pred147.02202.8352.3
HRAE-Lonly-Deeper28x-pred147.41108.4753.3
HRAE32x-pred148.52108.4759.4

Q2: Method Novelty.

The motivation of our design differs from existing hierarchical reasoning methods, such as HRM and HRM-Text. HRM/HRM-Text targets text reasoning, whereas HRAE targets robotic reasoning, which requires integrating multimodal information while iteratively refining actions. We redesign the H-module and L-module to enable efficient multimodal processing and make them more lightweight for faster action refinement. Additionally, we integrate the hierarchical recurrent expert with x-prediction for higher-quality action generation.

To evaluate the benefits of these adaptations, we compare HRAE with an HRM-style action expert in Table T2. With the same number of blocks and comparable parameter counts, HRAE reduces computation from 236.74 to 48.52 GFLOPs and improves the RoboTwin success rate from 56.5 to 59.4.

  1. In the HRM-style structure, both the H-module and L-module apply self-attention over concatenated image-text representations, state tokens, and action tokens (7.386 × 32 = 236.352 GFLOPs). This design results high computational cost and inference latency, while also harder to converge, resulting in lower action prediction accuracy.

  2. Our H-module uses state-action self-attention (0.973 × 8 = 7.784 GFLOPs), while our L-module combines cross-attention to text/image representations with state-action self-attention (1.669 × 12 + 1.692 × 12 = 40.332 GFLOPs). These concise and efficient module designs enable deep iterative refinement of action predictions while maintaining low inference latency, supporting fast and accurate action generation.

  3. We find that aligning recurrent latent refinement with the training objective (x-prediction) achieves higher-quality predictions.

Therefore, combining x-prediction with HRAE enables fast and accurate one-step action generation, whereas conventional GR00T requires multiple flow-matching denoising steps.

Table T2: Comparision of HRM-style action expert and HRAE.

ModelBlocksParams (M)GFLOPsRoboTwin
HRM-style32100.33236.7456.5
HRAE32108.4748.5259.4
AuthorsOpenReview ↗

Response to Reviewer G2Td Part 2/2

Q3: Clarification of the loss setting and flow-time sampling distribution.

  1. Loss setting.

We clarify that x-prediction is trained with unweighted L1 loss, whereas v-prediction is trained with unweighted MSE loss, without additional time-dependent loss weighting.

As shown in Table T6, replacing MSE loss with L1 loss improves the x-prediction success rate from 58.3 to 59.4 for HRAE and from 52.2 to 52.5 for GR00T (DiT). Accordingly, all our x-prediction experiments use L1 loss unless otherwise specified.

HRAE retains a substantial advantage when x-prediction is trained with MSE. Under the MSE setting, one-step HRAE with x-prediction achieves 58.3, outperforming ten-step GR00T (DiT) with v-prediction (52.4) and one-step HRAE with v-prediction (52.2).

Table T6: Comparison of prediction and loss functions on RoboTwin.

ModelFlow StepsPredictionLossRoboTwin
GR00T (DiT)10v-predMSE52.4
GR00T (DiT)10x-predMSE52.2
GR00T (DiT)10x-predL152.5
HRAE1x-predMSE58.3
HRAE1x-predL159.4
  1. Flow-time sampling distribution.

Both x-prediction and v-prediction use the same flow-time sampling distribution, Gaussian noise distribution, and interpolation process. Following StarVLA, we sample u∼Beta(1.5,1.0)u \sim \mathrm{Beta}(1.5, 1.0) and set t=ρ(1−min⁡(u,ρ))t = \rho\bigl(1-\min(u,\rho)\bigr), where ρ=0.999\rho = 0.999.

We have clarified these details in the revised manuscript.

Q4: Real-World Experiments.

As shown in Table T1, we expanded our real-world experiments to four tasks, with 20 trials per task. Our method demonstrates better performance than both baselines on all four tasks. We have revised the manuscript to include descriptions of two additional tasks in Figure 3 and updated evaluation results for all four tasks in Table 6.

Table T1: Real-World Experiments.

ModelFlow StepsPick up and place cubeStack blocksPlace cube left of cylinderPick up and place cylinder
GR00T-N1.71011/209/2013/2010/20
π₀108/208/2013/2012/20
HRAE117/2010/2016/2017/20
AuthorsOpenReview ↗

Response to Reviewer G2Td

Dear Reviewer G2Td,

Thank you for your time and thoughtful comments. We greatly appreciate your recognition of the practical relevance of our motivation and the promising results of our model.

As the discussion period will close on November 18 (AoE), we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points require further clarification—we would be happy to provide it.

Thank you again for your valuable feedback.

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

This paper introduces HRAE single-step action generation in flow-based VLA models. By integrating iterative refinement through low- and high-level recurrent modules and employing clean-action prediction, HRAE delivers competitive manipulation performance while reducing inference latency relative to multi-step VLA baselines.

Critical strengths

Empirical performance and efficiency. HRAE demonstrates strong performance across simulation, OOD generalization, and real-world manipulation tasks, while substantially reducing action-expert inference latency compared with multi-step baselines.

Critical weaknesses
  1. Lack of one- or few-step baselines: The experiments primarily compare HRAE against standard multi-step VLA models. Comparisons with existing one- or few-step methods [1, 2] would help better establish the effectiveness of the proposed approach.

  2. Lack of dexterous manipulation evaluation: The experiments do not include dexterous manipulation tasks requiring precise, high-frequency control. Given that reducing inference latency is a primary motivation, evaluating such tasks would help demonstrate the practical benefits of HRAE.

[1] Zhan, Guojian, et al. "Mean Flow Policy with Instantaneous Velocity Constraint for One-Step Action Generation." ICLR (2026).

[2] Luan, Wuyang, et al. "SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation." arXiv preprint arXiv:2604.05656 (2026).

Questions for authors
  1. Could the authors provide comparisons with existing one- or few-step action generation baselines?

  2. Could the authors evaluate HRAE on dexterous manipulation tasks to demonstrate the benefits of reduced inference latency?

4 comments
AuthorsOpenReview ↗

Response to Reviewer WekN

Thank you for your constructive review and comments on our paper. We address your concerns as follows and will incorporate in the revision.

Q1: Comparison with One-Step Methods.

Existing one-step approaches primarily train models to compress the flow-matching process into a single step through constraints or teacher models, whereas HRAE moves iterative computation inside each flow evaluation. HRAE is complementary to existing approaches. We compare HRAE with MVP [1], SnapFlow [2], and Let It Be Simple [3] on LIBERO in Table T4. We use the results reported in the SnapFlow and Let It Be Simple papers. For MVP, we adapt its official implementation to LIBERO and train it for 1000k offline reinforcement learning steps, following its main training settings.

  1. Among the methods without pretraining, HRAE-scratch (98.0) outperforms MVP (71.7) and Let It Be Simple (97.75).

  2. We also compare HRAE-pretrain with SnapFlow models distilled from π₀ and π₀.₅, denoted as SnapFlow-π₀ and SnapFlow-π₀.₅, respectively.

    ○ HRAE-pretrain achieves a success rate of 98.7, comparable to the 98.75 reported for SnapFlow-π₀.₅, but SnapFlow’s performance depends on the pretrained teacher model used for distillation.

    ○ Notably, HRAE-pretrain uses 2.88k hours of robot pretraining data, whereas π₀.₅ is reported to have been pretrained on 10k+ hours of robot data.

These results demonstrate the effectiveness of HRAE compared with existing one-step methods.

Table T4: Comparison of HRAE with other one-step methods on LIBERO. All methods use a single action-generation step at inference time.

MethodLiberoMethodLiberoPretrain Data (h)
MVP71.7SnapFlow-π₀81.010k+
Let It Be Simple97.75SnapFlow-π₀.₅98.7510k+
HRAE-scratch98HRAE-pretrain98.72.88k

[3] Yitong Chen, et al. Let it be simple: One-step action generation for vision-language-action models. arXiv preprint arXiv:2606.05737, 2026.

Q2: Dexterous Manipulation Evaluation.

We evaluate HRAE and StarVLA-GR00T on 10 dexterous manipulation tasks from the RoboCasa GR1 Tabletop benchmark. As shown in Table T5, HRAE-Scratch (43.6) outperforms StarVLA-GR00T (40.0) while reducing policy latency(108.87ms -> 60.72ms). We have updated Appendix A.4 of the manuscript with detailed information on the 10 dexterous manipulation tasks.

Table T5: Comparison of HRAE with StarVLA-GR00T on 10 dexterous manipulation tasks of RoboCasa GR1 Tabletop.

ModelFlow StepsPolicy Latency (ms)10 Robocasa tasks
StarVLA-GR00T1098.8040.0
HRAE-Scratch160.7243.6
AuthorsOpenReview ↗

Response to Reviewer WekN

Dear Reviewer WekN,

Thank you for your time and thoughtful comments. We greatly appreciate your recognition of our model’s strong performance and its substantially reduced action-expert inference latency.

As the discussion period will close on November 18 (AoE), we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points require further clarification—we would be happy to provide it.

Thank you again for your valuable feedback.

I thank the authors for their rebuttal, and it has addressed many of my concerns. I have raised my score as a result.

AuthorsOpenReview ↗

We appreciate your recognition and support.

Thank you for taking the time to read our paper and the responses. We are grateful for the improved score. Please feel free to ask if you have any further concerns.

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

This paper proposes the Hierarchical Recurrent Action Expert (HRAE) for flow-based vision-language-action models. HRAE combines low- and high-level latent-state updates, parameter sharing across recurrent cycles, and clean-action x-prediction to enable action generation with a single external flow evaluation. Experiments on LIBERO, LIBERO Plus, RoboTwin, and real-world manipulation report competitive task performance, improved robustness in several distribution-shift settings, and lower action-expert inference latency.

Critical strengths
  1. Experiments show that HRAE substantially reduces action-expert inference latency while achieving comparable or higher success rates across multiple tasks and demonstrating improved robustness under distribution shifts, such as those in LIBERO Plus.

  2. The low- and high-level modules share parameters across recurrent cycles, allowing the model to increase its effective computational depth without increasing the parameter count in proportion to the number of cycles.

Critical weaknesses
  1. The real-world experiments cover only two tasks, with each model evaluated for only 10 trials per task, leaving insufficient evidence to determine whether the observed success-rate advantages are consistent across repeated evaluations.

  2. HRAE's core hierarchical recurrent structure is highly similar to that of HRM/HRM-Text, with the new design elements primarily adapting this structure to multimodally conditioned continuous action generation. The paper does not sufficiently demonstrate the necessity and benefits of these adaptations beyond directly transferring the existing architecture.

Questions for authors
  1. What do the authors consider HRAE's main methodological innovation beyond directly adapting HRM/HRM-Text to VLA models?

  2. Both external flow evaluations and internal recurrent computation iteratively update states using shared network parameters. How do the authors explain the simultaneous improvements in inference speed and task performance observed for HRAE in the experiments?

  3. The supplementary material I examined appears to contain documentation and code for HiTag rather than HRAE. Could the authors confirm whether the correct supplementary material was uploaded and report the parameter count of the action expert used in the experiments?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NtEE Part 1/2

Thank you for your positive review and thoughtful questions, which we address as follows and clarify them in the revision.

Q1 : Real-World Experiments.

As shown in Table T1, we expanded our real-world experiments to four tasks, with 20 trials per task. Our method demonstrates consistent success-rate advantages across repeated evaluations. We have revised the manuscript to include descriptions of two additional tasks in Figure 3 and updated evaluation results for all four tasks in Table 6.

Table T1: Real-World Experiments.

ModelFlow StepsPick up and place cubeStack blocksPlace cube left of cylinderPick up and place cylinder
GR00T-N1.71011/209/2013/2010/20
π₀108/208/2013/2012/20
HRAE117/2010/2016/2017/20

Q2: Innovations compared to HRM/HRM-Text.

The motivation of our design differs from HRM/HRM-Text. HRM and HRM-Text target text reasoning, whereas HRAE targets robotic action reasoning, which requires integrating multimodal information while iteratively refining actions. We redesign the H-module and L-module to enable efficient multimodal processing and make them more lightweight for faster action refinement.

To evaluate the benefits of these adaptations, we compare HRAE with an HRM-style action expert in Table T2. With the same number of blocks and comparable parameter counts, HRAE reduces computation from 236.74 to 48.52 GFLOPs and improves the RoboTwin success rate from 56.5 to 59.4.

  1. In the HRM-style structure, both the H-module and L-module apply self-attention over concatenated image-text representations, state tokens, and action tokens (7.386 × 32 = 236.352 GFLOPs). This design results high computational cost and inference latency, while also harder to converge, resulting in lower action prediction accuracy.

  2. Our H-module uses state-action self-attention (0.973 × 8 = 7.784 GFLOPs), while our L-module combines cross-attention to text/image representations with state-action self-attention (1.669 × 12 + 1.692 × 12 = 40.332 GFLOPs). These concise and efficient module designs enable deep iterative refinement of action predictions while maintaining low inference latency, supporting fast and accurate action generation.

We find that aligning recurrent latent refinement with the training objective (x-prediction) achieves higher-quality predictions. Therefore, combining x-prediction with HRAE enables fast and accurate one-step action generation, whereas conventional GR00T requires multiple flow-matching denoising steps.

Table T2: Comparision of HRM-style action expert and HRAE.

ModelBlocksParams (M)GFLOPsRoboTwin
HRM-style32100.33236.7456.5
HRAE32108.4748.5259.4

Q3: Improvements in speed and performance.

The improvements in inference speed and task performance come from fast L-module refinement and efficient H-module integration. Specifically, the lightweight module design enables faster inference, while hierarchical iteration improves task performance. Besides, HRAE requires only one step (6.85 ms), whereas our GR00T baseline uses 10 steps (4.473 X 10 = 44.73 ms). As shown in Table T3, we compare HRAE and StarVLA-GR00T in terms of RoboTwin tasks, FLOPs per block, block count, FLOPs per flow step, ect.

  1. HRAE has more computationally efficient blocks, which enable deeper internal iterative refinement (32 blocks vs 16 blocks) within each flow-matching step.

  2. In HRAE’s L-module, cross-attention alternates between image and text features. We reduce the dimensionality of image-text features in advance and reuse the features for low-dimensional K/V projections (0.566 GFLOPs).

  3. In contrast, StarVLA-GR00T performs high-dimensional K/V projections (1.887 GFLOPs) on the image-text features for every cross-attention block, and additionally processes learnable future tokens.

Therefore, our efficient block designs enable deeper internal refinement at a lower computational cost, supporting accurate action generation with a single external flow evaluation and improving both inference speed and task performance.

Table T3: Comparison of StarVLA-GR00T and HRAE in terms of block design, inference speed, and task performance.

ModelParamsBlock GFLOPsBlock NumGFLOPs per StepStepsGFLOPs per InferenceRoboTwin
StarVLA-GR00T164.26Cross: 3.143; Self: 1.448Cross: 8; Self: 837.8410378.452.4
StarVLA-GR00T164.26Cross: 3.143; Self: 1.448Cross: 8; Self: 837.84137.8445.8
HRAE108.47L-Text: 1.669; L-Img: 1.692; H: 0.973L-Text: 12; L-Img: 12; H: 848.52148.5259.4
AuthorsOpenReview ↗

Response to Reviewer NtEE Part 2/2

Q4: Supplementary Material and Parameter Count.

We have updated the supplementary material for HRAE and reported the parameter count of the action expert used in our experiments in Table T3. We uploaded the core code and representative examples of failed task executions in the supplementary material. In addition, we have updated Table 7 in the revised manuscript to report the parameter count of the action expert.

AuthorsOpenReview ↗

Response to Reviewer NtEE

Dear Reviewer NtEE,

Thank you for your time and thoughtful comments. We greatly appreciate your recognition of our parameter-efficient recurrent design and competitive task performance.

As the discussion period will close on November 18 (AoE), we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points require further clarification—we would be happy to provide it.

Thank you again for your valuable feedback.

AuthorsOpenReview ↗

Request for Feedback on Our Responses

Dear Reviewer NtEE,

Thank you for your time and thoughtful comments.

Since the discussion is about to close in one day, we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points need further clarification—we are happy to help. If our responses have resolved your concerns, we would be grateful if you would kindly consider raising the score.

Thank you again for your valuable feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

Dear Reviewers,

We deeply appreciate the time and effort you have dedicated to providing detailed and insightful feedback on our work. This work has received a lot of reviews with detailed comments. Your comments have been valuable in improving the quality of our paper.

In response, we have carefully addressed all your concerns individually and made revisions to both our main paper and the supplementary material. The primary revisions include, but are not limited to, the following:

● We have updated Table 6 to report the results of our expanded real-world experiments, which now cover four tasks with 20 trials per task. Our method consistently achieves higher success rates across all four tasks.

● We have updated Table 7 to provide additional comparisons between our model and the baseline models under matched computational budgets or model capacity. These results support the effectiveness of our design.

● We have updated Section 3.1 to provide further clarification on the loss settings and flow-time sampling distributions used for x-prediction and v-prediction training. X-prediction and v-prediction use the same flow-time sampling distribution and are trained with standard L1 and MSE losses, respectively, without any time weighting.

● We have added Table 11 and a new section A.6 to compare x-prediction and v-prediction using the same standard unweighted MSE loss and the same flow-time sampling settings, applied to clean-action targets and velocity targets, respectively. Under these settings, 1-step HRAE with x-prediction still outperforms both 10-step GR00T (DiT) with v-prediction and 10-step GR00T (DiT) with x-prediction.

● We have added Table 3 to provide a detailed analysis of the innovations and effectiveness of our model designs compared to prior works HRM/HRM-Text. HRAE achieves significant improvements in both computational efficiency (reducing FLOPs by 79.5%) and performance while introducing a complete redesign of the H-modules and L-modules in the hierarchical recurrent architecture. Moreover, incorporating recurrence at both the H-level and L-level is a well-established method in hierarchical recurrent models and was not first introduced by the HRM.

● We have provided Table 2 to compare our model with other one-step methods. These results demonstrate the effectiveness of HRAE compared with existing one-step methods.

● We have added Table 5 to present evaluation results on dexterous manipulation tasks for our model and the GR00T baseline. HRAE-Scratch outperforms StarVLA-GR00T while reducing policy latency.

● We have added Table 10 and a new section A.5 to the appendix to discuss our improvements in inference speed and task performance. We compare the design and performance of our models and the GR00T baseline to demonstrate that our lightweight module design enables faster inference, while deep hierarchical iteration improves task performance.

● We have updated the supplementary material to include the implementation of HRAE and examples of task execution failures for HRAE and the baseline models.

We sincerely thank you again for your thoughtful reviews. We greatly appreciate that several reviewers acknowledge our contribution to parameter-efficient recurrent design and strong task performance. We hope our work will inspire further research on one-step action generation in flow-matching VLAs.

Sincerely yours,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.