acceptodds
Under review as a conference paper at ICLR 2027

StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models

Abstract

Vision-language-action (VLA) models increasingly rely on diffusion- or flow-matching-based action heads to generate continuous robot actions. These action heads typically process the denoising trajectory in a largely uniform manner. However, we observe that the conditioning focus naturally shifts across denoising stages: early stages combine language instructions and visual observations to establish a coarse action trajectory, whereas later stages place greater emphasis on current visual observations for action alignment. Based on this insight, we introduce StairVLA, a stage-aware hierarchical action generation framework that uses partially denoised actions as a natural interface between coarse long-horizon action generation and local refinement. A high-level VLA performs early denoising to produce a reusable long-horizon partially denoised action trajectory, while a lightweight refiner operates at a higher frequency to refine local action chunks using the latest observations. This design amortizes expensive high-level VLA computation while preserving frequent closed-loop correction. On LIBERO, our GR00T-style instantiation improves average success from 96.5% to 97.8% while reducing amortized inference latency from 115.0 ms to 44.2 ms per action chunk. More broadly, across two VLA backbones, simulation benchmarks, and real-robot tasks, StairVLA consistently reduces inference cost while maintaining strong task performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action models, robot learning, hierarchical action generation

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_JpA3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes StairVLA, a hierarchical VLA framework that splits flow-matching action generation into two stages. A high-level policy generates partially denoised long-horizon action trajectories, which are cached and reused across multiple control cycles. A lightweight refiner then uses updated observations to refine local action chunks. Experiments on LIBERO, LIBERO-Plus, and real-world robot tasks show reduced inference latency while maintaining competitive success rates.

Critical strengths
  1. Using partially denoised actions as the interface between high-level and low-level policies is an interesting idea. It avoids introducing explicit subgoals or other handcrafted intermediate representations.

  2. The method achieves considerable inference speedups while maintaining or slightly improving success rates on both GR00T-style and pi-style backbones.

  3. The evaluation covers multiple benchmarks, real-world experiments, and several useful ablations. The analysis of denoising stages and trajectory reuse also helps motivate the design.

Critical weaknesses
  1. The role of partially denoised actions is not sufficiently isolated. The refiner also receives compressed high-level VLM features, which already contain task-related information. It is therefore unclear how much task intent is actually transferred through the intermediate actions. The current ablations do not resolve this.

  2. The advantage over simpler coarse-to-fine approaches remains unclear. A high-level policy could generate a fully denoised trajectory, followed by a lightweight observation-conditioned refiner. The paper does not compare against this alternative, although the inference-time denoising level is already close to completion. This makes it difficult to assess the specific benefit of splitting the denoising process.

  3. The evidence for stage-dependent modality reliance is not entirely convincing. The attention trends differ between GR00T and pi-style models, and the attention analysis uses only 16 samples. The results suggest that conditioning changes across denoising stages, but do not firmly establish the proposed division between early task understanding and late visual refinement.

  4. The performance and efficiency comparisons are not fully controlled. The baselines use 8-action chunks while StairVLA uses 5-action chunks, making the reported per-chunk latency less directly comparable. StairVLA also introduces an additional refiner and training stage, so some performance gains may come from increased capacity or training rather than the proposed mechanism. The real-world results are encouraging, but the limited number of trials makes the improvements less conclusive.

Questions for authors
  1. Can the authors provide an ablation removing the cached high-level VLM features to clarify how much task information is carried by the partially denoised actions?

  2. How does StairVLA compare with a baseline that uses fully denoised long-horizon trajectories followed by the same observation-conditioned refiner? This seems important for establishing the benefit of the proposed interface.

  3. Can the authors provide stronger evidence for the claimed stage-dependent modality reliance, particularly given the different attention patterns observed across backbones?

  4. Can the authors report latency normalized by executed actions and account for high-level policy refresh overhead? It would also be useful to clarify how much of the improvement remains under matched model capacity and training budgets.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JpA3 (1)

We greatly appreciate the reviewer's positive feedback on the "interesting idea" of using partially denoised actions as the high-to-low-level interface, "considerable inference speedups", and the extensive evaluation across multiple benchmarks and real-world experiments. We address the remaining concerns below.

Q1 & Q2: Task-intent transfer and comparison with fully denoised trajectories.

We conduct two additional ablation experiments, with the results summarized below.

SettingPartial ActionsVLM FeaturesSuccess (%)Δ\Delta (pp)
VLM features only✗✓88.55-9.25
Partial actions only✓✗95.75-2.05
Fully denoised + refiner✗✓95.25-2.55
StairVLA (Ours)✓✓97.80--

The ablation results indicate that partially denoised actions can convey useful task information even without cached VLM features. Nevertheless, we have revised the statement in Section 1 to more carefully describe the role of partially denoised actions in transferring task information, alongside cached VLM features.

Why are VLM features alone insufficient? The VLM-features-only setting resembles conventional hierarchical VLA designs, which often require additional training mechanisms or intermediate supervision to effectively guide low-level policies.

Why is fully denoised refinement less effective? The strong fitting capability of the high-level VLA may limit effective residual learning after full denoising: although increasing GR00T's action horizon from 8 to 32 reduces closed-loop success from 96.5% to 92.4%, the per-action RMSE evaluated on training samples decreases from 0.0431 (H=8H=8) to 0.0237 (H=32H=32).

The corresponding experimental results and analysis have also been included in Table 3 of the revised manuscript.

Q3: Evidence for stage-dependent modality reliance.

We evaluate instruction and image masking at different denoising stages on GR00T over 400 LIBERO episodes. Here, s denotes the masking onset, with the corresponding conditioning masked from s to 1. The table reports success-rate changes (percentage points) relative to normal inference.

Mask Onset0.800.850.900.950.960.970.980.99
Instruction-35.50-30.50-18.00-4.75-1.75-1.00-1.00-0.25
Image-96.75-97.00-91.00-39.00-16.50-7.25-2.50-0.75

Late-stage denoising remains critical. Masking images during only the final 10% of denoising reduces success by 91.00 points, causing almost complete failure.

Visual conditioning dominates late-stage sensitivity. At 0.97, instruction masking reduces success by only 1.00 point, while image masking still causes a 7.25-point drop, consistent with our empirically selected α=0.97\alpha=0.97.

See revised Section 4.3 and Appendix B.1.4 for detailed results across both backbones. Although the patterns differ across backbones, the π\pi-style attention analyses (repeated with 1,024 samples, with consistent findings) and GR00T interventions jointly motivate reducing repeated language processing while retaining observation-conditioned refinement. Our main results further validate this design's performance--efficiency benefits.

AuthorsOpenReview ↗

Response to Reviewer JpA3 (2)

Q4: Fair latency comparison, model capacity, and training budget.

We use h=5h=5 on LIBERO primarily to match the action chunk size of Libra-VLA for a fairer comparison.

MethodChunk SizeSuccess (%)ms/chunkms/action
StarVLA-GR00T896.50115.014.38
Libra-VLA597.20122.024.40
StairVLA (Ours)597.8044.28.84

StairVLA reduces inference latency per executed action by 38.5% compared with StarVLA-GR00T.

In our real-world Fruit25 experiments, we use h=8h=8, matching the StarVLA baseline. The results were originally reported in Figure 10; we additionally provide per-action latency below.

MethodChunk SizeSuccess (%)ms/chunkms/action
StarVLA-GR00T872.8114.914.36
StairVLA (Ours)878.465.78.21

On Fruit25, StairVLA reduces per-action inference latency by 42.8% compared with StarVLA-GR00T.

Model capacity and training budget. Amortized over executed actions, StairVLA incurs lower average inference computation cost. To examine whether the performance gains arise merely from additional model capacity or training, we conduct a Deep GR00T ablation, doubling the action head from 16 to 32 layers. We follow the same two-stage training schedule as StairVLA, with 30k steps per stage; in the second stage, the VLM is frozen while the expanded action head is trained.

MethodAction HeadTraining StepsSuccess (%)Δ\Delta (pp)
StarVLA-GR00T16 layers30k96.50--
Deep GR00T32 layers30k + 30k95.70-0.80
StairVLA (Ours)16 layers + refiner30k + 30k97.80+1.30

These results suggest that simply increasing model capacity and training budget does not reproduce the performance gains of StairVLA, supporting the effectiveness of our hierarchical action generation design.

The corresponding experiments and analyses have been incorporated into Table 3 and the latency analysis in the revised manuscript and appendix.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Just a gentle follow-up on our rebuttal and additional experiments. We would really appreciate any feedback or further questions you may have.

Thank you again for your time and consideration!

Reviewer_Jzh4OpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents StairVLA, a stage-aware hierarchical action generation framework for flow-matching-based vision-language-action (VLA) models. Motivated by the observation that different denoising stages exhibit varying dependencies on language and visual conditioning, the authors decompose action generation into low-frequency high-level generation and high-frequency low-level refinement. The high-level VLA generates and caches a long-horizon partially denoised action trajectory, while a lightweight refiner leverages the latest visual observations, cached VLM features, and local action context to refine short-horizon action chunks. Experiments on LIBERO, LIBERO-Plus, two VLA backbones, and real-world robotic tasks demonstrate substantial reductions in average inference latency while maintaining or improving task success rates.

Critical strengths
  1. Meaningful and practically relevant problem. The paper addresses an important trade-off between inference efficiency and closed-loop control frequency in VLA models. Reducing the computational overhead of large VLA models while preserving responsiveness to changing observations is highly relevant to real-world robotic deployment.

  2. Simple and intuitive hierarchical design. The proposed framework directly uses partially denoised actions as the interface between high-level generation and low-level refinement, avoiding the need for explicitly designed intermediate representations such as subgoals or waypoints. This design naturally builds upon the existing flow-matching formulation and provides a relatively straightforward way to introduce hierarchical control into generative VLA models.

  3. Well-motivated refiner training strategy. The authors construct a refinement path that accounts for prediction errors in the high-level model, rather than relying solely on the ideal flow-matching trajectory. This design is technically reasonable, as the refiner is trained to correct intermediate actions produced by the actual high-level model, potentially improving its robustness to imperfect coarse predictions.

Critical weaknesses
  1. Limited novelty and relatively modest performance improvements. The overall design shares similarities with existing hierarchical and coarse-to-fine VLA approaches, particularly in its use of cached high-level information and lightweight low-level action refinement. The unique contribution of stage-aware denoising could be more clearly distinguished from these prior approaches. Moreover, the improvements on LIBERO are relatively modest (+1.3 and +0.5 percentage points) over already strong baselines, and additional statistical evidence would help establish their significance and consistency.

  2. Insufficient evidence for the stage-dependent hypothesis and task-intent transfer. While the paper presents analyses of attention distributions and prediction errors across denoising stages, these observations do not fully establish that different stages serve distinct functional roles. Furthermore, the low-level refiner receives both partially denoised actions and cached high-level VLM tokens, making it difficult to determine whether task-level information is primarily transferred through the intermediate actions or the cached VLM representations. This ambiguity weakens the evidence supporting one of the paper's central claims.

  3. Missing comparison with fully denoised action trajectories. The paper does not directly compare partially denoised actions against fully denoised actions as the input to the low-level refiner. Such a comparison would help clarify whether the proposed intermediate denoising representation offers advantages beyond using completed high-level action predictions followed by local correction. Without this baseline, the necessity and effectiveness of the proposed stage-aware decomposition remain insufficiently validated.

  4. Inconsistent attention patterns across VLA backbones. The π-style model exhibits increasing attention to image tokens during later denoising stages, whereas GR00T shows the opposite trend for image-token attention. Although the authors suggest that GR00T may encode visual information through contextualized grounding-related tokens, this interpretation is not directly verified. These differing observations make it difficult to draw a unified conclusion about the functional roles of different denoising stages across architectures.

  5. Fixed high-level trajectory reuse and limited long-horizon robustness. The framework refreshes the high-level trajectory at predefined intervals, without adapting to execution errors or changes in the environment. As shown in Appendix B.3, task success decreases noticeably as the trajectory reuse horizon increases, suggesting that the low-level refiner cannot fully compensate for the growing mismatch between cached trajectories and the current environment state. This limitation may reduce the method's effectiveness in long-horizon or highly dynamic manipulation scenarios.

  6. Limited complexity and statistical scale of real-world experiments. The real-world evaluation is primarily conducted on structured pick-and-place tasks in Fruit25 and a single pushing task in PushBlock, which may not sufficiently demonstrate the method's effectiveness in more complex manipulation scenarios. In addition, PushBlock includes only 10 trials per method, while some Fruit25 tasks are evaluated with only 5 trials. The relatively small evaluation scale raises concerns about the statistical reliability and generalizability of the reported performance improvements.

Questions for authors
  1. High-level/low-level interface and task-intent transfer. How would the performance change if the cached VLM tokens were completely removed from the refiner inputs, leaving only the partially denoised actions as the high-level conditioning? Conversely, could the refiner achieve comparable performance using only cached VLM tokens without partially denoised actions?

  2. Stage-dependent hypothesis. Could the authors conduct more direct intervention experiments by selectively removing language or visual conditioning at different denoising stages, and evaluate the resulting changes in final action accuracy and task success? Such experiments would provide stronger evidence for the proposed functional distinction between denoising stages.

  3. Denoising-stage selection and training–inference discrepancy. What motivates the choice of α = 0.97 as the boundary between high-level generation and low-level refinement? Given that denoising is nearly complete at this point, what types of corrections does the refiner primarily perform? Additionally, the refiner is trained using high-level outputs at α = 0.8 but typically evaluated at α = 0.97. Could this discrepancy introduce a distribution shift, and how sensitive is the performance to this difference?

  4. Real-world task complexity. Could the authors evaluate the proposed method on more challenging multi-stage manipulation tasks or scenarios requiring frequent replanning? Such experiments would help establish whether the benefits extend beyond relatively structured manipulation settings.

  5. Real-world inference latency. Could the authors report the worst-case or P95 inference latency during high-level trajectory refresh, in addition to the amortized latency? It would also be helpful to report the actual control-loop waiting time during real-world execution to better assess the method's practical real-time performance.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Jzh4 (1)

We greatly appreciate the reviewer's positive feedback on the "meaningful and practically relevant problem", "simple and intuitive hierarchical design", and "well-motivated refiner training strategy". We address the remaining concerns below.

W1: Limited novelty and modest performance improvements.

Novelty. CF-VLA introduces a learned action initialization with KL-based supervision to reduce denoising steps, while hierarchical VLAs employ specially designed interfaces to reduce VLM invocation frequency. In contrast, we achieve both. Motivated by the observed stage-dependent conditioning, we establish the hierarchy within the action generation process itself, with a simpler design, greater efficiency, and higher task success.

Significance of improvements. We have not yet conducted multi-seed training. However, LIBERO improvements are consistent across two VLA backbones and multiple trajectory reuse settings, suggesting the gains are not configuration-specific. More importantly, our contribution lies in improving the performance--efficiency trade-off. As VLMs continue to scale, avoiding redundant VLM invocations becomes increasingly valuable without sacrificing frequent observation-based feedback.

These distinctions have been further clarified in the revised Introduction and Related Work sections.

Q1: High-level/low-level interface and task-intent transfer.

We add LIBERO ablations, shown in revised Table 3.

SettingPartial ActionsVLM TokensSuccess (%)Δ (pp)
VLM tokens only✗✓88.55-9.25
Partial actions only✓✗95.75-2.05
StairVLA (Ours)✓✓97.80--

Partial actions alone achieve 95.75% success, indicating effective task-information transfer without VLM tokens. In contrast, VLM tokens alone achieve 88.55%, suggesting that feature-only hierarchical designs may require additional supervision or training mechanisms.

Q2: Evidence for the stage-dependent hypothesis.

We conduct instruction and image masking experiments at different denoising stages on GR00T over 400 LIBERO episodes. Here, s denotes the intervention onset, with modified conditioning applied from s to 1. The table reports success-rate changes (percentage points) relative to normal inference.

Mask Onset0.800.850.900.950.960.970.980.99
Instruction-35.50-30.50-18.00-4.75-1.75-1.00-1.00-0.25
Image-96.75-97.00-91.00-39.00-16.50-7.25-2.50-0.75

Late-stage action generation is substantially more sensitive to visual than language conditioning. From s = 0.95 to 0.97, instruction masking causes only a 4.75--1.00 pp success drop, compared with 39.00--7.25 pp for image masking.

See revised Section 4.3 and Appendix B.1.4 for both backbones. The π-style attention analyses and GR00T interventions motivate reducing repeated language processing while retaining visual feedback. Our main results further validate this design's performance--efficiency benefits.

Q3: Denoising-stage selection and training–inference discrepancy.

Denoising-stage selection. We select α = 0.97 based on the inference-time denoising-level ablation in Figure 7, where it achieves the highest success rate on LIBERO. Additionally, Q2 shows that masking language and visual conditioning at α = 0.97 reduces success by 1.00 and 7.25 pp, respectively, further supporting this choice.

Training--inference discrepancy. In our early exploration, we experimented with training distributions concentrated near the inference denoising level, but obtained unsatisfactory results.

SettingTraining αInference αSuccess (%)Δ (pp)
StarVLA-GR00T----96.50--
StairVLA (Matched Levels)U(0.95, 1)0.9594.05-2.45
StairVLA (Matched Levels)U(0.97, 1)0.9794.25-2.25
StairVLA (Matched Levels)U(0.98, 1)0.9893.85-2.65
StairVLA (Ours)U(0.80, 1)0.9797.80+1.30

Our preliminary observation is that with nearly fully denoised training inputs, the refiner tends to continue high-level denoising with limited correction from updated visual observations. We therefore adopt different training and inference settings: training presents more challenging correction tasks to encourage the refiner to learn observation-conditioned corrections, while inference uses more accurate high-level trajectories to reduce the difficulty of execution-time refinement.

We believe this training design deserves further investigation. Given the limited rebuttal period, we can only provide these early exploratory results as preliminary evidence, since their training hyperparameters and some model configurations were not fully consistent. We plan to conduct more controlled experiments and add the corresponding analysis in a subsequent manuscript revision.

AuthorsOpenReview ↗

Response to Reviewer Jzh4 (2)

Q4 & Q5: Real-world task complexity and closed-loop latency.

We introduce a five-stage long-horizon drawer manipulation task: opening the drawer, grasping a fruit, placing it inside, closing the drawer, and returning to the initial configuration. The following table reports stage completion during sequential execution and overall task performance.

MethodOpenGraspPlaceCloseReturnAvg.Complete
Full-Task Evaluation (5 Stages)
StarVLA-GR00T5/100/100/100/100/100.50/10
StarVLA-π\pi2/100/100/100/100/100.20/10
StairVLA (Ours)10/104/103/100/100/101.70/10
Conditional Evaluation (4 Stages)
StarVLA-GR00T--7/106/100/103/101.60/10
StarVLA-π\pi--10/106/101/103/102.00/10
StairVLA (Ours)--10/1010/105/103/102.83/10

Given the challenges of drawer opening, closing, and transitions between multiple stages, we additionally conduct a conditional evaluation starting with the drawer already open. StairVLA achieves the highest average number of completed stages in both settings. We also observe more consistent action execution, with fewer repetitive adjustments or stalls, potentially benefiting from long-horizon high-level action guidance.

We further measure real-world closed-loop execution latency in the same experiment.

MetricStarVLA-GR00TStarVLA-π\piStairVLA
Policy Execution (ms/call)93.70279.3654.53
Obs.-to-Command (ms)110.07296.4971.64
Obs.-to-Command P95 (ms)119.06311.26121.78
Command Rate (Hz)17.3312.3318.84

Notably, by reducing denoising steps and amortizing high-level VLM invocations, StairVLA achieves lower average closed-loop latency. Its overall P95 observation-to-command latency is 121.78 ms, comparable to StarVLA-GR00T (119.06 ms), although high-level refreshes may incur additional computation.

These results demonstrate that StairVLA improves task execution performance while reducing average closed-loop control latency and achieving a command update rate closer to the target 20 Hz.

Detailed experimental results are provided in Appendix C.8 of the revised manuscript.

W5: Fixed high-level trajectory reuse and limited long-horizon robustness.

We acknowledge this limitation. StairVLA currently employs a fixed high-level trajectory refresh schedule and does not adaptively trigger regeneration based on execution errors or environmental changes. We agree that longer reuse horizons or substantial environmental changes may lead to trajectory drift. Adaptive high-level regeneration is an important direction for future work.

However, we emphasize that the primary goal of StairVLA is not to completely eliminate long-horizon trajectory drift, but to decouple frequent visual feedback from high-level VLM inference. In conventional unified VLA architectures, generating actions from updated observations typically requires repeating the full VLM inference process. Consequently, the VLM invocation frequency is dictated by the need for frequent visual feedback, not necessarily by the need for high-level reasoning. StairVLA removes this constraint, allowing the two frequencies to be determined independently.

For example, with a target control frequency of 20 Hz and an action chunk size of 8, new observations are incorporated into action generation every 0.4 s, typically requiring a high-level VLM invocation at the same interval. With StairVLA, high-level trajectory reuse can extend the VLM invocation interval to 1.2 s, while the low-level refiner still incorporates updated visual observations every 0.4 s to correct local actions.

Thus, StairVLA does not assume that long-horizon execution can proceed without feedback. Instead, it separates frequent observation-conditioned local refinement from less frequent high-level action generation, reducing redundant VLM computation while preserving closed-loop control.

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Just a gentle follow-up on our rebuttal and additional experiments. We would really appreciate any feedback or further questions you may have.

Thank you again for your time and consideration!

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

This paper proposes StairVLA, a hierarchical action generation framework that separates early-stage coarse action generation from late-stage refinement. A high-level VLA generates and caches partially denoised long-horizon action trajectories, while a lightweight refiner uses updated visual observations to produce executable local action chunks. This design reduces repeated high-level inference while maintaining frequent visual feedback. Experiments on LIBERO, LIBERO-Plus, and real-world manipulation tasks demonstrate improved inference efficiency with competitive task performance.

Critical strengths

Interesting motivation and practical design. The paper explores stage-dependent behavior in generative VLA models and uses partially denoised actions as an interface between high-level planning and low-level refinement. This is a reasonable approach to reducing expensive VLM computation without introducing explicit subgoal representations.

Meaningful efficiency improvements and useful ablations. The method achieves substantial inference acceleration while maintaining strong task performance. The component ablations, reuse-horizon analysis, and real-world evaluations provide useful evidence for the effectiveness and practical value of the hierarchical design.

Critical weaknesses

Insufficient comparisons to isolate the main contribution. The performance gains may result from several factors, including reduced high-level inference frequency, trajectory caching, and the additional observation-conditioned refiner. Although the paper includes component ablations, it lacks direct comparisons with alternative hierarchical designs that use the same reuse schedule and refiner but different high-level representations, such as fully denoised action trajectories. Therefore, the specific benefit of using partially denoised actions as the high-to-low-level interface remains unclear. Comparisons with recent hierarchical VLA methods are also difficult to interpret because of differences in architectures, inference settings, and latency measurement protocols.

Limited evidence for the proposed stage-dependent mechanism and broader robustness. The attention analysis shows different trends across the two backbones, and the appendix acknowledges that attention allocation does not establish causal dependence on conditioning modalities. More direct evidence is needed to support the proposed separation between early task-level generation and late visual refinement. Furthermore, the real-world results are encouraging but limited: the long-horizon fruit-sorting task shows no improvement over the GR00T baseline, while PushBlock is evaluated with only ten trials per method. The relatively weak zero-shot performance under camera perturbations also raises questions about the robustness of the visual refiner.

Questions for authors

Can the authors compare StairVLA with a matched hierarchical baseline that caches fully denoised high-level trajectories and uses the same low-level refiner, prediction horizon, and reuse schedule? This would help isolate the contribution of the partially denoised action interface from general hierarchical inference and caching.

Appendix Table 4 shows that the StarVLA baselines return 8-action chunks, whereas StairVLA returns 5-action chunks. Can the authors report latency normalized by the number of executed actions, together with end-to-end control latency under matched execution frequencies? This would provide a clearer comparison of practical efficiency.

Can the authors provide more direct evidence for the stage-dependent conditioning hypothesis, such as controlled image/instruction masking experiments across both backbones at their actual inference denoising steps? In particular, how important is the proposed stage separation when the high-level trajectory is already denoised to 0.97 at inference?

Can the authors further evaluate robustness under execution perturbations or unexpected changes to the scene? Also, given the limited improvement on the real-world long-horizon sorting task and the sensitivity to camera shifts on LIBERO-Plus, how reliably can the low-level refiner correct deviations from a cached high-level trajectory?

3 comments
AuthorsOpenReview ↗

Response to Reviewer UovR

We greatly appreciate the reviewer's positive feedback on our "clear and easy to follow" presentation, "interesting motivation and practical design", and "meaningful efficiency improvements". We address the remaining concerns below.

Q1: Comparison with fully denoised hierarchical baselines.

We additionally evaluate a Fully Denoised + Refiner baseline under matched settings.

MethodAction HorizonLIBERO Success Rate (%)
GR00T896.5
GR00T (Top-only)3292.4
Fully Denoised + Refiner3295.3
StairVLA (Ours)3297.8

Why is fully denoised refinement less effective? The strong fitting capability of the high-level VLA may limit effective residual learning after full denoising: Although increasing GR00T's action horizon from 8 to 32 reduces closed-loop success from 96.5% to 92.4%, the per-action RMSE evaluated on training samples decreases from 0.0431 (H=8H=8) to 0.0237 (H=32H=32).

The additional experiment is included in the revised manuscript, Section 4, Table 3.

Q2: Fair latency comparison across chunk sizes.

We set the low-level action chunk size to 5 primarily to maintain consistency with Libra-VLA.

MetricStarVLA-GR00TLibra-VLAStairVLA (M=4)
LIBERO Success (%)96.597.297.8
Chunk Size855
Inference (ms/chunk)115.0122.044.2
Inference (ms/action)14.3824.408.84

After normalization, StairVLA reduces inference latency per executed action by 38.5% compared with StarVLA-GR00T under the same hardware and measurement protocol. The updated results are provided in the revised manuscript Appendix Table 5.

We further evaluate closed-loop latency on the real-world drawer task. StairVLA reduces mean observation-to-command latency by 34.9% compared with StarVLA-GR00T, and achieves a higher command rate. The details are included in the revised manuscript, Appendix C.8.

MetricStarVLA-GR00TStarVLA-π\piStairVLA
Policy Execution (ms/call)93.70279.3654.53
Obs.-to-Command (ms)110.07296.4971.64
Obs.-to-Command P95 (ms)119.06311.26121.78
Command Rate (Hz)17.3312.3318.84

Q3: Evidence for stage-dependent conditioning.

We evaluate instruction and image masking at different denoising stages on GR00T over 400 LIBERO episodes. The table reports success-rate changes (percentage points) relative to normal inference.

Mask Onset0.800.850.900.950.960.970.980.99
Instruction-35.50-30.50-18.00-4.75-1.75-1.00-1.00-0.25
Image-96.75-97.00-91.00-39.00-16.50-7.25-2.50-0.75

Late-stage denoising remains critical. Masking images during only the final 10% of denoising reduces success by 91.00 points, causing almost complete failure.

Visual conditioning dominates late-stage sensitivity. At 0.97, instruction masking reduces success by only 1.00 point, while image masking still causes a 7.25-point drop, consistent with our empirically selected α=0.97\alpha=0.97.

See revised Section 4.3 and Appendix B.1.4 for detailed results across both backbones. Although the patterns differ across backbones, the π\pi-style attention analyses and GR00T interventions jointly motivate reducing repeated language processing while retaining observation-conditioned refinement. Our main results further validate this design's performance--efficiency benefits.

Q4: Robustness to perturbations and trajectory deviations.

We conduct two additional real-world experiments:

  • Object Perturbations. We manually perturb target fruits during Fruit25 execution (40 trials per method). StairVLA achieves 35.0% success, outperforming GR00T (30.0%) and StarVLA-π\pi (27.5%).

  • Long-Horizon Manipulation. We introduce a five-stage drawer task. StairVLA completes 1.7 stages on average in full-task evaluation (vs. 0.5 and 0.2), and 2.8 in conditional evaluation (vs. 1.6 and 2.0).

These results further demonstrate StairVLA's advantages in adapting to scene changes and multi-stage manipulation. Detailed results are provided in Appendix C.7 and C.8 of the revised manuscript.

LIBERO-Plus Camera Perturbations. The original LIBERO training data uses fixed camera viewpoints, while LIBERO-Plus introduces unseen camera variations. Since our refiner relies on updated visual observations, it is more sensitive to such shifts under zero-shot transfer. After LIBERO-Plus fine-tuning, StairVLA achieves 95.5% on Camera, surpassing Libra-VLA (94.5%), demonstrating more effective action refinement after adapting to the new visual distribution.

Reviewer_UovROpenReview ↗

Thank you for your response.

The matched fully denoised trajectory + refiner baseline directly addresses my main question. With the same 32-step horizon, the improvement from 95.3% to 97.8% gives me clearer evidence for the value of the partially denoised interface. Reporting latency per executed action and observation-to-command latency on the robot also makes the practical efficiency gain easier to assess.

I found the GR00T masking experiment particularly informative. At the inference setting of 0.97, masking images reduced success by 7.25 percentage points, compared with 1 point for masking instructions. Could the authors clarify whether a comparable intervention was performed on the π-style backbone, or whether the evidence there comes from attention analysis? This affects how broadly I would interpret the proposed mechanism. The added experiments strengthen my assessment of the design, and I will maintain my rating of 3 (weak acceptance).

AuthorsOpenReview ↗

Response to Reviewer UovR (Follow-up)

We sincerely thank the reviewer for the thoughtful follow-up and positive assessment of our additional experiments. We greatly appreciate the reviewer's constructive feedback, which has helped us strengthen the experimental evidence and improve the clarity of our manuscript. Regarding the reviewer's follow-up question, we would like to further clarify the experimental evidence on the π-style backbone.

Yes, we conducted the same conditioning interventions on the π-style backbone as on GR00T. Specifically, we evaluated instruction masking and image masking at different denoising stages, with 400 LIBERO episodes per setting. The complete results are reported in Appendix B.1.4 (Table 4 and Figure 14) of the revised manuscript.

The following table summarizes the success-rate changes (percentage points) relative to the corresponding unmasked baselines.

Mask Onset0.800.850.900.950.960.970.980.99
Instruction-54.00-51.75-45.50-24.00-17.75-8.25-0.50+0.25
Image-95.50-95.00-96.50-59.25-36.00-12.00-2.25+0.50

The results show that image masking generally causes larger success-rate drops than instruction masking on the π-style backbone. For example, at s = 0.95, image and instruction masking reduce success by 59.25 and 24.00 percentage points, respectively. At s = 0.98, the corresponding drops are 2.25 and 0.50 percentage points. However, π-style retains greater sensitivity to instruction masking in the late denoising stages, making the separation between visual and language intervention sensitivity less pronounced than on GR00T.

Interestingly, the two backbones exhibit contrasting patterns across the two analyses:

  1. Conditioning interventions (as shown above and in Appendix B.1.4, Figure 14): GR00T exhibits a clearer separation between visual and language intervention sensitivity during late-stage denoising, whereas this separation is less pronounced on π-style.

  2. Attention analysis (Appendix B.1.2, Figure 11): We analyze how the cross-attention of action queries is distributed across different conditioning-token categories at different denoising stages. The π-style backbone exhibits a clear increase in attention to image tokens during late-stage denoising, whereas GR00T does not show an equally clear trend.

These differences may be related to the conditioning architectures of the two backbones. GR00T uses the final-layer VLM hidden states, whereas π-style conditions different action-head layers on hidden states from multiple VLM layers.

Overall, the two backbones exhibit stage-dependent changes in conditioning behavior from different perspectives—intervention sensitivity and attention allocation—providing strong empirical motivation for our stage-aware action generation design.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.