- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Marginal-Aligned Warm Start (MAWS), a training-free method for reducing the denoising steps of flow-matching VLA action heads to a single step. The authors study why one-step inference works on some tasks but fails on others, and argue that persistent prediction errors can accumulate during closed-loop execution. MAWS addresses this by shifting the previous action chunk, adding noise according to the flow-matching schedule, and using it to initialize the next prediction.
Critical strengths
- The paper is easy to follow, with clear and well-designed figures and tables.
- The proposed method makes sense and is simple and effective.
- The experiments are quite comprehensive, covering several simulation benchmarks, and real-world robotic tasks.
Critical weaknesses
-
The novelty of the proposed method is relatively limited. Using the previous action chunk as a warm start is not a new idea. Similar approaches have already been explored in previous works, including RTI-DP and Falcon. The main difference here is re-noising the previous chunk according to the flow-matching schedule. While this is a reasonable design, I am not sure it represents a substantial methodological contribution.
-
The experiments do not fully support the proposed motivation. The authors argue that MAWS works by reducing persistent residual errors during closed-loop execution. Although Section 3.2 provide some evidence, I am not convinced that this is the main reason for the performance improvements. Reusing the previous action chunk naturally introduces temporal smoothing, which could also explain the results. I would like to see a more direct comparison with simple temporal ensembling to better understand what the proposed marginal alignment actually contributes.
-
The practical acceleration is less impressive than the action-head numbers suggest. Although MAWS reduces the action-head computation from 10 steps to 1, the end-to-end latency in the real-world experiments only decreases from 67.9 ms to 54.7 ms. This is a relatively modest improvement compared to the reported 9.1× action-head speedup. While it helps meet the 15 Hz control deadline in the authors' setup, I am not sure how much additional benefit it would provide in practical deployments. In particular, techniques such as real-time chunking (RTC) can already overlap inference with action execution and hide much of the inference latency. I would like to see a comparison with RTC to understand whether MAWS offers a meaningful advantage in real-time control.
Questions for authors
Please see weaknesses
6 comments
Response to Reviewer 9x4s (2/4)
W2. Temporal smoothing: the gain is not an averaging effect
We ran the comparison the reviewer asked for on π0 Fractal (official protocol, 3 seeds, category-weighted average). The controls give smoothing the same chunk and the same weight as MAWS, applied after the head instead of inside the step, so whatever they recover is smoothing alone:
- TE: the temporal ensembling of ACT [1], which averages the predictions that overlapping chunks make for the same timestep. With H = 4 and r = 2 every timestep is predicted by exactly two chunks, so the executed chunk is (1−λ)·new + λ·previous over the overlapping positions, λ = 0.5 (ACT's near-uniform weight, and the retention s of MAWS).
- TE-EMA: the same over all positions with the MAWS prior (shifted chunk, repeated tail), carried forward, i.e. an exponential moving average over chunks.
| π0 Fractal, SR (%) | NFE | Coke | Move | Drawer | Apple | Avg |
|---|---|---|---|---|---|---|
| from noise (paper) | 10 | 97.3 | 77.4 | 66.4 | 42.6 | 77.2 |
| from noise (paper) | 1 | 97.3 | 78.2 | 64.0 | 16.4 | 73.6 |
| from noise + TE | 1 | 97.1 | 67.6 | 59.3 | 22.8 | 70.2 |
| from noise + TE-EMA | 1 | 93.1 | 69.0 | 58.0 | 18.5 | 68.3 |
| +MAWS (paper) | 1 | 99.0 | 82.4 | 67.7 | 25.6 | 77.4 |
Smoothing the one-step output lowers the average below one step from noise (70.2 and 68.3 vs 73.6), costing 9–11 points on Move Near and 5–6 on Drawer, while MAWS gains on every category with the same chunk and the same weight. Three observations separate the mechanisms:
- MAWS moves the output toward the full-step answer, not toward the prior: the prior sits 0.08–0.25 from the posterior mean, farther than the one-step output, yet the step taken from it lands closer than either (residual lowered by 17–35% on every task, Table 3). Averaging with the prior cannot move the output away from the prior.
- The executed actions are not what carries the gain: with r = 2 of H = 4, every tail fill of Sec. 5.4 (Fig. 5) shares the same executed positions and differs only in the two unexecuted tail positions, yet zeros there return success to the from-noise level and any fill that continues the motion gains 5–9 points.
- A smoothing effect would be monotone in how much of the prior is kept, whereas the sweep above peaks exactly at the training marginal.
Our sentence describing the loop as "the temporal ensemble of ACT inside the sampler" (Sec. 4.2) invites the reviewer's reading. The analogy was to the memory over chunks, not to the mechanism, and we will reword it.
[1] T. Z. Zhao, V. Kumar, S. Levine, C. Finn. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. RSS 2023.
Response to Reviewer 9x4s (3/4)
W3a. End-to-end latency: the head's share of the call sets the ceiling
The reviewer is right that 67.9 → 54.7 ms is modest next to the 9.1× head speedup. A VLA call has two stages, the VLM backbone and the action head, and MAWS acts on the head, so its end-to-end gain is capped at call / backbone. On the real-robot π0.5 (RTX 5090) the backbone is 39.7 of the 67.9 ms, so no head-side method can exceed 1.35×. MAWS removes 90% of the head (17.6 → 1.8 ms) and reaches 1.24×. The cap binds backbone acceleration just as hard: on π0 the head is 82% of the call, so an infinitely fast backbone would give at most 1.22×.
| Model (default steps from noise) | head share | head: default → MAWS, 1 step (ms) | E2E: default → MAWS, 1 step (ms) | E2E speedup |
|---|---|---|---|---|
| π0 (10) | 82% | 322.5 → 33.8 | 394.5 → 108.7 | 3.6× |
| Fast-WAM (10) | 83% | 516 → 54 | 622 → 164 | 3.8× |
| VLANeXt-B (10) | 79% | 372 → 43 | 470 → 142 | 3.3× |
| GR00T N1.7 (4) | 62% | 91.9 → 25.1 | 147.4 → 80.2 | 1.8× |
| π0.5 (10) | 31% | 37.0 → 6.0 | 120.8 → 89.8 | 1.3× |
| π0.5, real robot, RTX 5090 (10) | 26% | 17.6 → 1.8 | 67.9 → 54.7 | 1.2× |
Batch 1, isolated, A40 unless noted. Fast-WAM (a world action model, Wan2.2-5B video DiT with a flow-matching action expert) and VLANeXt-B (Qwen3-VL-2B with a flow head) were added for this rebuttal, with released checkpoints and the paper's rule for s. The end-to-end gain tracks the head's share: 3.3–3.8× where the head dominates, 1.2–1.3× where the backbone does. The real-robot number is the smallest in the paper because π0.5 has the smallest head share, not because MAWS does less there. The two sides compose. We combined MAWS with visual-token pruning of the backbone, SAFE-Pruner [2], on π0.5 / LIBERO (4-suite average, 500 episodes per suite, pruned rows including the periodic full-token keyframes):
| backbone | head | SR (%) | call (ms) | speedup |
|---|---|---|---|---|
| dense | from noise, 10 steps | 95.2 | 131 | 1.00× |
| pruned | from noise, 10 steps | 94.8 | 113 | 1.16× |
| dense | MAWS, 1 step | 97.8 | 85 | 1.54× |
| pruned | MAWS, 1 step | 96.3 | 71 | 1.84× |
Either side alone stalls on the other stage, together they nearly halve the call with success above the dense baseline. In synchronous control the robot idles for the whole call at every replan, so the saving reaches the task: the same trajectory finishes about 1.37× sooner on π0 Fractal, and the completion time of successful Fast-WAM LIBERO episodes drops from 11.4 to 6.6 s (Spatial) and 27.0 to 15.9 s (Long).
[2] S. Ma, C. Zhang, C. Wang, Y. Wang, Y. Wu, Z. Wang, J. Tian, Z. Zhu, Y. Tang. SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation. ECCV 2026. arXiv:2605.29662.
Response to Reviewer 9x4s (1/4)
We thank the reviewer for the careful reading and address the three concerns in turn, each with new experiments. Numbers taken from the paper are marked as such.
W1. Novelty: what MAWS contributes beyond chunk reuse
The novelty is not that the previous chunk is reused, which the paper credits to prior work (Sec. 2.3), but what turns that chunk into a valid starting point for a single step. MAWS contributes a diagnosis of when and why one step fails in closed loop, and the marginal constraint under which one step from the previous chunk suffices. The prior-chunk methods do not use that constraint, and without it the same chunk is the worst option in Table 2.
Diagnosis. One step from noise fails on some tasks and not on others. The split is invisible in open loop, where the one-step residual is small on every task (Sec. 3.1), and is decided in closed loop, where the residual persists and integrates in the state until it exceeds the task's tolerance (Sec. 3.2). A warm start has to lower that persistent residual, not smooth the output.
Constraint. A flow head is trained on the marginal x_s = s·a + (1−s)·ε, so at flow time s its input must be the answer plus noise of width exactly 1−s. A clean previous chunk is not on that marginal, and the prior-chunk methods pay for this in steps (Falcon needs 7.6 evaluations per call to reach 75.6 on π0 Fractal, one-step MAWS 77.4, Table 2). Re-noising has two ingredients, a prior that is a sample of the policy's own action distribution and noise of the width the marginal prescribes. Two new experiments remove one at a time.
Removing the width. We keep the prior and s fixed and scale only the noise, x = s·g + c·(1−s)·ε, on the two Fractal tasks where the warm start gains (π0, s = 0.5, one step, 3 seeds, 324 episodes per cell). If the gain came from the chunk itself, less noise would keep more of it. If it comes from meeting the marginal, success peaks at c = 1.
Success peaks at c = 1 on both tasks. Under-noised, the same prior falls back to the level of one step from noise, and with no noise at all it is the RTI-DP-style collapse of Table 2 (9.9 and 18.5). Over-noised, it gives part of the gain back. The gain comes from re-noising the chunk to the marginal, not from the chunk as such.
Removing the sample. A prior built from the current robot state, re-noised with the same s, has the right width but is not a sample of the action distribution. It never beats one step from noise (π0 Fractal 72.6 vs 73.6, ALOHA 2.0 vs 22.0) and collapses where it covers the whole chunk (CALVIN, r = H: 6.9 vs 25.1, with MAWS at 40.7). All 3 seeds.
Re-noising is a one-line operation, the shifted chunk entering the flow at s with fresh noise of width 1−s instead of entering clean, but it is not a detail. The same shifted chunk gives 9.9 without it and 77.4 with it at one evaluation per call, level with ten evaluations from noise (77.2), while the prior-chunk methods that skip it need several evaluations to come within two points (Table 2). To our knowledge none of them identifies the constraint, and it is what turns chunk reuse into a one-step sampler.