acceptodds
Under review as a conference paper at ICLR 2027

Quantile Head for Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models integrate pretrained Vision-Language Models (VLMs) with action heads for robot control. Common action heads have distinct limitations: point regression provides only a point estimate of the action distribution, while standard flow-matching samplers require costly iterative sampling. To address these limitations, we unify regression and flow matching under a shared objective and extend it to derive a quantile objective. This quantile objective guides the design of our Quantile Head, which predicts a median and positive gaps to form ordered marginal action quantiles in one forward pass. These quantiles support multiple sampling strategies without retraining and are jointly supervised to train the default median policy. Our local analysis of this joint supervision shows that, with calibrated nearby quantiles, fixed gaps, and matched correction speed, direct median updates have lower variance than under median-only supervision. Experiments show that this jointly supervised median policy achieves the highest average success rates among the compared methods on LIBERO, LIBERO-Plus, LIBERO-Pro, and two real-robot tasks, together with the shortest mean episode time among matched LIBERO baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Imitation Learning, Robot Manipulation, Action Distribution Modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_czhFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Quantile Head for vision-language-action models, which predicts ordered marginal action quantiles in a single forward pass. The method uses a median-centered representation with positive gaps and trains the quantiles jointly through a masked pinball loss. The median is used for default action decoding, while the additional quantiles enable optional sampling. The authors also provide a theoretical analysis of joint quantile supervision. Experiments on LIBERO, LIBERO-Plus, LIBERO-Pro, and real-robot tasks demonstrate promising performance.

Critical strengths
  1. The paper explores an interesting and practical direction for VLA action heads, combining efficient single-pass action prediction with quantile-based distribution modeling. The overall method is simple and well motivated.

  2. The experimental results are strong across multiple benchmarks. In particular, the matched ablations show that joint quantile supervision improves the median policy over median-only training. The real-robot experiments further support the practical effectiveness of the approach.

Critical weaknesses
  1. Unclear source of improvement from joint quantile supervision. Although jointly supervising multiple quantiles improves the median policy, it remains unclear whether the gains come from reduced gradient variance or more general auxiliary supervision. Theorem 2 assumes calibrated quantiles and fixed gaps, while the actual model learns these gaps jointly. The connection between the theoretical analysis and the empirical improvements could be better established.

  2. Limited modeling of joint action distributions. The method predicts marginal quantiles for individual action coordinates without explicitly modeling their joint distribution across dimensions and time steps. This may limit its ability to capture coordinated multimodal action trajectories. The obstacle-avoidance experiments in Figure 6 also show limited success with quantile sampling, raising questions about the practical advantages of the learned distribution beyond median prediction.

Questions for authors
  1. Can the authors provide further evidence that joint quantile supervision reduces the variance of median updates in practice, rather than simply acting as additional auxiliary supervision?

  2. How well do the assumptions of Theorem 2, particularly calibrated quantiles and fixed gaps, reflect the actual training dynamics with jointly learned gaps?

  3. Since the method models marginal action quantiles, how are dependencies across action dimensions and time steps handled during sampling, especially in multimodal action settings?

5 comments

We thank the reviewer for recognizing the simplicity and practical value of our approach, as well as the strong experimental evidence supporting its effectiveness. We appreciate the constructive suggestions and have revised the manuscript and appendix accordingly, with all revisions highlighted in blue.

Question 1 (Weakness 1): Can the authors provide further evidence that joint quantile supervision reduces the variance of median updates in practice, rather than simply acting as additional auxiliary supervision?

Response:

We have added training-time diagnostics in Fig. 8 that provide empirical support for Theorem 2's prediction that joint quantile supervision reduces the variance of median updates.

Specifically, we track B/a2B/a^2, the normalized center-score variance, for Quantile Head, Median-Only, and Detached Median Anchor. This quantity estimates the local median-update noise coefficient in Theorem 2. Normalization by a2a^2 accounts for differences in local correction strength; the calculation is detailed in Appendix B.2. At 12,000 training steps, Quantile Head has the lowest normalized variance among the three variants at all three displayed finite-difference step sizes, with reductions of 20.8%, 36.5%, and 51.3% relative to Median-Only at h=0.02,0.03,0.05h=0.02, 0.03, 0.05, respectively. Here, hh is the finite-difference step size used to estimate the local correction slope.

To examine the role of auxiliary supervision, the Detached Median Anchor control retains auxiliary quantile supervision of the shared features while blocking its direct gradients to the median output. The median-gradient rescaling used in this control is described under “Ablation Settings” in Appendix B. Quantile Head also exhibits lower normalized variance than this control. This comparison supports a contribution from directly coupling quantile supervision to median updates, in addition to the auxiliary supervision of shared features.

Question 2 (Weakness 1): How well do the assumptions of Theorem 2, particularly calibrated quantiles and fixed gaps, reflect the actual training dynamics with jointly learned gaps?

Response:

We have added diagnostics in Fig. 7 that support treating these assumptions as local approximations during late-stage training, even when the gaps are learned jointly.

(1) Calibrated quantiles. Figure 7(a) shows that the pooled empirical coverage of the median and its neighboring quantiles stabilizes near their nominal levels. At 12,000 training steps, the absolute coverage errors are at most 0.71 percentage points, providing evidence of approximate calibration at the pooled level on the diagnostic observations.

(2) Fixed gaps. Theorem 2 holds the gaps fixed when analyzing the local median-update direction, while the implemented model learns them jointly. Fig. 7(b) shows that per-update changes in the learned adjacent gaps remain small during the second half of training. At the final update (step 12,000), the mean absolute gap change is only 0.0496% of the mean gap size. These small per-update changes support treating the gaps as approximately fixed within a single update for the local analysis.

The diagnostic procedures are detailed in Appendix B.2. We also report variance diagnostics in Fig. 8, where the observed reduction in the normalized variance of median updates is consistent with Theorem 2's prediction.

Reference

[1] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. arXiv preprint arXiv:2502.19645.

Question 3 (Weakness 2): Since the method models marginal action quantiles, how are dependencies across action dimensions and time steps handled during sampling, especially in multimodal action settings?

Response:

We address coordination through shared action representations and structured sampling, supported by new diagnostics in Appendix B.6 (Fig. 21).

(1) Shared representations support coordinated predictions. Action tokens exchange information through bidirectional attention across the entire action chunk, while action coordinates are predicted from shared observation-conditioned features. Coordinate-wise quantile losses therefore jointly update a shared action expert, allowing supervision across dimensions and time steps to shape coordinated predictions. This design is consistent with OpenVLA-OFT, which combines bidirectional attention, parallel action-chunk prediction, and coordinate-wise L1 supervision for effective robot control [1].

(2) Structured sampling preserves pooled correlations in the tested settings. During stochastic decoding, each coordinate's sampled quantile rank is held fixed throughout the ten-step chunk, and correlated latent variables can control how sampled ranks vary across replanning calls. Our new diagnostics also vary the degree of rank sharing across coordinates, from independent to fully shared. Using a single checkpoint evaluated on 1,024 training inputs, all 99 tested stochastic settings closely preserve the demonstrations' pooled correlation structure: preservation scores are 0.9429–0.9787 across dimensions and 0.9913–0.9916 across time steps. These scores measure correlation errors relative to a shuffled-label reference, with a score of 1 indicating a match to the demonstration correlations. Median decoding also achieves high scores of 0.9787 and 0.9916, respectively. These diagnostics assess pooled correlations in the tested setting.

(3) Quantile Head retains multiple marginal action modes. In the controlled experiments with only two demonstration trajectories per setting, Figs. 15 and 18 show peaks near both demonstrated action values in the ambiguous coordinates. In these experiments, flow matching produces broader distributions, while L1 and L2 each produce a single intermediate value. This illustrates the benefit of retaining marginal distributional information alongside the coordination supported by the shared expert and sampling scheme. For task performance, Fig. 4 shows that the best tested sampling configuration improves LIBERO-Long success from 98.6% to 99.6%, providing a modest additional benefit over median decoding.

Dear Reviewer,

We would like to follow up on our response above as the discussion deadline approaches. Could you let us know whether our response addresses your main concerns, or if any points need further clarification? We would be happy to provide additional details.

Thank you for your time and consideration.

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

This paper introduces the Quantile Head, a single-pass action head for Vision-Language-Action (VLA) models that bridges point regression and flow matching. Building on a unified objective for the two paradigms, the authors derive a quantile regression formulation (Theorem 1: integrated pinball loss is equivalent to CDF supervision) and show, via a local analysis (Theorem 2), that jointly supervising auxiliary quantiles reduces the variance of direct median updates under calibrated, nearby quantiles with matched correction rates. The head predicts a median plus positive, separately learned left/right gaps, yielding 21 non-crossing marginal action quantiles (τ from 0.025 to 0.975) in one forward pass, trained with a masked pinball loss (10% random coordinate dropout, shared across all quantiles of a coordinate). With median decoding, it achieves 99.3% average success on LIBERO, state-of-the-art results on LIBERO-Plus (87.1% zero-shot / 89.1% SFT) and LIBERO-Pro (60.2%), and 75.0% average success across two real-robot tasks, while also enabling optional inference-time sampling from the same outputs without retraining.

Critical strengths
  1. Clean theoretical grounding. The unified regression/flow-matching objective and the CDF↔pinball equivalence (Theorem 1) give the quantile formulation a principled basis, while Theorem 2 provides a plausible mechanism—reduced median-update variance—for why joint quantile supervision helps, supported empirically by matched ablations (e.g., median-only 97.5% vs. full head 99.3% on LIBERO).
  2. Practical efficiency with distributional expressiveness. A single forward pass produces an ordered set of quantiles, avoiding the iterative integration cost of flow matching (mean episode time 5.88s vs. 12.57s for flow matching, a ~27.7% reduction even over L1 regression), while still representing per-coordinate action uncertainty and supporting sampling strategies at inference time.
  3. Strong empirical results. State-of-the-art or near-SOTA on LIBERO, LIBERO-Plus, and LIBERO-Pro, plus substantial real-robot gains over a fully fine-tuned π0.5 baseline (+16.5 points average), with consistent component ablations isolating the contributions of joint pinball supervision, the label mask, and the median gradient path.
  4. Honest evaluation of limitations. The paper openly reports that sampling gains are task-dependent (Figure 6 controlled scenes) and that sampling-strategy selection remains unsolved, which strengthens its credibility and points to useful future work.
Critical weaknesses
  1. Insufficient justification for the generalization gains. The paper's central theoretical result (Theorem 2) is an optimization-level statement: joint quantile supervision reduces the variance of median updates under idealized conditions. However, the headline claims concern generalization under distribution shift (LIBERO-Plus/Pro), and no argument is provided bridging lower update variance to out-of-distribution robustness. Moreover, the attribution is confounded: the method changes several things at once relative to baselines—the loss (pinball vs. L1/L2), the amount of supervision per sample (1 vs. 21 targets per coordinate), the coordinate-level label mask, and critically the training recipe (frozen VLM + prompt tuning vs. fully fine-tuned π0.5). Since freezing the VLM is known to preserve pretrained visual features and avoid catastrophic forgetting, the gains on visually perturbed benchmarks may stem largely from this recipe rather than from the quantile structure itself. Decisively, all isolation ablations (Table 4a/b/c) are conducted only on in-distribution LIBERO suites; there is no ablation on LIBERO-Plus/Pro showing that the quantile design, the mask, or the joint supervision specifically drives the OOD improvements. A K=1 ablation (which reduces the head to L1 + masking) and a frozen-VLM + L1 + mask control are needed to establish the net contribution of the quantile structure.
  2. Theorem 2 rests on assumptions that training clearly violates. Calibrated nearby quantiles, fixed gaps, matched local correction rates, and smooth densities do not hold when gaps are jointly learned and quantiles are not guaranteed calibrated. The result is therefore motivational rather than explanatory, and the causal chain from "reduced median-update variance" to "higher success rate" is never formally or empirically closed.
  3. Headline gains are within noise on saturated benchmarks. On LIBERO, several baselines already exceed 98%, making the 0.4-point margin over InternVLA-A1.5 plausibly within evaluation variance; no confidence intervals, seeds, or statistical tests are reported anywhere. The LIBERO-Pro table is also garbled in the provided text, making several baseline numbers difficult to attribute.
Questions for authors

See Weakness.

4 comments

We thank the reviewer for recognizing the theoretical grounding, practical efficiency, and empirical strengths of our work, as well as our transparent discussion of its limitations. We appreciate the constructive suggestions and have revised the manuscript and appendix accordingly, with all revisions highlighted in blue.

Weakness 1(a): Theorem 2 establishes variance reduction under idealized conditions, but its connection to the claimed OOD generalization gains remains unclear.

Response:

Theorem 2 establishes a local optimization advantage of joint quantile supervision under its stated assumptions. Our additional experiments examine the empirical relationship between this advantage and IID/OOD performance. We clarify this connection below.

(1) Empirical evidence. We have added local median-update variance diagnostics alongside IID/OOD performance in Fig. 8. The diagnostic measures normalized center-score variance (B/a2B/a^2), which estimates the local update-noise coefficient in Theorem 2. At 12,000 training steps, Quantile Head has the lowest mean normalized variance among the three variants, with reductions of 20.8%, 36.5%, and 51.3% relative to Median-Only at h=0.02,0.03,0.05h=0.02, 0.03, 0.05, respectively. Here, hh is the finite-difference step size used to estimate the local correction slope. Quantile Head also achieves the highest success rates among these variants on LIBERO (IID: 99.3%) and LIBERO-Pro (OOD: 60.2%) (Table 4). These findings support the predicted local optimization advantage and show that it is accompanied by improved IID/OOD performance.

(2) Theoretical explanation. Theorem 2 shows that joint quantile supervision reduces the variance of median updates under the stated assumptions. For a locally smooth risk function, lower update variance tightens the upper bound on expected risk after an update when the mean update and step size are matched [1, Lemma 4.2]. This provides a possible optimization-based explanation for the observed gains.

Weakness 1(b): Insufficient justification for the generalization gains.

Response:

(1) Method components. Pinball loss and multi-quantile supervision are integral to the theoretical formulation of Quantile Head, while coordinate-wise label masking is an additional design contribution. We have added controlled comparisons to assess their contributions to OOD performance.

(2) OOD component ablations. We have added OOD ablations averaged over three training seeds on LIBERO-Pro (Table 4) and explicitly labeled Median-Only as K=1K=1. These ablations include the requested comparisons between frozen and fine-tuned VLMs and between Quantile Head and L1. Under matched settings, Quantile Head improves OOD success by 4.0 percentage points over Median-Only, while label masking provides a further 0.6 percentage-point gain for Quantile Head. Quantile Head also outperforms L1 by 8.7 and 7.3 percentage points with frozen and fine-tuned VLMs, respectively, showing that its OOD advantage persists under both adaptation settings.

(3) VLM freezing versus fine-tuning. The additional OOD evaluation across three training seeds shows higher mean success rates with the full VLM fine-tuning configuration, while IID performance remains comparable. We use a frozen VLM with soft prompts for efficiency: this configuration requires approximately one-sixth as many trainable parameters as full fine-tuning.

A possible explanation is that freezing the VLM limits the adaptation of pretrained representations to robot control, whereas joint fine-tuning allows vision-language representations to adapt to the action-prediction objective. This may alleviate task-specific representation bottlenecks. This interpretation is consistent with Shachaf et al.'s analysis of the limitations of freezing the initial layers of deep linear networks [2]. Li et al. also report an OOD advantage for full fine-tuning in their cross-domain visual classification experiments [3]. These studies provide related theoretical and empirical context for our observations.

Weakness 2: Theorem 2 rests on assumptions that training clearly violates.

Response:

We have added diagnostics for all four assumptions in Fig. 7. These measurements support treating the assumptions as local approximations during late-stage training. Specifically:

(1) Calibration of nearby quantiles — Fig. 7(a). The pooled empirical coverage stabilizes near the nominal quantile levels as training progresses. At 12,000 training steps, the absolute coverage errors for the median and its two neighboring quantiles are at most 0.71 percentage points, supporting approximate calibration at the pooled level.

(2) Fixed gaps — Fig. 7(b). The adjacent quantile gaps become increasingly stable. At step 12,000, the mean absolute change in the gaps over one optimizer update is only 0.0496% of the mean gap size. This supports treating the gaps as locally fixed when analyzing a median update.

(3) Matched local correction rates — Fig. 7(c). At 12,000 training steps, the ratio of the mean correction slope computed from the three central quantiles to that computed from the median alone ranges from 0.991 to 0.997 across the tested finite-difference step sizes. These values are close to one, supporting approximately matched local correction rates.

(4) Local density regularity — Fig. 7(d). The pooled density estimates remain positive and vary little across nearby quantiles, ranging from 12.58 to 12.66 at h=0.02h=0.02 around the median at 12,000 training steps. This behavior is consistent with the local density regularity assumed in the analysis.

We have also examined the prediction that joint quantile supervision reduces the variance of median updates. At 12,000 training steps, Fig. 8 shows lower normalized variance for Quantile Head than for Median-Only and Detached Median Anchor at all three displayed finite-difference step sizes. The corresponding comparisons in Table 4 also show higher IID/OOD success rates, as detailed in our response to Weakness 1(a).

Weakness 3: Headline gains are within noise on saturated benchmarks.

Response:

We observe improvements on more challenging benchmarks as well as on LIBERO, and our main simulation results for Quantile Head are averaged over three training seeds. We clarify the benchmark coverage and evaluation protocol below.

(1) Gains on more challenging benchmarks. We include LIBERO as a standard reference benchmark, alongside LIBERO-Pro, LIBERO-Plus under zero-shot transfer and supervised fine-tuning (SFT), and real-robot evaluations. Quantile Head achieves the highest aggregate success rates among the compared methods on LIBERO-Pro and in both LIBERO-Plus settings. On LIBERO-Pro, it achieves 60.2% versus 53.3% for the strongest baseline. On LIBERO-Plus, it achieves 87.1% in zero-shot transfer and 89.1% with SFT, exceeding the strongest baseline in each setting by 1.4 and 5.0 percentage points, respectively. Averaged over the two real-robot tasks, Quantile Head achieves 75.0% success, compared with 58.5% for π₀.₅. These evaluations show improvements in settings with substantial room for improvement.

(2) Three-seed evaluation. Our benchmark results for Quantile Head are already averaged over three training seeds. Due to space constraints, we documented this protocol in the appendix: LIBERO (lines 992–996), LIBERO-Plus (lines 1004–1005), and LIBERO-Pro (lines 1006–1007 and 1010–1011). The quantile-count and mask-ratio ablations are also averaged over three seeds, as stated in lines 1064–1066 and 1071–1073 and in the captions of Figs. 9–10. The newly added component ablations likewise report three-seed averages. We have also highlighted this protocol in the captions of Tables 1–2. We report mean success rates without confidence intervals for consistency with the baseline studies, which likewise report only mean success rates.

Table 1(b): Formatting. Thank you for bringing this to our attention. We did not observe any misalignment in our local rendering: all 10 baseline entries and the Quantile Head entry appear correctly aligned. The discrepancy may stem from differences in rendering platforms. We have recompiled the PDF and hope that the updated version displays correctly on your platform.

References

[1] Léon Bottou, Frank E. Curtis, and Jorge Nocedal. Optimization Methods for Large-Scale Machine Learning. SIAM Review, 60(2):223–311, 2018.

[2] Gal Shachaf, Alon Brutzkus, and Amir Globerson. A Theoretical Analysis of Fine-tuning with Linear Teachers. NeurIPS, 2021.

[3]Kunyang Li, Jean-Charles Noirot Ferrand, Ryan Sheatsley, Blaine Hoak, Yohan Beugin, Eric Pauley, and Patrick McDaniel. On the Robustness Tradeoff in Fine-Tuning. ICCV, pp. 4898–4907, 2025.

Dear Reviewer,

We would like to follow up on our response above as the discussion deadline approaches. Could you let us know whether our response addresses your main concerns, or if any points need further clarification? We would be happy to provide additional details.

Thank you for your time and consideration.

Dear Reviewer,

To further address your concern about whether the local analysis in Theorem 2 reflects actual training, we conducted an additional experiment measuring the empirical variance of one-step changes in median predictions.

We recorded how much the predicted median changed on the same 120 fixed probe inputs before and after each measured optimizer step. The table below reports the empirical variance of these changes at six training checkpoints, averaged over three training seeds.

Table. Empirical variance of one-step changes in median predictions. Values are in units of 10−6 (normalized action unit)210^{-6}\,(\text{normalized action unit})^2.

Training stepMedian-OnlyDetached Median AnchorQuantile Head
00.8330.8040.781
2,00019.44819.31116.626
6,0004.9644.4854.115
8,0001.2581.2621.252
10,0000.8620.8230.720
12,0000.4240.3640.340

When averaged over the three seeds, Quantile Head exhibits lower empirical variance than both controls at all six evaluated checkpoints. At 12,000 steps, the reductions are 19.8% relative to Median-Only and 6.6% relative to Detached Median Anchor. The latter retains auxiliary quantile supervision while blocking its direct gradient to the median output.

We would welcome your feedback on this additional evidence. Thank you for your time and consideration.

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

The paper proposes a Quantile Head for vision-language-action (VLA) models that predicts ordered quantiles for each action coordinate in a single forward pass using a median anchor and positive cumulative gaps. Training jointly supervises these quantiles with a randomly masked pinball loss. Inference uses the median by default and also supports quantile sampling without retraining. The theoretical analysis establishes an equivalence between the integrated squared CDF loss and the integrated quantile loss, and shows that auxiliary quantiles can reduce median-update variance under specific local conditions. Experiments on three simulation benchmarks and two real-robot tasks support the benefits of joint quantile supervision for deterministic control.

Critical strengths
  1. The method is simple and predicts ordered quantiles in a single forward pass. Compared with Median-Only under the same configuration, it improves success rates by 1.8/4.0 percentage points on LIBERO/LIBERO-Pro, supporting the value of auxiliary quantile supervision.

  2. Experiments cover three simulation benchmarks and two real-robot tasks, with comprehensive component ablations. Average real-robot success improves from 58.5% to 75.0%.

Critical weaknesses
  1. The pooled statistics in Figures 7–8 do not establish conditional calibration or measure actual training-update variance.

  2. The real-robot comparison changes both the action head and adaptation strategy, leaving the contribution of quantile supervision unisolated.

  3. In the two-demonstration obstacle experiment, median decoding achieves 0% success and the best sampling setting reaches only 15%, failing to demonstrate reliable obstacle avoidance.

Questions for authors
  1. Can you provide evidence of conditional calibration or actual update variance to validate the noise-reduction mechanism?

  2. How does Median-Only perform on the real-robot tasks under otherwise identical settings?

3 comments

Thank you for serving as an emergency reviewer for our manuscript. We appreciate your time and feedback, especially on such short notice. We are preparing a point-by-point response to your comments and will post it as soon as possible.

Thank you for reviewing our manuscript on short notice as an emergency reviewer. We appreciate your recognition of the simplicity of our single-pass quantile head, the gains over Median-Only in the matched LIBERO and LIBERO-Pro comparisons, and the breadth of our simulation, real-robot, and ablation experiments.

Question 1. Can you provide evidence of conditional calibration or actual update variance to validate the noise-reduction mechanism? Weakness 1. The pooled statistics in Figures 7–8 do not establish conditional calibration or measure actual training-update variance.

**Answer: **

To address your question, we conducted an additional experiment on median-update variance. We recorded how much the predicted median changed on the same 120 fixed probe inputs before and after each measured optimizer step. At all six evaluated checkpoints, the empirical variance of these changes was lower for Quantile Head than for both Median-Only and Detached Median Anchor, averaged over three training seeds. This result is consistent with the noise-reduction prediction of Theorem 2.

Table. Empirical variance of one-step changes in the predicted median. Values are in units of 10−6 (normalized action unit)210^{-6}\,(\text{normalized action unit})^2.

Training stepMedian-OnlyDetachedQuantile Head
00.8330.8040.781
2,00019.44819.31116.626
6,0004.9644.4854.115
8,0001.2581.2621.252
10,0000.8620.8230.720
12,0000.4240.3640.340

In addition to the new experiments, we would like to clarify what the experiments in Figures 7 and 8 measure.

(1)Figure 7(a) reports the empirical coverage of the predicted conditional quantiles across 1,024 observations. Estimating coverage separately at a fixed observation is difficult because each observation provides only one expert action label per prediction step and motion coordinate. We therefore use a property of conditional calibration that we can measure directly with the available data: if a predicted quantile is calibrated at each input, its coverage across inputs should match its nominal level. Figure 7(a) tracks this relationship during training for the three quantiles near the median that are relevant to Theorem 2. At 12,000 training steps, their empirical coverage rates were 45.96%, 50.66%, and 55.42%, each within 0.71 percentage points of its respective target of 45.25%, 50%, and 54.75%.

(2)Figure 8 provides empirical evidence for the noise-reduction mechanism in Theorem 2 by measuring the variance of the loss gradient with respect to the median output. Panels (a–c) show the normalized gradient variances estimated across observations, and panels (d–f) show their relative values across methods. We report relative values because, at matched local correction rates, Theorem 2 gives Var⁡(Δm)=η2B/a2\operatorname{Var}(\Delta m)=\eta^2 B/a^2. When comparing two methods, the common factor η2\eta^2 cancels. Thus, under the theorem’s assumptions, the ratio of their B/a2B/a^2 values equals the variance ratio of their direct median updates.

Question 2. How does Median-Only perform on the real-robot tasks under otherwise identical settings? Weakness 2. The real-robot comparison changes both the action head and adaptation strategy, leaving the contribution of quantile supervision unisolated.

We added two matched real-robot experiments using Median-Only and Detached Median Anchor with the same adaptation strategy, training data, and evaluation protocol as Quantile Head. Median-Only achieved 30% success on Place apple on yellow plate and 82% on Remove cuboid from blue plate, averaging 56.0%. Quantile Head achieved 52% and 98%, respectively, improving success by 22 and 16 percentage points on the two tasks and 19.0 percentage points on average. With the adaptation strategy held fixed, these gains support the benefit of auxiliary quantile supervision on both tasks. Detached Median Anchor, which retains auxiliary quantile supervision while blocking its direct gradient to the median output, achieved 62.5% average success, compared with 75.0% for Quantile Head.

Real-robot success rates (%). The original fully fine-tuned π0.5\pi_{0.5} baseline is included for reference.

Real-robot taskπ0.5\pi_{0.5}Median-OnlyDetachedQuantile Head
Place apple on yellow plate31303652
Remove cuboid from blue plate86828998
Average58.556.062.575.0

Weakness 3. In the two-demonstration obstacle experiment, median decoding achieves 0% success and the best sampling setting reaches only 15%, failing to demonstrate reliable obstacle avoidance.

We added a real-data comparison in the obstacle setting by increasing the training set from one demonstration per detour to ten per detour (10 left, 10 right). With the same shared quantile sampling interval, τ∈[0.2,0.8]\tau \in [0.2, 0.8], Quantile Head’s task success rate increased from 15% to 87%, a gain of 72 percentage points. With 20 demonstrations, median decoding and L1L_1 regression both remained at 0%, while Flow Matching achieved 5%. These results suggest that limited demonstration coverage contributed to the low success rate in the original experiment. With more examples of both detours, quantile sampling enabled reliable completion of the pick-and-place task in the tested obstacle setting.

Table. Pick-and-place task success rates (%) in the obstacle setting.

Method2 demonstrations (1 left, 1 right)20 demonstrations (10 left, 10 right)
Quantile Head, shared sampling (τ∈[0.2,0.8]\tau \in [0.2, 0.8])1587
Quantile Head, median decoding00
L1L_1 regression00
Flow Matching05

Comments on the paper

1 comment

We thank the reviewers, the Area Chair, and the Senior Area Chair for their time and careful consideration of our work. We appreciate Reviewer xxwE's recognition of its theoretical grounding, practical efficiency, empirical strengths, and transparent discussion of limitations, as well as Reviewer czhF's recognition of the simplicity and practical value of our approach and the strong experimental evidence supporting its effectiveness. The reviewers' constructive comments have helped us improve the manuscript. In response, we have added clarifications and supplementary experiments, with all revisions highlighted in blue. For ease of review, we have combined the main text (the original nine pages plus one additional page of experiments) and the appendix into a single PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.