- 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
- 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.
- The results are promising across several evaluation settings, especially the action-expert latency reduction.
Critical weaknesses
- 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.
- 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.
- 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
- 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?
- 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?
- 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
Response to Reviewer G2Td Part 2/2
Q3: Clarification of the loss setting and flow-time sampling distribution.
- 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.
| Model | Flow Steps | Prediction | Loss | RoboTwin |
|---|---|---|---|---|
| GR00T (DiT) | 10 | v-pred | MSE | 52.4 |
| GR00T (DiT) | 10 | x-pred | MSE | 52.2 |
| GR00T (DiT) | 10 | x-pred | L1 | 52.5 |
| HRAE | 1 | x-pred | MSE | 58.3 |
| HRAE | 1 | x-pred | L1 | 59.4 |
- 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 and set , where .
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.
| Model | Flow Steps | Pick up and place cube | Stack blocks | Place cube left of cylinder | Pick up and place cylinder |
|---|---|---|---|---|---|
| GR00T-N1.7 | 10 | 11/20 | 9/20 | 13/20 | 10/20 |
| π₀ | 10 | 8/20 | 8/20 | 13/20 | 12/20 |
| HRAE | 1 | 17/20 | 10/20 | 16/20 | 17/20 |
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.
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).
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).
For model capacity, HRAE-Lonly-Deeper matches HRAE’s parameter count (108.47M), but Deeper StarVLA-GR00T has substantially more parameters (202.83M).
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.
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.
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.
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.