- 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
-
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.
-
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
-
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.
-
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
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 and .
Table R3. Temporal consistency and chunk-boundary discontinuity on RoboTwin Open Microwave (50 episodes per configuration).
| Method | Success (%) ↑ | Arm D1 ↓ | Arm D2 ↓ | Reversal (%) ↓ | Boundary Jump ↓ |
|---|---|---|---|---|---|
| FastWAM T10 | 74 | 6.001 | 5.659 | 53.47 | 6.369 |
| FastWAM T1 | 8 | 26.545 | 44.289 | 67.31 | 19.607 |
| FastWAM T1 + DCT | 0 | 11.995 | 12.808 | 36.90 | 23.565 |
| Raw-DMD | 20 | 12.789 | 21.196 | 70.50 | 10.407 |
| Raw-DMD + DCT | 52 | 6.884 | 6.248 | 35.16 | 11.181 |
| OneRobo w/o History | 60 | 5.714 | 4.051 | 32.28 | 10.238 |
| OneRobo | 64 | 5.543 | 3.883 | 31.74 | 8.204 |
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.
| Task | OneDP (42) | OneDP (123) | OneRobo (42) | OneRobo (123) |
|---|---|---|---|---|
| Handover Mic | 74 | 90 | 96 | 98 |
| Open Microwave | 70 | 66 | 64 | 68 |
| Put Bottles Dustbin | 72 | 54 | 64 | 76 |
| Scan Object | 88 | 74 | 88 | 86 |
| Stack Bowls Three | 78 | 86 | 90 | 88 |
| Avg. (5 tasks) | 76.4 | 74.0 | 80.4 | 83.2 |
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.
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.
Revised Table 5. Component ablation on SmolVLA LIBERO-Spatial.
Revised Figure 6(b). Success rates (%) under different DCT truncation levels on LIBERO-Spatial.
Revised Table 12. Projection-placement and channel ablation on SmolVLA LIBERO-Spatial, with history conditioning disabled.
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.
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.