acceptodds
Under review as a conference paper at ICLR 2027

SOAR: Temporal Structure-Guided One-Step Action Generation for Robotic Manipulation

Abstract

Efficient robotic manipulation requires temporally coordinated actions with low inference latency. Multi-step diffusion and flow policies progressively refine action trajectories but require costly iterative inference, whereas one-step policies gain efficiency at the expense of temporal refinement. We introduce SOAR, a structure-guided one-step generative policy that improves action quality while preserving one-step efficiency. SOAR predicts an observation-conditioned discrete cosine transform (DCT) representation of future actions, capturing temporal variations across multiple frequency bands to provide explicit structural guidance. Unlike velocity-parameterized one-step flow policies that recover clean actions indirectly from a predicted velocity field, SOAR directly generates clean action trajectories in action space. This aligns the model output with the executable control target and provides a direct interface for structural calibration. The predicted structure further corrects the generated trajectory according to their spectral discrepancy, without requiring an additional evaluation of the generative backbone. During training, the corresponding MeanFlow velocity is derived from the calibrated clean action and used for consistency learning, while an endpoint objective aligns training with the pure-noise condition used for one-step inference. Across 53 Adroit and Meta-World tasks, SOAR achieves 83.9% overall success with only 8.7 ms inference latency. When applied to the VLA model π₀.₅, SOAR further achieves 97.4% on LIBERO and 60.0% on RoboTwin 2.0. Real-world experiments and extensive ablations further validate its effectiveness and efficiency.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotic Manipulation, Generative Policy, One-Step Action Generation, Efficient Visuomotor Policy, Frequency-Domain Action Modeling, Flow Matching, Action Generation Reasoning

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_LaV3OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SOAR, a one-step policy for robotic manipulation that predicts an action trajectory's discrete-cosine-transform (DCT) structure, uses low-, mid-, and high-frequency features to condition a direct action generator, and calibrates its output using the predicted spectrum. MeanFlow consistency and an endpoint objective train the calibrated action while inference uses one generator evaluation. On 3D Adroit and Meta-World benchmarks, the paper reports 83.9% overall success across 53 tasks at 8.7 ms latency, compared with 75.9% for the reproduced one-step MP1 baseline. It also reports one-step results on LIBERO and RoboTwin 2.0 and real-world π0.5 comparisons: 84% versus 88% on a 25-configuration tidy-up task, and 9/10 versus 8/10 successful dynamic pick-and-place trials. Ablations on the Adroit tasks suggest contributions from multi-band conditioning, calibration, and endpoint alignment.

Critical strengths
  1. SOAR gives a coherent formulation for adding explicit temporal structure to a one-step action generator. The DCT prediction, multi-band conditioning, action-space calibration, and inference-endpoint objective are defined mathematically, and the endpoint objective is shown to reduce to direct action matching at the one-step boundary. This makes the method's connection between training and inference relatively clear.
  2. The empirical evaluation is broad for a visuomotor-policy paper: it includes 53 simulated tasks, two multi-task VLA benchmarks, latency measurements, ablations, and two real-robot settings. The strongest controlled result is the 53-task comparison against reproduced one-step baselines, where SOAR reports a substantial success-rate gain at similar inference latency. The real-robot results also make the efficiency claim practically relevant.
Critical weaknesses
  • Real-world robot evaluations contain only 25 tidy-up configurations and 10 dynamic pick-and-place trials. The tidy-up success differs by merely one case (21/25 vs 22/25), and dynamic pick-and-place differs by one successful trial (9/10 vs 8/10). Such small differences cannot support a reliable task success advantage. While the reduced empty-grasp rate is promising, the paper does not report per-trial variability or uncertainty estimates. Furthermore, cross-paper benchmark comparisons may adopt inconsistent evaluation protocols. These limitations weaken the claims on robustness and generalization, though the demonstrated one-step latency reduction remains valid.
  • Multiple key components are inherited from existing one-step MeanFlow methods, including the dispersive regularizer from MP1. The proposed spectral calibration is essentially a simple residual correction operating in DCT space. Current ablations only test component effects on three Adroit tasks. It remains unclear how much of the full benchmark performance gain originates from the novel spectral mechanism, rather than the established one-step training recipe or hyperparameter tuning interactions.
  • The overall novelty is limited; the proposed method appears to stack multiple existing techniques. The design uses DCT variation as a loss regularizer and smoothed actions as another loss regularizer, where the smoothing operation is implemented via the neural model. The design directly adds DCT-transformed condition features and raw condition features as residuals, which introduces unexplained information overlap across two different domains.
  • The frequency domain is partitioned into high, mid, and low frequency bands. The hyperparameter choice of 3 frequency bands lacks ablation support: it is unclear whether 2 bands are sufficient, or whether further splitting into 4 bands would yield better performance. Experimental justification for this partition is required.
  • SOAR’s benchmark performance is built upon MP1 / MeanFlow components. If baselines like DP, DP3, CP, and FlowPolicy are augmented with the same set of techniques, it remains unknown whether SOAR can still outperform them.
  • In Libero experiments, SOAR uses PI05 as the model backbone. Existing results show PI05 already maintains full accuracy at NFE=1 on Libero. This critical baseline comparison is missing from the paper. Libero results must be reported with mean and standard deviation across multiple random seeds. Additional experiments on Libero-Pro or Libero-Plus are required.
Questions for authors
  1. Are the five selected checkpoints evaluated on the same episodes used for checkpoint selection? If yes, please report results using a disjoint held-out evaluation set and the final training checkpoint, with uncertainty estimates across multiple training seeds. This directly impacts the credibility of benchmark performance gains.
  2. Can you provide paired per-configuration outcomes and uncertainty metrics for both real robot experiments, especially dynamic target displacement trials? This will verify whether the observed differences in success rate and empty-grasp rate are statistically stable enough to support the claimed robustness improvement.
  3. What are the principled and empirical differences between SOAR and FreqPolicy, FAFM, FreqFM, especially those methods already applying supervision over DCT action structures? Could you add a matched one-step baseline equipped with a DCT-domain objective, and isolate the individual contribution of target spectrum prediction and the proposed action-space calibration?
  4. Could you provide full-benchmark ablations across extra training seeds: remove spectral calibration while keeping identical MeanFlow, endpoint, and dispersive objectives, and compare against a parameter- and compute-matched baseline with direct action correction? This will quantify the genuine contribution of spectral calibration, decoupled from the inherited one-step training recipe.
  5. Please clarify the design choice of residual addition between DCT-transformed condition and raw condition features. Why is summing features from two different domains reasonable, and how do you mitigate potential information overlap?
  6. Provide ablation studies to justify the hyperparameter of frequency band splitting. Compare performance under 2-band, 3-band, and 4-band frequency partitions.
  7. Please run control experiments: augment DP, DP3, CP, and FlowPolicy with MP1/MeanFlow components, and compare SOAR against these upgraded baselines.
  8. Add the critical PI05 baseline at NFE=1 for Libero. Report Libero results with mean ± standard deviation over multiple seeds. Supplement experiments on Libero-Pro or Libero-Plus.

I am open to raise my score if author solve my concerns.

5 comments

We thank the reviewer for the constructive comments on evaluation reliability, real-robot uncertainty, frequency-domain methods, calibration ablations, frequency partitioning, baseline controls, and VLA generalization. We have added independent evaluations, extended ablations, and implementation details to address these concerns.

Q1

Response: To address potential checkpoint-selection bias, we evaluated the final SOAR and MP1 checkpoints on independent episodes with matched initial configurations and three training seeds. SOAR retains higher success across all task groups, suggesting that its improvements are not solely due to favorable checkpoint selection.

1. Original evaluation protocol. Following prior work, including MP1, we evaluated models every 200 epochs and averaged the five highest periodic scores per seed. These evaluations were also used for checkpoint selection and were not independent held-out tests. Initial states were not explicitly matched across checkpoints.

2. Independent final-checkpoint evaluation. We evaluated the final checkpoints without performance-based selection, using separate episodes and matched initial configurations. Results are means ± standard deviations across three training seeds.

BenchmarkMP1SOAR
Adroit (3)71.11 ± 3.85%77.11 ± 3.67%
Meta-World Easy (28)91.90 ± 0.82%94.05 ± 2.06%
Meta-World Medium (11)63.64 ± 1.82%70.91 ± 4.81%
Meta-World Hard (6)66.67 ± 6.94%77.78 ± 9.62%
Meta-World Very Hard (5)64.00 ± 9.62%72.00 ± 5.25%

SOAR outperforms MP1 across all groups under independent final-checkpoint evaluation. Although this does not eliminate selection bias from the original top-five results, it provides additional evidence for the reported improvements.

See Appendix A.5.2 and Table 13.

Q2

Response: To make the real-robot comparisons more transparent, we report individual outcomes for 40 Tidy-up configurations and 20 dynamic trials per policy, including success indicators and empty-grasp counts.

1. Tidy-up and dynamic success. We expanded the evaluations and added π0.5 (NFE=1) as a control.

Taskπ0.5 (10)π0.5 (1)SOAR (1)
Tidy-up (40 configurations)75.0%70.0%82.5%
Dynamic pick-and-place (20 trials)80.0%65.0%85.0%

2. Empty-grasp variability. In dynamic manipulation, SOAR records 0.35 ± 0.49 empty grasps per trial, compared with 0.95 ± 0.83 and 1.20 ± 1.01 for π0.5 at NFE=10 and NFE=1, respectively (mean ± sample standard deviation).

The individual Tidy-up outcomes and dynamic trial records are provided in Appendix A.4, Tables 9–10. These results support a descriptive improvement in dynamic performance.

Q3

Response: SOAR differs from existing frequency-domain policies by explicitly predicting an observation-conditioned target action spectrum and using it for both action generation and calibration. We clarify this distinction through comparisons with FreqPolicy, FAFM, and FreqFM, together with 20-task component ablations. A fully parameter- and compute-matched one-step baseline with a direct DCT-domain loss remains incomplete due to training constraints.

1. Differences from prior work. Frequency-domain supervision itself is not new. The main distinctions are:

  • FreqPolicy (Su et al., 2025) applies frequency-domain consistency constraints across Flow Matching timesteps.
  • FAFM performs Flow Matching in DCT coefficient space and uses temporal-derivative regularization.
  • FreqFM incorporates frequency information through a spectrum-matched source distribution, frequency-balanced objectives, and frequency-constrained guidance.
  • SOAR predicts the target temporal structure directly from observations. Multi-band conditioning (MC) uses this prediction to guide one-step action generation, while action calibration (AC) refines the generated trajectory using its discrepancy from the predicted structure.

Thus, SOAR focuses on predicting and utilizing temporal structure, rather than only imposing frequency-domain constraints on generated actions.

2. Component evidence. Across 20 Adroit and Meta-World tasks:

VariantAverage success
Direct-action baseline69.3%
+ SS69.5%
+ SS + MC70.5%
+ SS + MC + AC71.0%
SOAR (+ EA)71.9%

Structure supervision (SS) alone provides only a 0.2-point improvement, consistent with its role in training the structure head rather than directly improving task performance. In contrast, MC and AC yield larger gains by incorporating the predicted temporal information into generation and calibration. Removing MC or AC from SOAR reduces success to 70.7% and 70.3%, respectively.

Q4

Response: Due to time and computational constraints, we could not complete full-benchmark ablations across all 53 tasks with additional training seeds. Instead, we conducted targeted ablations on 20 Adroit and Meta-World tasks with three independent training seeds, comparing SOAR without calibration and with direct action-space correction. All variants retain one-step inference and the same generation backbone and training protocol. The results support the contribution of spectral calibration beyond the shared MeanFlow training objectives.

1. Controlled variants. We evaluate:

  • SOAR w/o AC: Removes spectral calibration while retaining MeanFlow consistency, inference-endpoint alignment, and Dispersive Loss.
  • SOAR w/ direct correction: Replaces spectral calibration with direct action-space correction under the same backbone and training settings.

2. Quantitative comparison.

VariantMW MediumMW HardOverall (20 tasks)
SOAR w/o AC72.3%62.2%70.3%
SOAR w/ direct correction72.3%60.9%70.0%
SOAR74.3%63.5%71.9%

SOAR improves overall success by 1.6 percentage points over the no-calibration variant and 1.9 points over direct correction. The gains are particularly evident on challenging Meta-World tasks.

These results support the contribution of spectral calibration beyond the shared MeanFlow, endpoint, and dispersive objectives.

Q5

Response: SOAR does not directly add DCT coefficients to raw observation features. Instead, learned projections map each predicted frequency band into the common conditioning space, followed by gated residual addition. This preserves the original observation features while allowing the model to regulate structural guidance.

1. Feature fusion. The structure head predicts key temporal information from observations rather than independent executable actions. Each predicted frequency band is first transformed by a learnable projection into the observation-conditioning feature space, where it is residually combined with the original features.

2. Residual gating and information overlap. Since both branches originate from the same observation, information overlap is possible. We therefore use projections and gates to regulate the contribution of structural information and mitigate information overlap. Each frequency band has a learnable gate initialized to zero, allowing its structural contribution to be introduced gradually during training. The original observation features remain available through the residual connection.

See Appendix A.1.1, Eq. (8).

Q6

Response: We added frequency-partition ablations comparing two, three, and four bands across 20 tasks and three training seeds. All configurations use the same four DCT coefficients and one-step inference. The original three-band configuration achieves the highest overall success, supporting its use in the evaluated short-horizon setting.

1. Frequency partitions. All variants use K=T=4 and NFE=1, with unchanged structure prediction and calibration formulations.

ConfigurationFrequency groups
2 bands{0,1} / {2,3}
3 bands (SOAR){0,1} / {2} / {3}
4 bands{0} / {1} / {2} / {3}

All variants follow the same training and evaluation protocol with three independent seeds.

2. Results.

ConfigurationOverall success (20 tasks)
2 bands70.9%
3 bands (SOAR)71.9%
4 bands71.2%

The three-band configuration achieves 79.8% on Adroit, 74.3% on Meta-World Medium, and 63.5% on Hard. Its Medium-task success exceeds the two- and four-band variants by 1.48 and 0.85 percentage points, respectively, while differences on Hard are small.

These results support the three-band choice for T=4.

See Appendix A.5.5, Table 16.

Q7

Response: To examine whether SOAR's gains mainly arise from MP1/MeanFlow training components, we augmented existing policy backbones with the same training ingredients and evaluated them on the 37-task benchmark. SOAR retains the highest overall average, indicating benefits beyond the shared MP1/MeanFlow objectives.

1. MP1-augmented baselines. We retain each baseline's observation encoder and generation backbone while introducing MeanFlow/JVP training, dual-time conditioning, the guidance target from our local MP1 implementation, and Dispersive Loss (weight 0.5). All methods use NFE=1.

The adapted models include DP + MP1, DP3 + MP1, and FlowPolicy + MP1. For CP, removing the teacher and distillation objective makes its adapted student equivalent to our DP3 + MP1 implementation. We therefore report their shared result rather than treating CP as an independent fourth baseline.

2. Benchmark comparison.

MethodMeta-WorldOverall (37 tasks)
DP + MP181.6 ± 4.4%79.9 ± 4.2%
DP3 + MP181.9 ± 4.6%81.4 ± 4.3%
FlowPolicy + MP180.6 ± 4.6%81.1 ± 4.3%
SOAR84.8 ± 0.7%84.4 ± 0.5%

SOAR exceeds these augmented baselines by 4.5, 3.0, and 3.3 percentage points overall.

Since all augmented baselines already incorporate the MP1/MeanFlow training ingredients, SOAR's overall gains cannot be explained solely by these shared components. The results further support using predicted temporal structure to guide action generation and calibration.

Implementation details and full results are provided in Appendix A.5.6 and Table 17.

Q8

Response: We added the critical π0.5 (NFE=1) baseline and three-seed evaluations on all 40 LIBERO tasks. We also extended the VLA comparison to RoboTwin 2.0 with longer action chunks. SOAR outperforms the reproduced one-step π0.5 baseline on both benchmarks. LIBERO-Pro and LIBERO-Plus evaluations remain incomplete due to training and benchmarking constraints.

1. LIBERO one-step and multi-seed comparisons. All models are evaluated using three independent training seeds and 50 episodes per task per seed.

ModelNFESuccess rate
π0.51096.50 ± 0.35%
π0.5296.10 ± 0.43%
π0.5195.48 ± 0.95%
π0.5-SOAR197.21 ± 0.34%

SOAR improves success by 1.73 percentage points over the one-step π0.5 baseline and 0.71 points over the ten-step baseline, while requiring only one action-expert evaluation.

2. RoboTwin 2.0 evaluation. We further evaluated 50 tasks using 2,500 clean demonstrations and 50-step action chunks under the clean-to-clean setting.

ModelNFESuccess rate
π0.51070.0%
π0.5269.1%
π0.5165.1%
π0.5-SOAR171.7%

Together, these results support SOAR's effectiveness across different VLA tasks and longer action horizons.

The updated comparisons are included in the Simulation Experiments with VLA Settings section. LIBERO statistics are reported in Appendix A.5.8, Table 19, and RoboTwin configurations in Appendix A.3.

Reviewer_fv8ZOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SOAR, a one-step action-generation framework built on MeanFlow. An observation-conditioned branch predicts the DCT representation of future actions, whose frequency-band features condition a generator that directly outputs a clean action chunk. A learnable spectral residual then calibrates the generated trajectory. Training combines frequency-domain supervision, action reconstruction, MeanFlow consistency, inference-endpoint alignment, and an inherited feature regularizer. Inference requires one evaluation of the generative backbone. The method is evaluated on Adroit and Meta-World, integrated into π₀.₅ for LIBERO and RoboTwin 2.0, and tested on real robot tidying and dynamic pick-and-place tasks. The paper reports improved simulation success and substantially reduced action-generation latency while maintaining competitive real robot performance.

Critical strengths
  • Broad empirical evaluation. The experiments cover 3D manipulation benchmarks, multi-task VLA settings, and physical robot tasks. The Adroit and Meta-World results include three training seeds.
  • Practical integration with π₀.₅. SOAR replaces the iterative action expert with one-step generation while maintaining competitive success rates, demonstrating potential deployment value.
  • Concrete real robot validation. Tidying and dynamic pick-and-place experiments evaluate both long-horizon execution and responses to target displacement, using matched configurations for the compared policies.
  • Useful latency reporting. The paper separates action-generation latency from end-to-end inference latency, showing that the acceleration remains meaningful after accounting for the unchanged vision-language backbone.
Critical weaknesses
  • Limited methodological novelty. The method builds on MeanFlow and inherited regularization, with the contribution primarily consisting of an auxiliary spectral branch, conditioning, calibration, and additional supervision.
  • The central hypothesis is insufficiently validated. The paper does not directly establish that missing temporal structure is the main limitation of one-step policies, or that the proposed components specifically resolve this limitation.
  • The need for duplicate trajectory prediction remains unclear. A complete DCT representation already determines the action sequence, and the calibration reduces to a scalar linear combination of two trajectories. Comparisons with a time-domain auxiliary branch, a frequency-only generator, and capacity-matched alternatives are missing.
  • The ablations provide limited mechanistic evidence. They cover only three Adroit tasks, with Hammer saturated across all variants. Structure supervision and conditioning are initially added together, and several component gains are small relative to the reported variability.
  • The evaluation protocol requires clarification. Averaging the five highest evaluation scores during training may produce optimistic estimates. The separation between model selection and final testing, and consistency of this protocol across baselines, should be clearer.
  • Real robot results do not isolate the proposed design. The comparison against ten-step π₀.₅ supports acceleration, but lacks a simpler one-step baseline or component ablations. With only 25 tidying trials and 10 dynamic trials, the evidence for improved task performance is also limited.
Questions for authors
  1. Central hypothesis. What evidence shows that insufficient temporal structure is a specific limitation of one-step generation? Can this be measured directly, rather than inferred from task success?

  2. Choice of DCT. Does DCT outperform a time-domain representation or another orthogonal basis under matched model capacity and training conditions? Which benefit comes specifically from frequency-based organization rather than additional supervision?

  3. Two-branch prediction. What performance is obtained by directly executing the inverse DCT of the structure prediction? Since this branch already represents a complete trajectory, what evidence establishes the need for the second generator and their scalar fusion?

  4. Scope of ablations. Do the component gains persist beyond the three Adroit tasks, particularly with longer action chunks and in VLA settings? Can structure supervision, conditioning, and endpoint supervision be isolated more fully?

  5. Real robot comparisons. Would a simpler one-step action expert achieve similar latency and dynamic-task performance? A matched one-step baseline would help separate the benefit of reduced inference time from the proposed spectral design.

  6. Evaluation protocol. Are the evaluations used to select the five highest scores separate from final testing, and is the same procedure applied to reproduced baselines? Please also clarify the two π₀.₅ LIBERO results, 96.5% and 96.9%, and provide uncertainty estimates for the VLA comparisons.

  7. Reproducibility. Please specify the loss weights in Equation (1), the structure-head architecture and DCT configuration, and the real robot temporal regularization and event-centered sampling settings. Which settings are shared across benchmarks?

6 comments

We thank the reviewer for the constructive comments on temporal structure, DCT representation, the two-branch design, broader ablations, real-robot comparisons, evaluation protocols, and reproducibility. We have added trajectory-level analyses, component ablations, and experimental details to address these concerns.

Q1

Response: To directly examine temporal structure in one-step generation, we introduce temporal shape error (TSE) and amplitude-spectrum error (ASE), measuring trajectory-shape and Fourier magnitude-spectrum discrepancies relative to expert motions. Comparisons of DP3 at different inference steps and SOAR before and after calibration show that temporal discrepancies can be reduced through iterative refinement or structural calibration.

1. DP3 one-step versus ten-step. Across 53 Adroit and Meta-World tasks, we analyzed 318 saved final-checkpoint evaluation trajectories from three training seeds, without additional rollouts.

Metric (×10⁻³ m²)NFE=1NFE=10Reduction
TSE2.2062.0576.8%
ASE0.9960.85214.4%

Both errors decrease across all five task groups when NFE increases from 1 to 10.

2. SOAR structural calibration. Calibration reduces TSE from 3.438 to 2.092 (39.2%) and ASE from 1.483 to 0.936 (36.9%), in the same units, without additional backbone evaluations. Both improvements hold across all five groups.These results provide direct trajectory-level evidence rather than relying solely on task success.

Details are provided in Appendix A.5.1, Tables 11–12.

Q2

Response: We adopt DCT because it explicitly organizes temporal variations in action trajectories by frequency, separating slowly varying motion patterns from shorter-scale action changes. This allows the structure head to learn key temporal information for guiding action generation and calibration, rather than independently predicting accurate executable trajectories. Unlike direct time-domain representations, DCT naturally supports multi-band structural guidance. Although other orthogonal bases can also represent temporal structure, our choice of DCT is motivated by its simple, real-valued frequency representation rather than an assumption of inherent superiority.

1. Role of DCT supervision. DCT supervision trains the observation-conditioned structure head to predict informative temporal patterns. Multi-band conditioning (MC) uses the predicted structure to guide action generation, while action calibration (AC) refines the generated trajectories based on spectral discrepancies. Thus, the direct role of structure supervision is to train the structure head.

2. Component ablation results. Across 20 Adroit and Meta-World tasks, the average success rates are:

VariantAverage success
Direct-action baseline69.3%
+ SS69.5%
+ SS + MC70.5%
+ SS + MC + AC71.0%
SOAR (+ EA)71.9%

Adding structure supervision (SS) alone improves success by 0.2 percentage points, consistent with its role in training the structure head. Incorporating MC and AC produces larger gains by utilizing the predicted structure. Removing MC or AC from the complete SOAR reduces success to 70.7% and 70.3%, respectively.

These results indicate that the main benefits arise from utilizing predicted temporal structure for action generation and calibration, supporting the contributions of the individual components.

Q3

Response: We directly evaluate structure-only reconstruction, base actions, and calibrated actions to examine the necessity of both branches. The results show that the structure head cannot independently produce sufficiently accurate executable actions, but its predictions provide useful guidance for generation and calibration.

1. Three-output evaluation. Using the final SOAR checkpoints, we obtain:

BenchmarkStructure onlyBase actionCalibrated action
Adroit17.67%59.22%74.56%
Meta-World24.03%69.07%80.60%

The low structure-only success rates show that structural prediction cannot replace the generator. Calibration further improves success over the base action by 15.34 and 11.53 percentage points on Adroit and Meta-World, respectively.

2. Complementary roles. The predicted frequency bands guide generation through residual conditioning, while calibration corrects discrepancies between the generated actions and predicted structure.

Although the complete DCT allows reconstruction of a full trajectory, the structure head is not designed as an independent action policy. Thus, SOAR's contribution lies in the complementary use of temporal structure for generation and calibration.

The evaluation protocol and scalar-fusion equivalence are provided in Appendix A.5.3, Table 14.

Q4

Response: To examine whether SOAR's component gains extend beyond Adroit and short-horizon manipulation, we expanded the ablations to Meta-World Medium and Hard tasks and LIBERO-LONG. Due to time and computational constraints, we prioritized these targeted evaluations rather than exhaustive full-benchmark ablations. The results distinguish the contributions of structure supervision (SS), multi-band conditioning (MC), action calibration (AC), and inference-endpoint alignment (EA) in 3D manipulation, and further demonstrate the effectiveness of MC, AC, and EA in longer-horizon VLA settings.

1. Extended 3D ablations. We evaluated 20 tasks, including three Adroit tasks and 17 Meta-World tasks (11 Medium and six Hard), using three independent training seeds.

SOAR improves success from 71.7% to 74.3% on Meta-World Medium and from 60.9% to 63.5% on Hard. Across all 20 tasks, average success increases from 69.3% for the direct-action baseline to 71.9% for SOAR.

Progressive ablations show that SS, MC, AC, and EA contribute successive gains of 0.2, 1.0, 0.5, and 0.9 percentage points, respectively. SS is designed to train the structure head to predict key temporal information, rather than directly improve task performance. The larger gains from MC and AC indicate that the predicted structure becomes more useful when incorporated into action generation and calibration.

2. Longer-horizon VLA ablations. We further evaluated a SOAR-integrated VLA policy on LIBERO-LONG with prediction horizon H=10 and five executed actions per inference. All variants use NFE=1.

VariantSuccess rate
SOAR96.0%
w/o MC93.0%
w/o AC94.0%
w/o EA93.0%

Removing MC, AC, or EA reduces success by 3.0, 2.0, and 3.0 percentage points, respectively, confirming their contributions in the evaluated longer-horizon VLA setting.

Together, these results show that the benefits of structural conditioning, action calibration, and endpoint alignment extend beyond short-horizon 3D manipulation to VLA settings with longer action chunks. The extended 3D ablations also clarify that the main performance gains arise from utilizing predicted temporal structure rather than structural supervision alone.

Results are reported in Section 4.4, Table 5, and Appendix A.5.7, Table 18.

Q5

Response: To distinguish reduced inference steps from SOAR's design benefits, we added a real-robot comparison with π0.5 (NFE=1), alongside π0.5 (NFE=10) and π0.5-SOAR (NFE=1). The results show that simply reducing π0.5's inference steps does not reproduce SOAR's success rates, dynamic-task performance, or action-generation latency.

1. Tidy-up. Across 40 configurations, SOAR achieves 82.5% success, compared with 75.0% for π0.5 (NFE=10) and 70.0% for π0.5 (NFE=1). Relative to one-step π0.5, SOAR reduces action-generation latency by 54.4% while improving success by 12.5 percentage points.

2. Dynamic pick-and-place. We evaluated 20 trials per model, each involving four target displacements.

Metricπ0.5 (10)π0.5 (1)SOAR (1)
Placement success80%65%85%
Empty grasps per trial0.95 ± 0.831.20 ± 1.010.35 ± 0.49
Total empty grasps19247

Empty-grasp values are means ± sample standard deviations.

All methods use the same observations, action representation, and deployment platform.

The results are included in the real-world experimental section, result figure, and appendix.

Q6

Response: We clarified the top-five evaluation protocol, added independent final-checkpoint evaluations of SOAR and MP1, and distinguished previously reported from reproduced π0.5 LIBERO results. We also provide multi-seed LIBERO statistics.

1. Top-five evaluation and independent testing. For Adroit and Meta-World, SOAR and all reproduced 3D baselines follow the same procedure: three independent training seeds, evaluation every 200 epochs, and averaging the five highest periodic evaluation scores per task and seed, followed by statistics across seeds.

These periodic evaluations are also used for checkpoint selection and therefore are not independent held-out tests. To address potential selection bias, we additionally evaluated the final SOAR and MP1 checkpoints on new episodes with matched initial configurations. SOAR retains its advantage in these independent evaluations (Appendix A.5.2, Table 13).

2. LIBERO result clarification. The 96.9% π0.5 result is previously reported, whereas 96.5% is our reproduced π0.5 result at NFE=10. Reproduced baselines are marked with an asterisk (*) and used for direct comparisons with SOAR.

3. VLA uncertainty. We evaluated the reproduced π0.5 configurations and π0.5-SOAR on LIBERO using three independent training seeds, with 500 episodes per suite (2,000 episodes across four suites per seed).

ModelNFESuccess rate
π0.51096.50 ± 0.35%
π0.5296.10 ± 0.43%
π0.5195.48 ± 0.95%
π0.5-SOAR197.21 ± 0.34%

SOAR exceeds the reproduced NFE=10 and NFE=2 baselines by 0.71 and 1.11 percentage points, respectively, while requiring only one action-expert evaluation.

Due to time and computational constraints, we did not repeat the full RoboTwin 2.0 experiments across multiple training seeds. These limitations are clarified in Appendix A.5.8, Table 19.

Q7

Response: We added the loss coefficients, structure-head architecture, DCT configuration, and real-robot temporal regularization and event-centered sampling details. We also distinguish shared settings from benchmark-specific configurations to improve reproducibility.

1. Loss coefficients. In Eq. (1), λstruct=0.05, λact=0.01, λdis=0.5, and λend=0.05, with a MeanFlow coefficient of 1. These weights are shared across benchmarks.

2. Structure head and DCT. The 3D structure head uses three linear layers with two 512-dimensional hidden layers and Mish activations. All benchmarks use an orthonormal temporal DCT-II with the complete spectrum:

BenchmarkK
Adroit / Meta-World4
LIBERO10
RoboTwin 2.050
Real robot16

3. Real-robot training. First- and second-order temporal regularization on normalized joint actions uses weights of 0.15 and 0.05, respectively, with a Huber threshold of 1. The gripper is excluded. The losses are averaged over random-flow-time and one-step endpoint predictions, without dividing temporal differences by the sampling interval.

For event-centered sampling, a window is classified as an event window if a future gripper target differs from the current gripper state within the 16-step horizon. Training uses a 1:1 mixture of event-window and ordinary-window distributions; ordinary windows may also contain events.

4. Cross-benchmark settings. Shared settings include the outer loss weights, DCT-II convention, complete-spectrum representation, and 512-dimensional hidden width. Structure-head depth and activation, encoders, action horizons, and training schedules are benchmark-specific.

Details are provided in Appendix A.1.5 and the corresponding implementation appendices.

Reviewer_9mdfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SOAR, a one-step (NFE = 1) action generation policy built on MeanFlow. SOAR predicts the clean action chunk directly and derives the MeanFlow velocity from it. A structure head predicts the DCT coefficients of the action chunk from the observation. These coefficients are used both to condition the U-Net and to calibrate the generated action. An endpoint loss aligns training with the one-step inference condition. SOAR reaches 83.9% success on 53 Adroit and Meta-World tasks with 8.7 ms latency. When it replaces the action expert of π0.5, it reaches 97.4% on LIBERO and 60.0% on RoboTwin 2.0.

Critical strengths
  • The results are strong and consistent for a one-step policy on 53 Adroit and Meta-World tasks, especially on Medium and Hard tasks. Latency is only 0.6 ms higher than MP1.
  • The method transfers to π0.5 with a single action-expert evaluation. In the real-robot test, it reduces action-generation latency from 27.8 ms to 2.3 ms with similar success.
Critical weaknesses
  • Limited novelty. Eq. 6 follows pixel MeanFlow and JiT, which are not cited, and the adaptive frequency loss is close to FreqPolicy.
  • With K = T, Eq. 4 reduces to a linear blend between A_base and the deterministic prediction IDCT(C^s), not a real frequency-domain correction. With a horizon of 4, the "multi-band" structure has only 4 coefficients.
  • L_end and the blend both push the output toward E[x|c], which conflicts with the multimodality motivation. A regression baseline and a sample-diversity analysis are missing.
  • The ablation is only on Adroit, and the differences are within one standard deviation. The gain over MP1 may come mainly from x-prediction.
  • The performance gains are marginal and not clearly attributable to SOAR. π0.5-SOAR improves over π0.5 by only 3.0 points on RoboTwin 2.0 and 0.9 points on LIBERO. In the real world it is lower on Tidy-up. Since the main claim is efficiency, the key baseline is π0.5 with 1–2 Euler steps under the same training budget, which is missing. On the 37-task benchmark, SOAR beats OMP by only 2.1 points, and OMP is missing from the 53-task table.
Questions for authors
  1. Please report the learned tanh(ρ), the gate values, K, and the λ weights. If α is close to 1, SOAR is close to a regression policy.
  2. Please add a deterministic regression baseline, and an analysis of sample diversity or a task with multimodal demonstrations.
  3. Please run ablations on Medium and Hard Meta-World tasks, including variants without L_end and without x-prediction. Please also add OMP to the 53-task table.
  4. Please compare with π0.5 using 1–2 Euler steps under the same training budget on LIBERO and RoboTwin 2.0.
  5. Please clarify how SOAR differs from pixel MeanFlow and FreqPolicy.
3 comments

We thank the reviewer for the constructive comments on learned parameters, deterministic regression, extended ablations, low-NFE VLA comparisons, and the relationship to prior work. We have added parameter analyses, component evaluations, sample diversity measurements, and additional benchmark comparisons to address these concerns.

Q1

Response: We report the learned calibration coefficient, conditioning gates, DCT configuration, and loss weights from the trained checkpoints. We also analyze the exact calibration rule and its effect on conditional diversity. The measured coefficients remain far from α=1, providing no evidence that SOAR approaches deterministic regression through this mechanism.

1. Learned parameters. Appendix A.1.5 reports the best EMA checkpoints for three Adroit and 50 Meta-World tasks, with three training seeds per task (9 and 150 checkpoints). Values are means ± sample standard deviations across task–seed checkpoints.

ParameterAdroitMeta-World
α := tanh(ρ)0.118051 ± 0.0812760.075216 ± 0.053296
α range[0.007725, 0.220070][0.015644, 0.273769]
Low-frequency gate−0.002653 ± 0.0045630.000244 ± 0.009540
Mid-frequency gate0.000278 ± 0.005642−0.001454 ± 0.004646
High-frequency gate0.001601 ± 0.003492−0.003051 ± 0.005112

The gates are unconstrained learned scalars applied directly to projected features, without sigmoid or tanh. All checkpoints use K=T=4, with frequency groups {0,1}, {2}, and {3}. In Eq. (1), λstruct=0.05, λact=0.01, λdis=0.5, and λend=0.05; the MeanFlow coefficient is 1.

2. Regression behavior. With a complete DCT, Eq. (4) is exactly equivalent to an affine combination of the base action and deterministic structure prediction, weighted by 1−α and α. We agree that α approaching 1 would make the output approach deterministic regression. However, all measured values are far from this limit.

For a fixed observation and checkpoint, calibration scales the base generator's conditional covariance by (1−α)². The checkpoint-averaged scaling factors are 0.7837 on Adroit and 0.8580 on Meta-World. These measurements do not support collapse toward deterministic regression through α≈1.

Q2

Response: We added a deterministic regression baseline across 20 Adroit and Meta-World tasks using three training seeds, together with a conditional sample diversity analysis on Meta-World Basketball. The results show that SOAR achieves higher task success than deterministic regression while preserving sample variation after calibration.

1. Deterministic regression baseline. We use the same observation encoder and action U-Net as SOAR, with fixed zero-action and time inputs, trained using MSE without noise sampling. Both policies require one network evaluation (NFE=1) and follow the same task-specific training and evaluation budgets.

BenchmarkRegressionSOAR
Adroit (3)78.91 ± 2.70%79.8%
MW Medium (11)64.39 ± 1.65%74.3%
MW Hard (6)54.22 ± 2.28%63.5%
Overall (20)63.52 ± 1.25%71.9%

SOAR outperforms deterministic regression by 8.38 percentage points overall. This comparison evaluates the complete SOAR framework rather than individual component contributions.

2. Conditional sample diversity. On Basketball, we generated 32 samples for each of five fixed observations using different noise initializations (160 samples). We compared deterministic regression with SOAR before and after calibration.

Metric (×10⁻³)RegressionBefore ACAfter AC
Overall pairwise distance04.2383.834
Temporal-shape distance01.9931.945
Mean reference MSE22.88046.1970.733

Calibration retains 90.5% of overall sample diversity and 97.6% of temporal-shape diversity while reducing mean reference MSE by 98.4%. Unlike deterministic regression, calibrated SOAR retains nonzero conditional sample variation.

These results distinguish SOAR from deterministic action regression in both task performance and sampling behavior. Detailed protocols and results are provided in Appendix A.5.4 Deterministic Regression and Sample Diversity Analysis.

Q3

Response: We extended the component ablations to all 11 Medium and six Hard Meta-World tasks, alongside the three Adroit tasks, covering 20 tasks in total. The results support inference-endpoint alignment and show that direct action prediction achieves higher overall success than velocity prediction, although their relative advantages vary across tasks. We also clarify why OMP is not included in the 53-task comparison.

1. Extended ablations. The "+ SS + MC + AC" variant removes the inference-endpoint objective (L_end), while "SOAR w/ u-prediction" replaces direct clean-action prediction with velocity prediction, retaining the same structural guidance and calibration.

VariantMW MediumMW HardOverall (20 tasks)
SOAR74.3%63.5%71.9%
Without L_end73.6%62.4%71.0%
SOAR w/ u-prediction70.7%65.7%70.7%

Removing L_end reduces overall success by 0.9 percentage points, supporting inference-endpoint alignment. Velocity prediction performs better on Hard tasks, while direct action prediction achieves higher overall success (71.9% vs. 70.7%). Thus, the relative benefit of each parameterization depends on the task.

2. OMP comparison. The official OMP implementation was not publicly available, preventing reproduction under the same 53-task protocol. Its published 37-task results do not cover the additional 16 tasks.

The extended component ablations are included in the revised manuscript.

Q4

Response: To compare SOAR with low-NFE π0.5 under matched training budgets, we reproduced π0.5 using 1, 2, and 10 Euler steps on LIBERO and RoboTwin 2.0. Within each benchmark, the reproduced baselines and SOAR use the same training data, optimization steps, and evaluation protocol. SOAR consistently outperforms the 1–2-step baselines while requiring only one action-expert evaluation.

1. LIBERO comparisons. All methods are evaluated across 40 tasks using three independent training seeds, with a training budget of 30K optimization steps.

ModelNFESuccess rate
π0.51096.50 ± 0.35%
π0.5296.10 ± 0.43%
π0.5195.48 ± 0.95%
π0.5-SOAR197.21 ± 0.34%

SOAR improves success by 1.73 and 1.11 percentage points over π0.5 with NFE=1 and NFE=2, respectively.

2. RoboTwin 2.0 comparisons. We jointly train a single policy across 50 manipulation tasks with 50-step action chunks, using 100K optimization steps and clean-to-clean evaluation.

ModelNFESuccess rate
π0.51070.0%
π0.5269.1%
π0.5165.1%
π0.5-SOAR171.7%

SOAR outperforms the one-step and two-step baselines by 6.6 and 2.6 percentage points, respectively.

Across both benchmarks, SOAR also achieves higher success than the reproduced ten-step π0.5 baseline. These comparisons support the effectiveness of structure-guided one-step generation beyond simply reducing Euler integration steps.

The LIBERO multi-seed results and RoboTwin configurations are provided in the corresponding experimental appendices.

Q5

Response: We clarify SOAR's relationship to pixel MeanFlow, JiT, and FreqPolicy. SOAR adopts the data-prediction formulation of pixel MeanFlow but extends it to robotic action generation with explicit temporal structure prediction, conditioning, and calibration. Its contribution lies in how predicted structure participates in one-step generation, rather than in the inherited MeanFlow formulation or DCT supervision alone.

1. Relationship to pixel MeanFlow and JiT. We acknowledge that the data-to-average-velocity conversion in Eq. (6) shares its algebraic form with pixel MeanFlow and will explicitly cite pixel MeanFlow and JiT.

SOAR uses an observation-conditioned structure head to predict key temporal information about future actions. The predicted frequency components guide clean-action generation through multi-band conditioning and subsequently calibrate the generated trajectory. The corresponding MeanFlow velocity is derived from the calibrated clean action for consistency learning, while an additional endpoint objective at (t,r)=(1,0) aligns training with one-step inference.

2. Difference from FreqPolicy. FreqPolicy (Su et al., 2025) imposes frequency-domain consistency across different Flow Matching timesteps. In contrast, SOAR explicitly predicts the target action spectrum from observations and uses it for both conditioning and calibration.

Although both methods employ adaptive spectral weighting, SOAR weights supervised spectrum-prediction errors relative to target coefficient energy, rather than discrepancies between velocity predictions at different flow times.

Therefore, the key distinction is the prediction and utilization of observation-conditioned temporal structure, rather than frequency-domain supervision itself. We will clarify these relationships near Eq. (6) and in Appendices A.1.1–A.1.2.

Dear Reviewer,

We understand that the review period is particularly busy. As the discussion deadline approaches, we would sincerely appreciate any final feedback on our rebuttal and additional experiments, if your schedule permits.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.