acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Precision Forcing for Generalist Vision-Language-Action Policies

Abstract

Coupling a pretrained vision-language model (VLM) with a diffusion or flow-matching action expert has become the mainstream vision-language-action (VLA) paradigm for embodied policies, yet they remain unreliable once a task demands precise manipulation, e.g., submillimeter cable insertion. We attribute this weakness to a training objective that averages flow-matching errors across contexts and flow times, allocating little supervision to the few decisive moments determining task success. Along the generative path, inference integrates a locally supervised velocity field, so equally penalized errors at different flow times can affect the executed action unequally. Consequently, precision-task performance deteriorates as physical tolerance tightens by orders of magnitude. We call this supervision imbalance across contexts and flow times Flow-Time Starvation. The remedy is to jointly align supervision allocation and representation learning with context-dependent precision demand, enabling the policy to recognize when greater precision is needed. We instantiate this precision-forcing strategy as Precision-aware Reweighting and Internalization for Skilled Manipulation, which scores precision demand offline, strengthens flow supervision where errors are consequential and action signals are visible, and internalizes that demand in the VLM condition representation. Simulation and real-robot experiments show that preserves or improves general-task performance while raising mean end-to-end success from to across drive connector insertion, bimanual connector mating, and screw tightening, substantially extending the precision boundary of existing embodied policies without new sensors or observation modalities.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robotic manipulation, flow matching, imitation learning

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

This paper proposes an accuracy-aware training method for VLA in high-precision scenarios that does not require changes to the model architecture or inference process. The authors found that existing Flow Matching methods apply similar training penalties to prediction errors in the VLA model across different operational phases and flow times, without fully considering the impact of these errors on the final task success rate. To this end, the authors propose PRISM, a VLA training method designed for fine-grained operations. The method estimates accuracy requirements for different operational stages based on the robot’s taught trajectory and adjusts the training loss accordingly, considering flow time and signal-to-noise ratio. Based on RLBench, LIBERO, and real-world robot experiments, PRISM has demonstrated strong performance and practical value. They address a significant and meaningful issue with current VLA models in fine-motor tasks. I believe the main contribution of this paper lies in incorporating accuracy requirements into the VLA training process and demonstrating the effectiveness of this design through experiments. The overall structure of the essay is clear, allowing readers to grasp the very issue the author raises. However, the paper still raises some issues that warrant further discussion. We raise these concerns in the Critical Weaknesses section and hope the authors will address them further.

Critical strengths
  1. The greatest strength of this article is that the research question is representative of real-world applications. The author links training supervision to the accuracy requirements of different operational phases, making the approach highly targeted. Compared to simply scaling up the model or increasing the amount of training data, this method offers an alternative approach to improving the VLA’s operational capabilities.

  2. This method does not increase inference overhead and can be easily integrated into existing VLA models. PRISM's PPS labels are generated offline before training, and CPR's prediction head is used only during training. It is a very flexible solution that offers a direction for future exploration.

  3. Results from experiments on real devices show that the improvements are concentrated in the fine-tuning phase, but gains were also observed on RLBench and LIBERO. This suggests that the authors did not observe any significant trade-off between fine-tuning performance and general performance.

Critical weaknesses

I think there is one crucial concern regarding the methodology. The core premise of PRISM is that different operational stages have different accuracy requirements. Therefore, different training weights should be assigned. However, the PPS in Section 3.2 is primarily based on the accuracy requirements for estimating the robot’s motion speed, end-effector displacement, and gripper state. This method is more about determining whether the robot is in a slow, stable operational phase rather than actually measuring physical accuracy requirements. In real-world applications, certain high-speed contact operations may also have strict accuracy requirements. In Figure 5(a), the authors report a high correlation between the CPR predictions and the PPS labels. However, this only demonstrates that the model can predict the accuracy labels designed by the authors. I don't think this proves that the PPS labels accurately reflect the actual physical tolerances. If PPS merely identifies slow-motion phases, then PRISM’s core mechanism could also be interpreted as a training weighting method that takes motion phases into account. I believe this issue is well worth considering in terms of future efforts to truly generalize these methods to the real world for performing precise operations. It also raises the question of whether this paper has overstated its contributions.

In addition, the authors need to provide results for two further experimental sections. This would further supplement the results that the cumulative ablation in Table 4 cannot demonstrate. First, we hope the authors will provide performance results using only Flow-time Weighting without PPS. Second, how effective would it be to apply weighting directly to the low-speed operation phase without using the full PPS?

Questions for authors

PPS primarily relies on motion speed and accuracy requirements to estimate gripper states. So, there are 2 questions.

  1. Could the authors further validate the relationship between PPS and real-world physical tolerances?
  2. Does this method remain effective when motion speed and accuracy requirements do not align?
4 comments
AuthorsOpenReview ↗

Author Response 1/2 | W1 & Q1 & Q2: PPS beyond slow motion and its link to tolerance

We thank the reviewer for the thoughtful review and for recognizing the practical value of precision-aware training. All revisions are highlighted in blue in the updated paper.

  1. PPS is not a slow-motion detector: PPS multiplies a within-trajectory motion rank by gripper engagement and by two gates, one that suppresses rapid closing and one that suppresses isolated gripper motion while the arms are still (Section 3.2). Slow motion alone therefore does not give a high score. Among 24 manually labeled slow preparation or free-motion samples, PPS selects none, while speed only and masked speed select 9 and 8 (Appendix G.4, Figure 13). In training, weighting slow phases directly reaches 70% insertion success, versus 95% with PPS (Appendix G.2, Table 16).

  2. Link to physical tolerance (Q1): Figure 5(a) tests whether CPR learns PPS, so we additionally validated PPS against references it never sees.

    • Manual labels of socket approach, alignment and contact on 180 held-out time points give AUROC 0.858, versus 0.707 and 0.718 for the speed rules. Within the top 25%, 82.0% of PPS selections fall in these contexts, versus 54.5% (Appendix G.4).
    • Measured tolerances of 43.75, 10 and 0.15 mm show that PPS puts 82.47% of its top-25% frames in the 0.15 mm stage, versus 60.61% and 62.52% (Appendix F.1).
    • Task success drops from 95% to 40% when PPS is shuffled within each trajectory, so the gain depends on PPS marking the precision-critical frames.

    Appendix C, Figure 7, shows the same pattern across nineteen tasks.

  3. Fast precision actions (Q2): Three properties keep PRISM effective.

    • Fast fine actions are still relatively slow. Our skilled operators collected the demonstrations through VR teleoperation at a natural pace, without slowing down for fine actions (Section 4.6). Even so, fine actions remain slower than large coarse motions in the same trajectory, and PPS only needs this relative difference because it ranks motion within each trajectory. Median insertion motion is 0.495× pickup and 0.545× handover motion (Appendix F.1).
    • No sample loses supervision. Since wi∈[1,1+γ]w_i\in[1,1+\gamma], a precision frame with a lower score still keeps the full standard FM weight, so its supervision remains at least as strong as the baseline objective.
    • Large-motion precision also benefits. Palletizing on the Sculptor robot requires precise relative placement during large arm and base motion, and PRISM raises success from 55% to 80%, above 70% for DAgger (Section 4.7). Appendix H.2, Figure 16, visualizes the corresponding task through two demonstrations with camera observations and full archived suction-adapted PPS traces.
  4. Contribution: As you summarize, PRISM brings precision demand into VLA training without changing the architecture or inference. PPS estimates this demand from signals present in every demonstration, and its validity rests on manual labels, measured tolerances and trained controls. With it, end-to-end success rises from 35/35/45% to 95/80/85% on the three MSSM tasks, while simulation performance is preserved or improved.

AuthorsOpenReview ↗

Author Response 2/2 | W2: Flow-time-only and low-speed trained controls

  1. Both requested controls are now trained: On drive insertion, we trained the two requested variants and several further controls with the same data, initialization and update budget as PRISM (Appendix G.2, Table 16). Each uses 20 trials with 95% Wilson intervals, and time and attempts average successful trials.

    VariantSR [95% CI]Time (s)Attempts
    π0.5 baseline0.35 [0.18, 0.57]474.1
    Flow time only, no PPS/CPR/SNR0.45 [0.26, 0.66]464.1
    Flow time + SNR, no PPS/CPR0.45 [0.26, 0.66]474.2
    PRISM without CPR0.75 [0.53, 0.89]443.2
    Randomized PPS0.40 [0.22, 0.61]514.3
    Unmasked low speed0.70 [0.48, 0.85]412.7
    Full PRISM0.95 [0.76, 0.99]382.1
  2. Flow-time weighting alone: Clean-end weighting without PPS, CPR or the SNR gate gives 45%, and adding the SNR gate also gives 45%. Flow-time emphasis therefore yields only a modest gain over the 35% baseline, and reaching 95% requires context-dependent demand.

  3. Weighting slow phases directly: Replacing PPS with the unmasked low-speed score in both PSW and CPR gives 70% with 2.7 attempts. Full PRISM reaches 95% with 2.1 attempts and the shortest time (38 s). This 25-point gap comes from the gripper cues and the active mask, which keep extra supervision away from slow preparation and handling (Figure 13).

  4. Further controls: Shuffled PPS gives 40%, and removing CPR gives 75%. Mean-normalized weights keep LIBERO at 97.45% versus 97.50% and mating at 16/20 (Appendix G.1). The gain therefore depends on correct demand assignment, not on loss scale.

  5. Why PPS serves both PSW and CPR: PPS tells CPR when local precision is needed and tells PSW where to add supervision. The cumulative ablation raises success from 35% to 45/70/85/95% as CPR, precision weighting, clean-end emphasis and SNR gating are added (Table 4). Table 5 shows the same benefit without VLM state inputs, with π0.5 rising from 30% to 95% and π0 from 25% to 80%.

Thanks for the response, which has addressed my concerns. I will maintain my positive assessment.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely appreciate your response and the positive feedback. Thank you for your time and valuable insights!

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

This paper identifies Flow-Time Starvation, where the training objective averages flow-matching errors across task contexts and flow times, allocating limited supervision to precision-critical moments. It proposes PRISM to align supervision allocation and representation learning with context-dependent precision demand, enabling the policy to recognize when greater precision is needed. Experiments on RLBench, LIBERO, and real-robot manipulation tasks report performance improvements.

Critical strengths

S1. The paper points out that flow-matching training averages errors across task contexts and flow times, leaving limited supervision for precision-critical interactions. This provides a reasonable motivation for PRISM.

S2. Experiments on VLA backbones and three real-robot tasks report consistent performance gains, supporting the practical value of PRISM.

Critical weaknesses

W1. PPS estimates precision demand from slow motion and gripper engagement, but their relationship to actual physical tolerance is not sufficiently validated. Eq. (11) also combines joint and end-effector displacements without clarifying their different units and scales.

W2. Proposition 1 analyzes physical tolerance and flow-time error propagation, while PSW uses PPS, clean-end emphasis, and SNR gating. The paper seems does not clearly justify how this weighting reflects the theoretical endpoint sensitivity. Moreover, Proposition 1 is primarily based on error linearization and existing mathematical tools, it does not establish a novel theoretical framework.

W3. Eq. (3) defines tolerance-normalized risk in task space using inverse-squared tolerance weighting, while Figure 2(b) uses action-space errors scaled by inverse tolerance. It remains unclear how these results validate the proposed Flow-Time Starvation mechanism.

Questions for authors

Q1. How well does PPS reflect actual physical precision requirements? In Eq. (11), are joint and end-effector displacements normalized before being combined?

Q2. Could the authors clarify how the PSW weighting in Eq. (13) relates to the endpoint sensitivity derived in Proposition 1? Does this weighting change the optimal velocity field of standard flow matching?

Q3. How does Figure 2(b) support Flow-Time Starvation when it uses a different error measure from Eq. (3)?

5 comments
AuthorsOpenReview ↗

Author Response 1/3 | W1 & Q1: PPS validity and displacement scales

We thank the reviewer for the constructive review. All revisions are highlighted in blue in the updated paper.

  1. PPS is validated against three independent references: Demonstrations do not record tolerance, so PPS estimates precision demand from relative motion and gripper cues. We check it against references that PPS never sees.

    • Manual labels. On 180 held-out time points labeled for socket approach, alignment and contact, PPS reaches AUROC 0.858 [0.808, 0.911], versus 0.718 for masked speed. Its top-25% selections contain 82.0% of these contexts, versus 54.5% (Appendix G.4).
    • Measured tolerances. PPS puts 82.47% of its top-25% frames in the 0.15 mm insertion stage, versus 62.52% for masked speed, and the gain is positive in all ten episodes (Appendix F.1).
    • Task success. PRISM reaches 95% insertion success, versus 70% with low-speed scores and 40% with shuffled PPS (Appendix G.2, Table 16).

    Appendix H.2, Figure 16, further shows two Sculptor palletizing demonstrations with camera observations and full archived suction-adapted PPS traces.

  2. Units and normalization in Eq. 11: Eq. 11 mixes the raw joint and EEF displacements as si=αqΔqi+αeΔeis_i=\alpha_q\Delta q_i+\alpha_e\Delta e_i with αq=αe=0.5\alpha_q=\alpha_e=0.5. The word normalized referred to the coefficients summing to one, not to each channel, so Section 3.2 now calls sis_i a weighted mixture. The mixture is then converted to a percentile rank within each trajectory, so the final motion score is dimensionless and unaffected by any common rescaling. The same coefficients are used for every dual-arm dataset, including the nineteen tasks in Appendix C. Appendix B now lists every step, including the activity threshold, gripper normalization, cue scaling, the neutral inactive score and seven-frame smoothing.

  3. Why the gripper cues matter: Gripper engagement separates manipulation from free motion. The acquisition gate suppresses rapid closing, and the quiescent-hand gate suppresses isolated gripper motion while the arms are still. Among 24 slow preparation or free-motion samples, PPS selects none, while speed only and masked speed select 9 and 8. Figure 13 shows both speed rules highlighting ordinary handling that PPS suppresses.

AuthorsOpenReview ↗

Author Response 2/3 | W2 & Q2: PSW, Proposition 1, and the weighted optimum

  1. How PSW follows Proposition 1: In Proposition 1, endpoint sensitivity depends on Ac(τ)=Λ(c)1/2JcΦc(0,τ)A_c(\tau)=\Lambda(c)^{1/2}J_c\Phi_c(0,\tau), the product of a context factor set by tolerance and geometry and a propagation factor set by flow time. Neither can be observed in demonstrations, so Eq. 13 uses one surrogate for each,

    wi=1+γ ρi (1−τi)p (1−e−SNRi/τ0).w_i=1+\gamma\,\rho_i\,(1-\tau_i)^p\,(1-e^{-\mathrm{SNR}_i/\tau_0}).

    PPS ρi\rho_i represents the context factor and (1−τi)p(1-\tau_i)^p represents propagation. The SNR gate (Eq. 9) delays extra weight until the action is visible, because a small precision chunk can stay hidden in noise longer than a larger routine chunk at the same flow time.

  2. Empirical support: Equal-energy perturbations give the largest endpoint response at the clean end for all three stages and both energies (Appendix E.1, Figure 11), so the flow-time factor is shared across stages as assumed. PPS puts 82.47% of its top-25% frames in the 0.15 mm insertion stage, whose inverse tolerance is 291.67 times that of pickup (Appendix F.1). Adding CPR, ρ\rho, clean-end emphasis and SNR gating raises insertion success from 35% to 45%, 70%, 85% and 95% (Table 4).

  3. Theoretical contribution: Proposition 1 deliberately uses standard tools, linearization and a weighted Cauchy Schwarz bound, so that the result is easy to verify. Its novelty lies in what it bounds. It brings tolerance Λ\Lambda, task geometry JJ and flow-time propagation Φ\Phi into one sensitivity term for flow-matching policies, which explains why a uniform FM loss is misaligned with physical consequences. The bound is also tight, with equality for residuals proportional to pτ−1Ac⊤hmax⁡p_\tau^{-1}A_c^\top h_{\max}, so Sc(B)\mathcal S_c(B) is the exact worst-case leverage of a region. This diagnosis, which we call Flow-Time Starvation, directly motivates the design of PRISM.

  4. Effect on the optimal field: PSW keeps the FM target U=ϵ−z∗U=\epsilon-z^*. With inputs X=(xτ,τ,c)X=(x_\tau,\tau,c) and weight WW, the weighted optimum is

    vW∗(X)=E[WU∣X]/E[W∣X],v_W^*(X)=\mathbb E[WU\mid X]/\mathbb E[W\mid X],

    which differs from the standard conditional mean only by Cov⁡(W,U∣X)/E[W∣X]\operatorname{Cov}(W,U\mid X)/\mathbb E[W\mid X] (Appendix F.2). The clean-end factor depends only on the input τ\tau, while ρi\rho_i and the chunk scale σf,i\sigma_{f,i} are fixed for each anchor frame and its demonstrated chunk. The difference is therefore zero whenever samples sharing an input have the same precision demand and action scale, and the bounded weight range [1,1+γ][1,1+\gamma] limits it otherwise. Positive weights also keep every exactly realizable zero-error solution. PSW thus reallocates finite capacity, and the residual analysis supports this role. On the evaluated checkpoints, raw residuals fall in all five PPS bins, including 2.25% in the lowest-demand bin and 2.10% in pickup, where a biased target would show first. Pickup success stays at 100%, and handover rises from 90% to 100% (Table 3).

AuthorsOpenReview ↗

Author Response 3/3 | W3 & Q3: Figure 2(b) as a conservative instance of Eq. (3)

  1. What Figure 2(b) measures: We add equal-energy velocity perturbations in five flow-time intervals, record the endpoint action error per unit energy, and multiply it by each stage's inverse tolerance relative to pickup. The protocol uses 60 held-out observations, the native ten-step Euler sampler and two energies (Appendix E.1). Figure 11 shows the unweighted curves with 95% bootstrap intervals.

  2. Relation to Eq. (3): If a stage allows error δ\delta in each constrained direction, then Λ(c)=δ−2I\Lambda(c)=\delta^{-2}I and Eq. (3) becomes δ−2 E∥Jcez∥2\delta^{-2}\,\mathbb E\|J_ce_z\|^2. Figure 2(b) sets JcJ_c to the identity, since the physical map is unknown, and weights by δ−1\delta^{-1} instead of δ−2\delta^{-2}. Because every stage tolerance is at most the pickup tolerance, the squared ratio is never smaller, so Figure 2(b) is a conservative version of Eq. (3). With the squared weighting, the insertion factor grows from 291.67 to about 8.5×1048.5\times10^4 and the handover factor from 4.375 to about 19.1, which only strengthens the fine-stage peak. The weighting scales whole stages and cannot shift the peak in flow time, because the unweighted response already peaks at τ∈[0,0.2]\tau\in[0,0.2] for every stage.

  3. Support for Flow-Time Starvation: The claim has two parts. First, equal FM errors have unequal endpoint effects across flow time, which Figure 11 shows without any tolerance weighting. Second, this burden concentrates in tight-tolerance stages, which Figure 2(b) shows under a weighting milder than Eq. (3). Both parts therefore also hold under Eq. (3). Appendix E.1 now includes this derivation.

Reviewer_7L2wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that standard flow-matching objectives distribute supervision across contexts and flow times even though their errors can have different effects on physical task success. It formalizes a local, tolerance-normalized endpoint-error bound and proposes PRISM, a training-only method with three components: Proprioceptive Precision Scoring (PPS) assigns frame-level precision-demand labels from robot motion and gripper cues; Precision-Scaled Weighting (PSW) reweights flow-matching loss using PPS, flow time, and an action-SNR visibility gate; and Conditional Precision Regularization (CPR) trains the VLM condition representation to predict PPS. The paper evaluates π0.5 and StarVLA in RLBench-10 and LIBERO, and π0.5 on three real-robot tasks. On the robot tasks, PRISM raises end-to-end success from 0.35/0.35/0.45 to 0.95/0.80/0.85 across connector insertion, bimanual connector mating, and screw tightening. The model architecture and inference procedure remain unchanged.

Critical strengths
  1. The method addresses a practically important problem while keeping deployment simple. PRISM changes training supervision without adding sensors, changing the policy architecture, or introducing inference-time computation. This makes it attractive for existing VLA policies if the gains transfer to other tasks and platforms.

  2. The empirical results are encouraging across simulation and hardware. The paper reports RLBench-10 mean gains of 12 percentage points for π0.5 and 10 points for StarVLA, as well as smaller gains on the near-saturated LIBERO suites. On three robot tasks, mean end-to-end success rises from 38.3% for π0.5 to 86.7% for PRISM, compared with 65.0% for the DAgger variant. The paper also includes cumulative component ablations and reports matched initialization, data, update budget, and inference settings for baseline and PRISM runs.

Critical weaknesses
  1. The evidence does not fully establish the proposed precision mechanism. Proposition 1 bounds local, linearized endpoint sensitivity, but the paper does not derive the particular PPS, clean-end, and SNR weighting rule in Eq. 13 from that bound or show that it approximates the sensitivity term in practice. PPS is a behavioral proxy based on motion ranks and gripper cues. CPR’s correlation with PPS on held-out episodes shows that the representation predicts its training labels, but does not independently establish that those labels track physical tolerance, alignment error, or failure risk. The reported ablations support the components on drive insertion, but controls such as mean-normalized weights or shuffled PPS labels would help distinguish precision-specific allocation from changes in overall loss scale or generic emphasis on selected frames.

  2. The physical evaluation has limited statistical and platform coverage. Each method is evaluated in 20 trials per task on one dual-arm platform, and the paper does not report uncertainty intervals or results across independent training runs. The three robot tasks use fixed tolerances; a success-versus-tolerance evaluation would more directly support the claim that PRISM extends the precision boundary. The reported gains are promising, but the current study leaves their statistical stability and broader generality uncertain.

Questions for authors
  1. Could you compare the proposed PSW with mean-normalized weights and a shuffled or matched-random PPS control? This would help separate selective precision-aware allocation from a change in overall gradient scale and test whether PPS itself drives the gains.

  2. What independent evidence supports PPS as a measure of physical precision demand, beyond predicting its training labels? For example, can you relate PPS to annotated alignment/contact phases, measured geometric tolerance, or success under controlled perturbations, and assess its sensitivity to trajectory speed and pauses?

  3. Could you report per-trial outcomes and uncertainty intervals for the physical results, clarify whether trial configurations were randomized or paired across methods, and provide results across independent training runs? A success-versus-tolerance curve on one task would also directly support the precision-boundary claim.

8 comments
AuthorsOpenReview ↗

Author Response 1/4 | W1: From Proposition 1 to the PSW weights

We thank the reviewer for the careful and constructive review. All revisions are highlighted in blue in the updated paper.

  1. PSW follows the two factors of Proposition 1: Proposition 1 shows that the same flow-matching error can have very different physical consequences. Its effect is governed by Ac(τ)=Λ(c)1/2JcΦc(0,τ)A_c(\tau)=\Lambda(c)^{1/2}J_c\Phi_c(0,\tau), which has two parts. The context factor Λ(c)1/2Jc\Lambda(c)^{1/2}J_c depends on tolerance and task geometry, and the propagation factor Φc(0,τ)\Phi_c(0,\tau) depends on flow time. Neither is recorded in demonstrations, so Eq. 13 uses one observable surrogate for each,

    wi=1+γ ρi (1−τi)p (1−e−SNRi/τ0).w_i=1+\gamma\,\rho_i\,(1-\tau_i)^p\,(1-e^{-\mathrm{SNR}_i/\tau_0}).

    PPS ρi\rho_i stands for the context factor, (1−τi)p(1-\tau_i)^p stands for the propagation factor, and the SNR gate waits until the action is visible in the noisy input. Since wi∈[1,1+γ]w_i\in[1,1+\gamma], PSW only adds supervision and never removes it from any sample.

  2. Both surrogates match what we measure: We checked each factor on held-out data.

    • Propagation. We perturb the velocity field with equal energy in each flow-time interval on 60 held-out observations. The endpoint response is largest in the clean-end interval [0, 0.2] for all three stages and both energies (Appendix E.1, Figure 11). The profile is shared across stages, which is exactly what a separate flow-time factor assumes.
    • Context. The stage tolerances of 43.75, 10 and 0.15 mm give inverse-tolerance ratios of 1, 4.375 and 291.67. PPS places 82.47% of its top-25% frames in the 0.15 mm insertion stage (Appendix F.1) and agrees with independent manual labels with AUROC 0.858 (Appendix G.4).

    Combining the two, Figure 2(b) peaks at the clean end of insertion, which is exactly where Eq. 13 places its largest weights.

  3. Each factor contributes to success: The cumulative ablation raises insertion success from 35% to 45% with CPR, 70% with ρ\rho, 85% with clean-end emphasis and 95% with SNR gating (Table 4). On the evaluated checkpoints, PRISM also lowers paired residuals by 2.97% [1.97, 3.99] and improves all five PPS bins (Appendix F.2, Table 12).

AuthorsOpenReview ↗

Author Response 2/4 | W1 & Q1: Mean-normalized weights and shuffled PPS

  1. Mean-normalized weights retain the gains: We divide the FM weights by their batch mean, w~i=wi/wˉB\tilde w_i=w_i/\bar w_{\mathcal B}, so that the batch-average FM weight is one, while retaining the original CPR targets and its loss coefficient of 0.1. The results are essentially unchanged from the original PRISM setting. LIBERO success is 97.45% with mean-normalized weights and 97.50% with the original weights, both above 96.85% for π0.5. Bimanual mating success remains 16/20 under both settings (Appendix G.1, Tables 14 and 15).

  2. Shuffled PPS removes most of the gain: We permute PPS once within each trajectory before training. This keeps the score distribution and the loss scale, but breaks the link between scores and frames. The shuffled scores replace PPS in both PSW and CPR. The table reports drive insertion over 20 trials with 95% Wilson intervals, and time and attempts average successful trials (Appendix G.2, Table 16).

    VariantSR [95% CI]Time (s)Attempts
    π0.5 baseline0.35 [0.18, 0.57]474.1
    Flow time only, no PPS/CPR/SNR0.45 [0.26, 0.66]464.1
    Flow time + SNR, no PPS/CPR0.45 [0.26, 0.66]474.2
    PRISM without CPR0.75 [0.53, 0.89]443.2
    Randomized PPS0.40 [0.22, 0.61]514.3
    Unmasked low speed0.70 [0.48, 0.85]412.7
    Full PRISM0.95 [0.76, 0.99]382.1
  3. PPS itself drives the gain: With the same score distribution and loss scale, shuffling lowers success from 95% to 40%. We also observe less smooth execution, alongside longer completion time and more attempts (51 s, 4.3 attempts). The improvement therefore comes from PPS marking the right frames, not from a larger loss or generic reweighting. We attribute the small remaining gain over the baseline (8/20 versus 7/20) to the retained flow-time-weighted loss, consistent with the modest 45% success of flow-time weighting alone.

AuthorsOpenReview ↗

Author Response 3/4 | W1 & Q2: Independent evidence for PPS, speed, and pauses

We test PPS against three references it never sees, namely manual labels, measured tolerances and task success.

  1. Manual labels: We manually labeled 180 time points from ten held-out insertion episodes using all three cameras, marking socket approach, alignment and contact. PPS identifies these contexts with AUROC 0.858 [0.808, 0.911], versus 0.707 for speed only and 0.718 for masked speed. Among its top-25% selections, 82.0% fall in these contexts, versus 54.5% for both speed rules, a paired gain of 27.5 points [19.0, 36.0]. Treating the 12 uncertain samples as either class still gives 80.0% for PPS versus 54.0% and 56.0% (Appendix G.4, Table 18).

  2. Measured tolerances: Drive insertion has stage tolerances of 43.75 mm for pickup, 10 mm for handover and 0.15 mm for insertion. Over all 6,692 held-out frames, PPS puts 82.47% of its top-25% frames in the 0.15 mm stage, versus 60.61% and 62.52% for the speed rules. The paired gain of 19.95 points [17.81, 21.83] is positive in all ten episodes and at every budget from 10% to 40% (Appendix F.1, Tables 10 and 11). In bimanual mating, PPS places 94.63% of its selections in the dual-grasp window, versus 67.19% and 72.78% (Appendix G.5).

  3. Task success: Shuffling PPS within each trajectory lowers insertion success from 95% to 40%, and replacing PPS with low-speed scores gives 70% (Appendix G.2, Table 16). Correct demand assignment is what turns insertion failures into successes, while pickup stays at 100%.

  4. Speed: PPS does not require fine actions to be slow in absolute terms. Our skilled operators collected the demonstrations through VR teleoperation at a natural pace, without slowing down for fine actions (Section 4.6). Even so, fine actions remain slower than coarse motions in the same trajectory, and PPS ranks motion within each trajectory, so this relative difference is enough. Median insertion motion is 0.495× that of pickup and 0.545× that of handover (Appendix F.1).

  5. Pauses: Inactive frames are excluded from the ranking and receive a neutral score of 0.5 (Appendix B). Among 24 manually labeled slow preparation or free-motion samples, PPS selects none, while speed only and masked speed select 9 and 8. Figure 13 shows the same pattern along full trajectories.

  6. Other tasks: Appendix C, Figure 7, shows PPS traces for nineteen tasks, including connector mating, screw tightening, USB insertion, puzzle assembly, water pouring and box stacking. PPS rises in fine-operation intervals and stays low in coarse-only tasks. Appendix H.2, Figure 16, additionally shows two Sculptor palletizing demonstrations, each with ten camera observations and a full archived suction-adapted PPS trace.

Comments on the paper

3 comments

Dear Reviewers,

Thank you again for your reviews of this submission. The authors have now posted their responses, and the author–reviewer discussion period ends on November 18, 2026 (AoE).

Please read the authors' rebuttal and let them know whether it addresses your concerns. A short reply is enough, for example:

  • which of your concerns are now resolved;
  • which remain open, and what evidence or clarification would resolve them;
  • whether you plan to update your rating.

If the rebuttal changes your assessment, please update your review and rating before the discussion period ends. If you have already engaged, thank you, and no further action is needed unless the authors reply again.

Your participation is important for a fair and well-informed decision.

AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their careful and constructive feedback. We are encouraged that the reviewers found the problem practically important and representative of real-world applications (7L2w, ttNt), the motivation of Flow-Time Starvation reasonable (SkwT), PRISM simple to deploy without changing the architecture or inference (7L2w, ttNt), and the gains consistent across simulation and real robots (7L2w, SkwT, ttNt). All revisions are highlighted in blue in the updated paper. The main changes are summarized below.

  1. Independent validation of PPS (7L2w, SkwT, ttNt; Appendices F.1, G.4): The main shared question was whether PPS reflects physical precision demand. We validated PPS against references it never sees. On 180 manually labeled time points, PPS identifies approach, alignment and contact with AUROC 0.858, versus 0.707 and 0.718 for speed rules. Over 6,692 held-out frames, it places 82.47% of its top-25% frames in the 0.15 mm insertion stage, versus 60.61% and 62.52%. PPS also selects none of 24 slow preparation or free-motion samples, while the speed rules select 9 and 8, so PPS is not a slow-motion detector.
  2. New trained controls that isolate the mechanism (7L2w, ttNt; Appendices G.1, G.2, Table 16): On drive insertion, we trained all requested controls with the same data, initialization and update budget as PRISM. Flow-time weighting alone gives 45%, flow time with SNR gives 45%, low-speed weighting gives 70%, shuffled PPS gives 40% and PRISM without CPR gives 75%, while full PRISM reaches 95%. Mean-normalized weights keep LIBERO at 97.45% versus 97.50% and mating at 16/20. The gain therefore comes from PPS marking the right frames, not from loss scale, flow-time emphasis or slow-phase weighting.
  3. New cross-platform validation on a second embodiment (7L2w, ttNt; Section 4.7, Appendix H): To address single-platform coverage, we repeated the full training and evaluation pipeline on a different robot, the SEEN·E Sculptor. It is a mobile robot with a single six-joint arm and a suction end effector, so it differs from the original dual-arm gripper platform in embodiment, number of arms, mobility and end effector. On box palletizing, the robot stacks 12 randomly piled boxes, and a wrong orientation or an offset above 3 cm fails the trial. Over 20 paired trials per method, PRISM reaches 80%, versus 55% for π0.5 and 70% for DAgger with 20 extra corrective demonstrations. Suction engagement replaces gripper closure, and all other PRISM rules stay unchanged. PRISM therefore remains effective on a new platform, demonstrating its cross-platform generality.
  4. Statistical analysis of the physical results (7L2w; Appendices F.3, G.3, I): We now report all 180 per-trial outcomes, 95% Wilson intervals and paired McNemar tests. PRISM improves over π0.5 with p = 4.2e-7 pooled and over DAgger with p = 0.0072 pooled. Retraining on RLBench with three seeds gives a paired gain of 0.115 ± 0.005, so run-to-run variation is small compared with the gain.
  5. Success across tolerances (7L2w; Appendix D): On three connectors with measured tolerances of 0.3, 0.2 and 0.1 mm, PRISM gains 30, 35 and 35 points over π0.5, so its advantage holds as tolerance tightens.
  6. Theoretical clarifications (7L2w, SkwT; Appendices E.1, F.2): We explain how Eq. 13 maps onto the context and propagation factors of Proposition 1 and verify both factors empirically. We show that the bound is tight, that PSW changes the optimal velocity field only through a bounded covariance term, and that Figure 2(b) is a conservative instance of Eq. (3).
  7. Presentation (SkwT; Section 3.2, Appendix B): Eq. 11 is now described as a weighted mixture whose percentile rank is dimensionless. Appendix B lists every PPS step, and Section 4.6 describes the natural-speed VR demonstrations.

We hope these revisions address the reviewers' concerns and would be glad to clarify any remaining questions.

AuthorsOpenReview ↗

Rebuttal Summary for AC

Dear Area Chair,

We thank you for handling our submission. All three reviewers rated the paper weak acceptance and agreed that the problem is important, the method is easy to deploy and the gains are consistent. Their concerns were requests for clarification and additional evidence, and we addressed each of them with new experiments or analysis.

  • Reviewer 7L2w raised four concerns. (i) For PPS validity, we validated PPS against manual labels (AUROC 0.858), measured tolerances and task success. (ii) For a possible loss-scale effect, mean-normalized weights leave results essentially unchanged (97.45% versus 97.50% on LIBERO), while shuffled PPS lowers insertion success from 95% to 40%, so the gain comes from PPS itself. (iii) For the link between Eq. 13 and Proposition 1, we mapped Eq. 13 onto the context and propagation factors and verified both empirically. (iv) For statistical stability and generality, we added per-trial outcomes, Wilson intervals, paired McNemar tests (p = 4.2e-7 pooled against π0.5), three RLBench seeds (paired gain 0.115 ± 0.005) and a success-versus-tolerance study from 0.3 to 0.1 mm. We also evaluated PRISM on a second robot platform, a mobile single-arm robot with a suction end effector, where it reaches 80% versus 55% for π0.5 and 70% for DAgger with all other rules unchanged.
  • Reviewer SkwT raised four concerns. (i) For PPS validity, we provided the same independent evidence. (ii) For units in Eq. 11, we clarified that the final motion score is a dimensionless percentile rank and documented every PPS step in Appendix B. (iii) For the relation between PSW and Proposition 1, we explained the two-factor mapping, clarified that the novelty lies in what the bound captures, showed that the bound is tight, and showed that PSW changes the optimal field only through a bounded covariance term. (iv) For Figure 2(b), we proved that it is a conservative instance of Eq. (3).
  • Reviewer ttNt raised three concerns. (i) For whether PPS only detects slow motion, PPS selects none of 24 slow non-precision samples, while the speed rules select 9 and 8. (ii) For fast precision actions, PPS ranks motion within each trajectory and therefore needs only a relative speed difference, and since wi≥1w_i\ge 1 no frame loses supervision. (iii) For the requested controls, flow-time-only weighting reaches 45% and low-speed weighting reaches 70%, versus 95% for full PRISM.

In summary, the rebuttal shows that (1) PPS tracks physical precision demand against independent references, (2) the gain comes from correct precision-demand assignment and not from loss scale or slow-phase weighting, (3) the real-robot gains are statistically significant and hold as tolerance tightens, and (4) PRISM remains effective on a second robot platform. We believe these results resolve the reviewers' concerns, and we thank you for your consideration.

Sincerely,

PRISM Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.