acceptodds
Under review as a conference paper at ICLR 2027

OneRobo: Temporally Structured Distribution Matching for One-Step Robot Policies

Abstract

Diffusion and flow-matching based robot policies model continuous action distributions but incur inference latency from iterative sampling. Few step distillation reduce this overhead, yet existing methods overlook the within-chunk temporal structure and cross-chunk execution consistency required for robot control. We introduce OneRobo, a one-step distillation framework for robot policies that preserves temporal structure. To accommodate the distinct dynamics of arm motion and continuous gripper control, we adopt channel-specific structured generation: arm generation and its distillation directions are constrained to a low-frequency discrete cosine transform (DCT) subspace to suppress unnecessary temporal variations, while teacher supervision on gripper changes across timesteps promotes stability. To improve cross-chunk consistency, we incorporate aligned historical-action initialization and learnable adapters into training. These mechanisms form a differentiable action-generation pipeline, with distribution matching and auxiliary supervision jointly applied to the final structured outputs. Across 50 RoboTwin simulation tasks with FastWAM, OneRobo reduces action generation from ten steps to one, with only an drop in average success rate relative to the multi-step teacher and an inference speedup. Additional experiments on and multiple real-world tasks further demonstrate its broad effectiveness.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Policies, Diffusion Distillation, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yHjnOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents OneRobo, a method for distilling multi-step flow-matching robot policies (VLAs and WAMs) into one-step action generators using Distribution Matching Distillation. The authors argue that existing one-step distillation methods can achieve low pointwise action error while introducing high-frequency oscillations and frequent direction reversals, and that they ignore consistency between consecutive action chunks. OneRobo makes three modifications to DMD. First, arm trajectories are restricted to a low-frequency subspace spanned by the first K DCT-II bases, and this projection is applied to the initial noise, the predicted velocity, the output, and the distribution-matching update. Second, gripper channels are left unconstrained and are instead supervised to match the frozen ten-step teacher’s command values and first- and second-order temporal differences. Third, the unexecuted portion of the previous chunk is used as a history prior, both by blending it into the initial noise and by injecting it as additive action-token embeddings. The method is evaluated with Fast-WAM on all 50 RoboTwin tasks, with π0.5 and SmolVLA on LIBERO, and with π0.5 on two real-world tasks on an ALOHA-AgileX platform. On RoboTwin, the one-step student reaches 91.16% (Easy) and 89.72% (Hard) average success, compared with 92.80% and 91.96% for the ten-step teacher, with a 3.23× speedup. Component ablations and DCT truncation sweeps are reported on SmolVLA LIBERO-Spatial.

Critical strengths
  1. The paper identifies a concrete and practically relevant failure mode of one-step policy distillation. Table 1 and Fig. 2 show that a distilled student can match the teacher closely in action MSE while its direction-reversal rate rises from about 5% to 27%. This is consequential because it shows that the pointwise objectives and metrics commonly used to evaluate distilled policies can hide execution-level degradation, and it gives a clear rationale for the proposed design.

  2. The channel-specific treatment of arm and gripper actions is well motivated and empirically supported. Table 12 shows that applying the DCT projection inside DMD training to arm channels only outperforms inference-time filtering, output smoothing, and all-channel projection, both in success rate and in gripper temporal error. This supports the claim that the constraint must be built into distillation rather than applied as post-processing. The method also adds only 0.028M parameters and keeps a single network evaluation at deployment.

Critical weaknesses
  1. The quantitative results are internally inconsistent and the reported gains are small relative to the likely variance. The same configuration, OneRobo with SmolVLA on LIBERO-Spatial, is reported as 75.10% in Table 3 and 81.6% in Table 5, and the arm-only projected DMD variant is reported as 79.80% in Table 12. In Table 3, OneRobo is the weakest one-step method on this suite. The 6.5-point discrepancy exceeds every effect claimed in the component ablation (+0.2, −0.6, and +1.2 for individual components). In addition, the LIBERO standard deviations come from re-evaluating a single checkpoint with different seeds rather than from independent training runs, RoboTwin results have no variance estimates, and the π0.5 LIBERO setting is saturated (the undistilled one-step model already achieves 96.62% against 96.70%). Under these conditions it is difficult to establish that OneRobo reliably outperforms the strongest baseline, OneDP.

  2. The central claim of improved temporal fidelity is not measured where it matters most. Smoothness metrics appear only in Table 1, whose backbone, benchmark, and task set are not specified, and in Table 12. Table 12 itself shows that the configuration with the lowest reversal rates does not achieve the highest success. No reversal or continuity metrics are reported on RoboTwin or on the real robot, and the inter-chunk discontinuity that the history component is designed to reduce is never quantified. Combined with the fact that the individual components closely follow prior work (frequency-domain constraints as in FreqPolicy, previous-chunk conditioning as in RTC and OFP, and DMD2), this leaves the contribution largely dependent on empirical evidence that is currently not conclusive.

Questions for authors

Why is the same configuration (OneRobo, SmolVLA, LIBERO-Spatial) reported as 75.10% in Table 3, 81.6% in Table 5, and 79.80% in Table 12? Please specify the checkpoint, evaluation seeds, and protocol used for each table.

Which backbone, benchmark, and tasks are used in Table 1? Please give a precise definition of the direction-reversal rate (formula, threshold, and averaging over joints), and explain why Table 1 and Table 12 report reversal rates on very different scales.

Can you report RoboTwin and ablation results over multiple independent training runs, so that the 1 to 2 point differences between OneRobo and OneDP can be assessed?

The baselines use different numbers of student updates, learning rates, and optimizers (Tables 7 and 8). Were they tuned to a comparable degree? Why does OFP reach only 65.62% on π0.5 LIBERO when all other methods are near 96.6%?

Were the real-world trials for all methods run under the same scene and lighting conditions? The OneRobo rollouts in Fig. 8 appear to have a different setup from the teacher and DMD rows.

Other than above, there are a few requests that could strengthen the paper:

  • Report smoothness and chunk-boundary discontinuity metrics on RoboTwin and the real robot.

  • Compare against training-time RTC for the history component, and against a two- or three-step teacher on the speed and success trade-off.

  • Include OneDP in the real-world comparison.

5 comments

We thank the reviewer for the detailed and constructive feedback. We appreciate the concerns regarding result consistency, statistical reliability, temporal fidelity, and experimental fairness. We address each point below.

W1 & Q1: Inconsistent quantitative results and insufficient statistical evidence

A1: We apologize for the reporting inconsistencies. The 75.10% reported for OneRobo on SmolVLA LIBERO-Spatial in Table 3 was erroneous. Moreover, the original Tables 3, 5, and 12 used different evaluation seeds, and Table 12 evaluated projection variants without history conditioning rather than the full OneRobo configuration. To ensure consistency, **we have consolidated and re-evaluated the results using the same three evaluation seeds **, including two seeds previously used in Tab.3 and one used in Tab.5 and Tab.12 experiments.

The corrected results are presented below.

Revised Table 3. Success rates (%) on LIBERO with SmolVLA, reported as mean ± standard deviation across three evaluation seeds.

MethodSpatialObjectGoalLIBERO-10Overall
Teacher (10-step)71.20 ± 1.4485.73 ± 0.9574.07 ± 1.8134.13 ± 1.5566.28 ± 0.28
Teacher (1-step)80.27 ± 0.7682.00 ± 1.4066.60 ± 0.2019.40 ± 0.9262.07 ± 0.13
OneDP78.13 ± 1.7082.80 ± 6.7670.87 ± 3.4128.20 ± 4.0465.00 ± 3.13
FreqPolicy79.80 ± 1.4084.73 ± 0.7066.27 ± 0.3123.93 ± 1.1063.68 ± 0.13
OneRobo79.93 ± 1.9487.33 ± 5.8471.13 ± 1.9429.67 ± 3.1867.02 ± 2.35

Revised Table 5. Component ablation on SmolVLA LIBERO-Spatial.

ConfigurationSuccess Rate (%) ↑
DMD77.53 ± 1.03
DMD + Arm Subspace78.00 ± 0.72
DMD + Gripper Supervision78.07 ± 1.33
DMD + History Conditioning79.13 ± 1.86
Full OneRobo79.93 ± 1.94

Revised Figure 6(b). Success rates (%) under different DCT truncation levels on LIBERO-Spatial.

DCT Modes (K)Success Rate (%) ↑
20.00 ± 0.00
422.73 ± 0.76
871.33 ± 0.70
1278.53 ± 0.31
1679.93 ± 1.94
2078.93 ± 1.50
3278.87 ± 0.70
3678.80 ± 1.11

Revised Table 12. Projection-placement and channel ablation on SmolVLA LIBERO-Spatial, with history conditioning disabled.

ConfigurationSuccess (%) ↑Translation Reversal (%) ↓Rotation Reversal (%) ↓Gripper D1 Error ↓
No DCT77.53 ± 1.031.41410.75110.14283
Post-hoc DCT75.67 ± 1.701.36400.63890.14548
Output-only Projection76.93 ± 1.291.35150.72520.14607
Noise + Output Projection76.80 ± 1.741.26880.53690.13869
Projected DMD (all-channel)77.40 ± 1.561.39650.51350.14022
Projected DMD (arm-only)78.00 ± 0.721.22680.50340.11503

Under the unified evaluation protocol, the full OneRobo configuration consistently achieves 79.93% ± 1.94% on LIBERO-Spatial in both Table 3 and Table 5. Table 12 separately evaluates projection mechanisms without history conditioning; its arm-only projected DMD result (78.00% ± 0.72%) is consistent with the corresponding component configuration in Table 5. The updated results also show that OneRobo achieves 67.02% average success on SmolVLA LIBERO, compared with 65.00% for OneDP and 63.68% for FreqPolicy.

To further assess robustness on RoboTwin, we additionally evaluate several one-step methods, including ManiCM, over three evaluation runs with matched environment seeds.

Table R4. Success rates (%) on five RoboTwin Clean tasks, reported as mean ± standard deviation across three evaluation runs.

TaskFastWAM T10FastWAM T1OneDPFreqPolicyOneRoboManiCM
Handover Mic99.3 ± 1.254.7 ± 8.382.0 ± 9.297.3 ± 1.294.0 ± 2.088.7 ± 3.1
Open Microwave76.7 ± 3.110.0 ± 2.072.7 ± 4.60.7 ± 1.267.3 ± 4.264.7 ± 3.1
Put Bottles Dustbin94.0 ± 2.013.3 ± 3.165.3 ± 13.364.7 ± 4.272.0 ± 9.269.3 ± 9.2
Scan Object95.3 ± 2.344.0 ± 6.982.7 ± 5.058.0 ± 5.387.3 ± 1.274.7 ± 6.4
Stack Bowls Three81.3 ± 6.152.7 ± 9.572.7 ± 7.660.0 ± 6.984.0 ± 5.377.3 ± 7.0
Avg.89.3234.9475.0856.1480.9274.94

On these five tasks, OneRobo achieves the highest average success among the evaluated one-step methods (80.92%), compared with 75.08% for OneDP and 74.94% for ManiCM. We further report independent-training-seed comparisons in W5.

W2 & Q2: Insufficient evidence connecting temporal fidelity to task success

A2: We agree that lower reversal rates do not necessarily imply higher task success. Our objective is to suppress unnecessary temporal oscillations while preserving meaningful action dynamics. We first clarify that the original Table 1 reports temporal metrics for FastWAM on the RoboTwin Stack Bowls Three task. Table 12 instead evaluates SmolVLA on LIBERO-Spatial and separately reports translation and rotation reversal rates. To provide additional evidence, we evaluate temporal consistency and chunk-boundary discontinuity on RoboTwin Open Microwave, reporting Arm D1, Arm D2, Direction Reversal Rate, and Boundary Jump. For the additional RoboTwin evaluation, Boundary Jump measures the mean absolute difference between the last executed arm command of one chunk and the first command of the next chunk, using H=32H=32 and E=24E=24.

Table R3. Temporal consistency and chunk-boundary discontinuity on RoboTwin Open Microwave (50 episodes per configuration).

MethodSuccess (%) ↑Arm D1 ↓Arm D2 ↓Reversal (%) ↓Boundary Jump ↓
FastWAM T10746.0015.65953.476.369
FastWAM T1826.54544.28967.3119.607
FastWAM T1 + DCT011.99512.80836.9023.565
Raw-DMD2012.78921.19670.5010.407
Raw-DMD + DCT526.8846.24835.1611.181
OneRobo w/o History605.7144.05132.2810.238
OneRobo645.5433.88331.748.204

Arm D1, Arm D2, and Boundary Jump are reported in units of ×10⁻³.

Compared with Raw-DMD, OneRobo reduces Arm D1 from 12.789 to 5.543, Arm D2 from 21.196 to 3.883, and Boundary Jump from 10.407 to 8.204, while improving task success from 20% to 64%. Importantly, history conditioning reduces Boundary Jump from 10.238 to 8.204 (19.9%) relative to OneRobo without history, providing direct quantitative evidence of improved continuity at action-chunk transitions. We also emphasize that temporal smoothness alone does not guarantee task success. For example, applying post-hoc DCT to FastWAM T1 substantially reduces reversal rates but does not improve success. This supports our motivation for incorporating temporal structure into distillation rather than relying solely on output smoothing.

Furthermore, our controlled perturbation experiments, detailed in the response to Reviewer z6Td, demonstrate that high-frequency and low-frequency action errors with nearly identical MSE can have different impacts on task success, particularly on precision-sensitive tasks.

We additionally provide comparative execution videos for Open Microwave and Adjust Bottle under matched evaluation seeds, together with two real-world manipulation tasks, in the supplementary onerobo.html file. These videos provide qualitative evidence of the action jitter and unstable execution observed in one-step DMD2.

W3 & Q3: Baseline training configurations and OFP performance

A3: We thank the reviewer for raising this concern. Within each benchmark, all methods use the same training dataset, effective batch size, and nominal outer-step budget. However, the learning rates, optimizers, and student update frequencies are method-specific, as detailed in Tables 7 and 8. We acknowledge that equal outer-step budgets do not necessarily imply identical optimization budgets. In particular, DMD-based methods alternate between distribution-estimator and student updates, whereas other methods employ different optimization schedules. Regarding OFP's relatively low performance on π0.5 LIBERO, we have rechecked our implementation and evaluation results, and the low performance persists under our adapted configuration. We followed the original OFP training procedure where applicable, while adapting it to the π0.5 action generator.

W4 & Q4: Consistency of real-world evaluation conditions

A4: We agree that consistent evaluation conditions are essential for a fair comparison. All real-world methods were evaluated under the same initial scene configuration and lighting conditions, with the same robotic platform and evaluation procedure.

W5 & Q5: Independent training runs and additional evaluations

A5: We appreciate the request for independent training runs. To complement the evaluation-seed results in Table R4, we independently retrained OneDP and OneRobo using two different training seeds (42 and 123) and evaluated both methods on the same five RoboTwin Clean tasks. Each trained checkpoint is evaluated the same 50 episodes per task.

Table R5. Success rates (%) across two independent training seeds on five RoboTwin Clean tasks.

TaskOneDP (42)OneDP (123)OneRobo (42)OneRobo (123)
Handover Mic74909698
Open Microwave70666468
Put Bottles Dustbin72546476
Scan Object88748886
Stack Bowls Three78869088
Avg. (5 tasks)76.474.080.483.2

Across these two independent training runs, OneRobo achieves 80.4% and 83.2% average success, compared with 76.4% and 74.0% for OneDP, respectively. This experiment also shows higher average success in both evaluated training runs.

Reviewer_Y3zUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

OneRobo distills generative robot policies into one-step policies by combining DCT-based arm-trajectory constraints, gripper-specific temporal supervision, and history-conditioned action generation.

Critical strengths

The experimental evaluation is extensive.

Critical weaknesses
  1. The manuscript requires more careful presentation. Figures contain crowded annotations (fig. 1), and some labels and references are inconsistent. For example, Figure 6(c) refers to (H−E)/H(H-E)/H in its title but to the history support/decay horizon elsewhere. Tables 2 and 3 are placed within the method derivation, before their discussion in the experimental section, which disrupts the reading flow.

  2. The novelty appears incremental. DCT-based action representations and frequency-aware policy learning are already well explored, including in FAST and FreqPolicy. Although the paper distinguishes its projected distillation formulation from frequency-consistency losses and post-processing, I remain unconvinced that the combination provides a sufficiently substantial conceptual advance.

  3. On LIBERO the success and efficiency advantages are both negligible.

  4. The evidence for improved inter-chunk continuity under realistic asynchronous execution remains limited. History-conditioning ablations and inference-latency measurements do not directly establish robustness to delayed observations, variable inference latency, or action-chunk transitions during continuous execution. Execution videos would also help assess the claimed reductions in jitter and improvements in continuity.

Questions for authors

please address the concerns above

4 comments

We thank the reviewer for the careful assessment and constructive suggestions. We address each concern below.

W1: Manuscript presentation and inconsistent notation

A1: We thank the reviewer for these helpful suggestions. We will (1) simplify Figure 1 to improve readability, (2) correct inconsistent figure labels and notation, and (3) move Tables 2 and 3 to the experimental section alongside their corresponding discussions.

Regarding Figure 6(c), we clarify that HH denotes the prediction horizon, EE the number of executed actions, and OO the history support/decay horizon. While (H−E)/H(H-E)/H represents the fraction of overlapping actions between consecutive chunks, O/HO/H specifies the normalized history decay horizon. These are distinct quantities. Figure 6(c) varies the history decay configuration rather than the execution horizon; therefore, we will correct its horizontal-axis label to History Decay Horizon Ratio (O/HO/H) and clarify the distinction in the caption and text.

W2: Incremental novelty relative to FAST and FreqPolicy

A2: We acknowledge that DMD and frequency-domain action modeling are established techniques. OneRobo does not claim a new frequency representation or a new generic DMD objective. Instead, its contribution lies in integrating temporal structure directly into one-step distribution matching through three complementary designs:

  1. Subspace-consistent distillation: Unlike FAST's frequency-based action tokenization or FreqPolicy's frequency-consistency regularization, OneRobo constrains noise, generated arm actions, distribution-estimator fitting, and DMD update directions within a shared trajectory subspace.

  2. Channel-specific dynamics: Arm trajectories are constrained to low-frequency components, while gripper transitions are preserved through teacher-guided supervision of action values and temporal differences.

  3. Cross-chunk continuity: Aligned historical actions condition both student generation and distribution estimation, jointly addressing intra-chunk dynamics and inter-chunk continuation.

Importantly, our experiments demonstrate benefits beyond simply combining DMD with DCT filtering. On RoboTwin Open Microwave (Table R3 in our additional experiments), post-hoc DCT improves DMD success from 20% to 52%, whereas the full OneRobo achieves 64%. Furthermore, the original Table 12 shows that projected DMD achieves 78.0% success, compared with 75.7% for inference-only DCT and 76.8% for noise-and-output projection.

These results support the effectiveness of incorporating temporal constraints into distribution-matching training rather than relying solely on output filtering. We will further clarify this distinction in the revised manuscript.

Table R3. Post-hoc DCT and OneRobo on RoboTwin Open Microwave (50 episodes; external RTC disabled).

MethodSuccess ↑Arm D1 ↓Arm D2 ↓Reversal ↓Boundary Jump (D1) ↓
FastWAM 10NFE74%6.0015.65953.47%6.369
FastWAM 1NFE8%26.54544.28967.31%19.607
FastWAM 1NFE + DCT0%11.99512.80836.90%23.565
Raw-DMD20%12.78921.19670.50%10.407
Raw-DMD + DCT52%6.8846.24835.16%11.181
OneRobo w/o History60%5.7144.05132.28%10.238
OneRobo64%5.5433.88331.74%8.204

W3: Limited success gains on LIBERO

A3: We agree that the performance gains on π0.5 LIBERO are limited, primarily because the benchmark is nearly saturated: direct one-step inference already achieves 96.62% success, compared with 96.70% for the 10-step teacher. SmolVLA provides a more challenging LIBERO setting, where direct step reduction causes greater performance degradation. We correct the inconsistent SmolVLA results in Table 3. Beyond LIBERO, OneRobo achieves 91.16% average success across 50 RoboTwin Clean tasks, compared with 89.16% for OneDP, while providing a 3.23× speedup over the 10-step teacher.

To further evaluate challenging tasks and strengthen the baseline comparisons, we additionally compare OneRobo with ManiCM and other one-step methods on five RoboTwin Clean tasks, using three evaluation runs with 50 episodes per task and matched environment seeds.

Table R4. Success rates (%) on five RoboTwin Clean tasks, reported as mean ± standard deviation across three evaluation runs.

TaskFastWAM T10FastWAM T1OneDPFreqPolicyOneRoboManiCM (5k)
Handover Mic99.3 ± 1.254.7 ± 8.382.0 ± 9.297.3 ± 1.294.0 ± 2.088.7 ± 3.1
Open Microwave76.7 ± 3.110.0 ± 2.072.7 ± 4.60.7 ± 1.267.3 ± 4.264.7 ± 3.1
Put Bottles Dustbin94.0 ± 2.013.3 ± 3.165.3 ± 13.364.7 ± 4.272.0 ± 9.269.3 ± 9.2
Scan Object95.3 ± 2.344.0 ± 6.982.7 ± 5.058.0 ± 5.387.3 ± 1.274.7 ± 6.4
Stack Bowls Three81.3 ± 6.152.7 ± 9.572.7 ± 7.660.0 ± 6.984.0 ± 5.377.3 ± 7.0
Avg.89.3234.9475.0856.1480.9274.94

OneRobo achieves the highest average success rate among the evaluated one-step methods (80.92%), outperforming ManiCM (74.94%) and OneDP (75.08%) by 5.98 and 5.84 percentage points, respectively. In particular, it achieves 84.0% success on Stack Bowls Three, compared with 77.3% for ManiCM and 72.7% for OneDP.

These results provide additional evidence of OneRobo's effectiveness on challenging manipulation tasks where direct one-step inference exhibits substantial performance degradation.

W4: Limited evidence for inter-chunk continuity

A4: We thank the reviewer for this valuable suggestion. We agree that history-conditioning ablations and inference-latency measurements alone do not directly establish inter-chunk continuity. Figure 6(c) examines the effect of different history configurations on task success. To provide more direct quantitative evidence, we additionally evaluate arm temporal metrics and chunk-boundary discontinuity on the RoboTwin Open Microwave task, using H=32H=32 and E=24E=24. We introduce Boundary Jump (D1), defined as the mean absolute difference between the last executed arm command of the previous chunk and the first executed arm command of the new chunk in the Tab.R3.

Compared with OneRobo without history conditioning, the full model reduces Boundary Jump from 10.238 to 8.204 (19.9%), while improving success from 60% to 64%. These results provide direct quantitative evidence that history conditioning can improve continuity at action-chunk transitions under the evaluated execution protocol. We also include comparative execution videos of two RoboTwin tasks (Open Microwave and Put Bottles Dustbin) under matched evaluation seeds, together with two real-world manipulation tasks, in the supplementary onerobo.html file. These videos illustrate the action jitter and unstable execution observed in one-step DMD2, which are reduced by OneRobo.

We thank the reviewer for the valuable feedback. We would appreciate any further thoughts on our responses and additional experiments, and would be happy to clarify any remaining concerns.

We sincerely thank you for your thoughtful review. We hope our responses and additional experiments have addressed your concerns. If you have any further questions or comments, we would be happy to discuss them.

Reviewer_z6TdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces the OneRobo framework, which aims to address robotic arm jitter and discontinuities between action chunks that arise after the distalation of flow-matching-based robotic policies (VLA/WAM).It builds on the DMD2 framework by adding three new modules: DCT-based low-frequency subspace truncation for the arm channels, multi-order difference supervision from the teacher for the gripper, and history conditioning on the unexecuted actions from the previous chunk.

Critical strengths
  1. Distilling the action heads of VLA and WAM remains practically valuable in the current setting, as it can effectively improve execution efficiency and reduce latency.
  2. The experiments are extensive in scope, including three backbones (FastWAM, π0.5, and SmolVLA), all 50 tasks in RoboTwin, and real-world robot experiments.
  3. For the RoboTwin benchmark, OneRobo yields moderate performance gains over the baselines.
Critical weaknesses
  1. The explanation and analysis of the research question are insufficient. The jitter and rotation flipping observed at inference time in one-step generative models obtained via distillation or shortcut training do exist. However, the authors provide no evidence of how this issue actually affects the success rate of downstream robotic tasks. The failures of DMD may also stem from mode collapse or out-of-distribution issues. More critically, the simulation results show that, in most experiments, the success rate appears largely uncorrelated with temporal accuracy.
  2. DCT truncation can effectively remove noise and restrict the student to learning only low-frequency information, but it also artificially discards meaningful high-frequency components. In this sense, it is not substantially different from applying post-hoc low-pass filtering.
  3. The proposed OneRobo model offers limited novelty compared with DMD and FreqPolicy, as it largely appears to be a combination of distribution matching distillation and frequency-domain truncation.
  4. The experiments lack comparisons with mainstream consistency distillation methods such as Consistency Policy and ManiCM.
  5. Table 1 lacks specific details about the experimental setup, and the axes in Figure 2 are not labeled.
Questions for authors
  1. The authors should provide more thorough experiments to establish a causal relationship between the temporal metrics and task success rate. Without such evidence, there remains a gap between the design motivation of the method and the task objective.
  2. Comparing only against DMD-family baselines is not sufficient. It would be valuable to include other one-step generative approaches, such as consistency models.
5 comments

We thank the reviewer for the constructive assessment and thoughtful suggestions. We address each concern below.

W1 & Q1: Relationship between temporal fidelity and task success

A1: We agree that temporal metrics alone do not establish their impact on task success. To investigate this relationship, we conduct controlled perturbation experiments using the same FastWAM 10-NFE teacher on two RoboTwin tasks.

Specifically, we inject frequency-specific perturbations into the decoded absolute joint targets of the teacher:

Ut,j′=UT,t,j+ασjϵt,j,U'_{t,j}=U_{T,t,j}+\alpha\sigma_j\epsilon_{t,j},

\frac{ w_t\sum_{k\in\mathcal K_f}B_{t,k}c_{k,j} }{ \sqrt{ \frac{1}{24}\sum_{\tau=0}^{23} \left(w_\tau\sum_{k\in\mathcal K_f}B_{\tau,k}c_{k,j}\right)^2 } }.$$ where $\alpha\in\{0.5,1,2\}$ controls perturbation strength, $c_{k,j}\sim\mathcal N(0,1)$, and $B$ is the orthonormal DCT-II basis with $H=32$. We use $\mathcal K_{\mathrm{low}}=\{1,\ldots,7\}$ and $\mathcal K_{\mathrm{high}}=\{16,\ldots,31\}$. The joint-wise scale $\sigma_j$ is estimated from a fixed set of teacher trajectories and shared across tasks. Only the 12 arm channels are perturbed; gripper commands remain unchanged. By normalizing each perturbation over the 24-step execution horizon, both frequency groups have identical perturbation energy before joint-limit clipping: $$\mathrm{MSE}= \frac{\alpha^2}{12}\sum_{j=1}^{12}\sigma_j^2.$$ The measured MSE differences between paired low- and high-frequency settings remain approximately 0.5% or less. **As shown in Tables R1 and R2, the effect of temporal perturbations is task-dependent.** On Open Microwave, at nearly identical perturbation MSE (0.7565 vs. 0.7570), high-frequency perturbations reduce success to 24%, compared with 50% under low-frequency perturbations. In contrast, Adjust Bottle remains relatively robust, achieving 98% and 100% success, respectively. These results provide controlled evidence that temporal error structure can affect task success even at comparable action MSE, supporting our motivation for temporally structured distillation. We additionally provide comparative simulation and real-world execution videos in the supplementary `onerobo.html` file to illustrate the observed execution instability. **Table R1. Controlled perturbations on RoboTwin Open Microwave (50 episodes per configuration).** | Configuration | Success ↑ | Perturbation MSE | Arm D1 ↓ | Arm D2 ↓ | Reversal ↓ | |:--|--:|--:|--:|--:|--:| | FastWAM T10 | 74% (37/50) | 0 | 6.001 | 5.659 | 53.47% | | Low 0.5 | 70% (35/50) | 0.1887 | 6.632 | 6.060 | 43.28% | | High 0.5 | 74% (37/50) | 0.1883 | 10.156 | 14.105 | 62.83% | | Low 1.0 | 50% (25/50) | 0.7565 | 7.627 | 6.822 | 40.36% | | High 1.0 | 24% (12/50) | 0.7570 | 15.053 | 25.235 | 66.64% | | Low 2.0 | 20% (10/50) | 3.0514 | 11.086 | 8.809 | 35.26% | | High 2.0 | 14% (7/50) | 3.0404 | 27.385 | 47.985 | 68.04% | **Table R2. Controlled perturbations on RoboTwin Adjust Bottle (50 episodes per configuration).** | Configuration | Success ↑ | Perturbation MSE | Arm D1 ↓ | Arm D2 ↓ | Reversal ↓ | |:--|--:|--:|--:|--:|--:| | FastWAM T10 | 100% (50/50) | 0 | 17.142 | 6.657 | 39.04% | | Low 0.5 | 100% (50/50) | 0.1945 | 17.868 | 7.018 | 34.91% | | High 0.5 | 100% (50/50) | 0.1942 | 21.004 | 15.207 | 50.87% | | Low 1.0 | 100% (50/50) | 0.7742 | 18.991 | 7.558 | 30.67% | | High 1.0 | 98% (49/50) | 0.7782 | 25.864 | 26.273 | 53.34% | | Low 2.0 | 100% (50/50) | 3.0967 | 21.527 | 9.121 | 26.58% | | High 2.0 | 92% (46/50) | 3.0928 | 36.484 | 49.251 | 56.69% |

W2: DCT truncation versus post-hoc low-pass filtering

A2: We agree that DCT truncation can remove task-relevant high-frequency components; this concern motivates integrating a trajectory-subspace constraint throughout DMD training, rather than applying DCT only after generation. In OneRobo, the arm-action noise, generated actions, distribution-estimator training targets, and DMD update directions are consistently constrained to the same subspace. Appendix A.1.1 characterizes this subspace property. And Table 12 shows that projected DMD (arm-only) achieves 78.0% success, versus 75.7% for inference-only DCT and 76.8% for noise-and-output projection. We further evaluate FastWAM on RoboTwin Open Microwave:

Table R3. Post-hoc DCT and OneRobo on RoboTwin Open Microwave (50 episodes; external RTC disabled).

MethodSuccess ↑Arm D1 ↓Arm D2 ↓Reversal ↓Boundary Jump (D1) ↓
FastWAM 10NFE74%6.0015.65953.47%6.369
FastWAM 1NFE8%26.54544.28967.31%19.607
FastWAM 1NFE + DCT0%11.99512.80836.90%23.565
Raw-DMD20%12.78921.19670.50%10.407
Raw-DMD + DCT52%6.8846.24835.16%11.181
OneRobo w/o History60%5.7144.05132.28%10.238
OneRobo64%5.5433.88331.74%8.204

Post-hoc DCT improves Raw-DMD success from 20% to 52%, showing that filtering can help, but remains below the full OneRobo (64%). Conversely, directly filtering FastWAM 1NFE reduces its reversal rate without improving success (8% to 0%). Thus, smoother outputs alone do not guarantee better task execution. Since the full OneRobo also contains gripper supervision and history conditioning, we use the controlled projection-placement ablations in the Table 12 to isolate the role of training-time projection.

W3: Novelty relative to DMD and FreqPolicy

A3: We acknowledge that DMD and frequency-domain action modeling are established techniques. OneRobo does not claim a new DCT basis or a new generic DMD objective. Its contribution is a temporally structured, channel-aware formulation of one-step distribution matching:

  1. Subspace-consistent distillation: Unlike post-hoc filtering or frequency-consistency regularization, OneRobo constrains projected noise, arm-action generation, distribution-estimator fitting, and DMD update directions within a shared trajectory subspace (Table 12 and Table R3).
  2. Channel-specific dynamics: Arm trajectories are constrained to low-frequency components, while native gripper actions are supervised using teacher-guided values and first- and second-order temporal differences to retain task-relevant transitions.
  3. Cross-chunk continuation: Aligned unexecuted actions from the previous chunk condition both the generator and the distribution estimator through history-aware initialization and learned embeddings.

Importantly, our experiments demonstrate benefits beyond simply combining DMD with DCT filtering. As shown in Table R3, applying post-hoc DCT to DMD improves success from 20% to 52% on RoboTwin Open Microwave, whereas OneRobo achieves 64%. Similarly, Table 12 shows that projected DMD achieves 78.0% success, outperforming inference-only DCT (75.7%) and noise-and-output projection (76.8%).

These results support the effectiveness of integrating temporal constraints into the distillation process rather than relying solely on output filtering. We will further clarify this distinction in the revised manuscript.

W4 & Q2: Lack of comparisons with Consistency Policy and ManiCM

A4: We thank the reviewer for this suggestion. Our original Tables 2 and 3 already include FreqPolicy, which employs frequency-domain consistency learning rather than DMD. To further strengthen the comparison, we additionally evaluate ManiCM alongside other one-step baselines on five RoboTwin Clean tasks. We conduct three evaluation runs, each using 50 episodes per task, with matched environment seeds across methods.

Table R4. Success rates (%) on five RoboTwin Clean tasks. Results are reported as mean ± standard deviation across three evaluation seed runs.

TaskFastWAM T10FastWAM T1OneDPFreqPolicyOneRoboManiCM
Handover Mic99.3 ± 1.254.7 ± 8.382.0 ± 9.297.3 ± 1.294.0 ± 2.088.7 ± 3.1
Open Microwave76.7 ± 3.110.0 ± 2.072.7 ± 4.60.7 ± 1.267.3 ± 4.264.7 ± 3.1
Put Bottles Dustbin94.0 ± 2.013.3 ± 3.165.3 ± 13.364.7 ± 4.272.0 ± 9.269.3 ± 9.2
Scan Object95.3 ± 2.344.0 ± 6.982.7 ± 5.058.0 ± 5.387.3 ± 1.274.7 ± 6.4
Stack Bowls Three81.3 ± 6.152.7 ± 9.572.7 ± 7.660.0 ± 6.984.0 ± 5.377.3 ± 7.0
Avg.89.3234.9475.0856.1480.9274.94

As shown in Table R4, OneRobo achieves the highest average success rate among the evaluated one-step methods (80.92%), outperforming ManiCM (74.94%) and OneDP (75.08%) by 5.98 and 5.84 percentage points, respectively.

In particular, OneRobo achieves 84.0% success on Stack Bowls Three, compared with 77.3% for ManiCM and 72.7% for OneDP. It also achieves the best one-step performance on Put Bottles Dustbin and Scan Object.

These additional results strengthen our empirical comparison beyond DMD-family methods and demonstrate the effectiveness of OneRobo on challenging multi-stage manipulation tasks.

W5: Missing experimental details and axis labels

A5: Thank you for pointing out these presentation issues. We will specify in the main-paper Table 1 caption that the reported temporal metrics are evaluated using FastWAM on the RoboTwin Stack Bowls Three task. We will also add the axes to Fig. 2: Action Step (horizontal) and Joint action (vertical).

Comments on the paper

4 comments
~Olivia_Brown3OpenReview ↗

Distribution gap in History Conditioning (Training vs. Deployment)

We are particularly interested in the history-conditioned generation. We noticed in Appendix A.1.4 (Eq. 40) that during training, the historical plan A−A^- is bootstrapped using a detached, single-step generation of the student. However, during deployment, the history is formed by recursively generated and partially executed plans.

We are curious: in your experience, does this gap between the training-time prior and the deployment-time prior lead to noticeable error accumulation or drift over long-horizon rollouts?We would love to know if you observed any stability issues related to this, or if you have any strategies to mitigate this during training (e.g., using ground-truth actions or teacher plans for bootstrapping).

We thank you for your interest in our history-conditioned generation and for raising this important question. During training, we construct the historical prior using a detached student prediction under the previous observation, without history conditioning. During deployment, the prior is instead obtained from previously generated and partially executed action chunks. To reduce sensitivity to potentially inaccurate historical predictions, we adopt two designs. First, 10% history dropout encourages the student to generate actions without relying exclusively on historical predictions. Second, a short, linearly decaying, arm-only history mask limits the influence of previous predictions, allowing current observations to guide new action generation. Our additional RoboTwin experiments provide evidence of improved local continuity. On Open Microwave, history conditioning reduces the chunk-boundary discontinuity from 10.238 to 8.204 (19.9%) and improves success from 60% to 64%. These results suggest that history conditioning improves continuity at chunk transitions. We agree that explicitly evaluating recursive-history drift would be a valuable direction for further investigation.

Thank you for your reply.

We sincerely thank the Area Chair and all reviewers for their careful evaluation and constructive feedback. The discussion has helped us clarify the methodological contributions of OneRobo, strengthen the experimental evidence, and improve the consistency and reproducibility of our results.

We would like to summarize the main contributions and the additional evidence provided during the rebuttal.

1. Main Contribution: Temporally Structured Distribution Matching

We acknowledge that DMD, DCT-based representations, and history conditioning are established techniques. The central contribution of OneRobo is to incorporate execution-relevant temporal structure directly into one-step distribution matching, rather than applying frequency-domain filtering as a post-processing operation.

Specifically, OneRobo consistently constrains projected noise, generated arm actions, distribution-estimator fitting, and DMD update directions within a shared trajectory subspace. This differs from conventional DMD in the original action space and from frequency-domain post-processing or auxiliary consistency regularization. We further distinguish arm and gripper dynamics through channel-specific supervision and incorporate aligned historical actions into both student generation and distribution estimation.

2. Additional Evidence Addressing the Main Concerns

Based on the reviewers' suggestions, we conducted several additional experiments:

  • Temporal fidelity and task success (Tables R1–R2). Controlled perturbations demonstrate that action errors with nearly identical MSE can have different effects on task success. On RoboTwin Open Microwave, high-frequency perturbations reduce success to 24%, compared with 50% under low-frequency perturbations at nearly identical MSE (0.7570 vs. 0.7565). Adjust Bottle is comparatively robust, indicating that the effect is task-dependent. These results strengthen the motivation for preserving temporal action structure.

  • Beyond post-hoc DCT filtering (Table R3 and Table 12). On Open Microwave, applying DCT filtering to Raw-DMD improves success from 20% to 52%, whereas full OneRobo achieves 64%. More importantly, the controlled projection-placement ablation in Table 12, with history conditioning disabled, shows that arm-only projected DMD achieves 78.00%, compared with 75.67% for inference-only DCT and 76.80% for noise-and-output projection. These results support incorporating trajectory-subspace constraints during distillation rather than relying solely on post-hoc filtering.

  • Inter-chunk continuity (Table R3). We introduce an executed-action chunk-boundary metric to directly quantify discontinuities during replanning. History conditioning reduces Boundary Jump from 10.238 to 8.204 (19.9%) while improving Open Microwave success from 60% to 64%. This provides quantitative evidence of improved local continuity.

  • Baseline coverage and statistical reliability (Tables R4–R5). We add ManiCM and evaluate six methods on five RoboTwin Clean tasks using three matched evaluation seeds. OneRobo achieves 80.92% average success, compared with 75.08% for OneDP and 74.94% for ManiCM. We additionally conduct two independent training runs for OneRobo and OneDP. OneRobo achieves 80.4% and 83.2% average success, compared with 76.4% and 74.0% for OneDP, respectively, providing preliminary evidence of robustness across training seeds.

3. Result Consistency and Experimental Scope

We also addressed the reported inconsistencies in SmolVLA LIBERO results by consolidating evaluations under the same three evaluation seeds. The corrected full OneRobo configuration achieves 79.93% ± 1.94% on LIBERO-Spatial in both Tables 3 and 5. Table 12 evaluates projection variants without history conditioning, explaining the difference from the full configuration.

Across all 50 RoboTwin Clean tasks, OneRobo achieves 91.16% average success, compared with 92.80% for the ten-step teacher and 89.16% for OneDP, while providing a 3.23× inference speedup. On SmolVLA LIBERO, OneRobo achieves 67.02% average success, compared with 65.00% for OneDP. We acknowledge that gains on the nearly saturated π0.5 LIBERO benchmark are limited and that the additional independent-training comparisons cover five rather than all 50 RoboTwin tasks.

We have incorporated the additional experiments, corrected results, and detailed temporal-metric definitions into the revised manuscript and appendix. Supplementary execution videos provide further qualitative comparisons.

We appreciate the reviewers' continued engagement. We hope these clarifications help distinguish OneRobo's training-time formulation from a straightforward combination of existing techniques and provide a clearer basis for assessing its methodological and empirical contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.