acceptodds
Under review as a conference paper at ICLR 2027

TRAC: Trajectory-Reliable Alignment Training for Action Chunks in VLA Models

Abstract

Action chunking predicts a sequence of future actions for execution between policy queries. Small action errors can accumulate and alter the intended motion, making relations across the sequence important for reliable manipulation. However, standard point-wise prediction losses sum penalties at individual action positions without explicitly evaluating cumulative deviations or coordinated variation. We introduce TRAC, Trajectory-Reliable Alignment Training for Action Chunks in VLA Models, which treats action chunks as structured action-space trajectories. TRAC complements flow matching with two objectives: weighted prefix consistency aligns accumulated translations with demonstrations throughout the chunk, while a local residual-structure prior encourages coordinated variation among neighboring actions. Both objectives operate on clean-action estimates, preserving the backbone's architecture and inference procedure. On LIBERO, TRAC achieves 96.55% success with SmolVLA, improving the original policy by percentage points and a clean-action baseline by 3.40 points. Further evaluations show lower cumulative prediction errors and gains across chunk lengths, with additional improvements on LIBERO-Plus, in an adapted GR00T policy, and on four real-robot tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Action Chunking, Trajectory-Level Supervision, Vision-Language-Action Models

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_o1EmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRAC, which adds two objectives to a flow-matching VLA on its clean-action estimates: a weighted prefix-consistency loss on accumulated translations and a local residual-structure prior on trend-removed action windows, so that relations within a chunk are supervised rather than only per-action error. On LIBERO with SmolVLA, TRAC improves a clean-action baseline by 3.4 points, with further results on various simulation and real-world tasks.

Critical strengths
  1. The prefix-consistency objective is correct and fully characterized. A positive-definite quadratic metric on translation errors whose unrestricted optimum is the unchanged conditional mean, which makes its effect transparent and its cost zero at inference.
  2. The evaluation is comprehensive, covering two backbones, a perturbed benchmark, a horizon sweep, held-out prediction error, and a real robot, and the prefix gain reproduces on the GR00T adaptation.
Critical weaknesses
  1. The evidence for TRAC cannot reliably separate its objectives from the clean-action reparameterization. The held-out error analysis, the SmolVLA results on LIBERO-Plus, and the real-robot results all compare TRAC with the original SmolVLA instead of the clean-action baseline.
  2. The proposed structural prior is not evident to justify its complexity. On SmolVLA, it has almost no impact on success, either alone or on top of prefix supervision. Moreover, the baselines in those comparisons use a single seed. Its gain on the GR00T adaptation is larger, but none of its settings in Table 7 are ablated, so it is unclear which parts of it matter.
Questions for authors
  1. Can you report Table 3 and Figure 4 for the clean-action baseline and for prefix-only, together with the per-step error in Appendix E.1? The question is whether the ratio of prefix error to per-step error falls under TRAC, since that ratio is what better composition means.
  2. What do the clean-action baseline and the prefix-only variant reach on LIBERO over three seeds, and what does the clean-action baseline reach on LIBERO-Plus and on the four real tasks?
  3. How are the settings within Table 7 chosen and on what data, and what fraction of windows incur a nonzero structural penalty at convergence?
  4. Which robot, cameras, control rate, and chunk length were used in Section 4.3, and is the SmolVLA column of Table 2 the velocity head or the clean-action head?
  5. The Loss Computation and Aggregate parts of Figure 2 do not render in any PDF viewer I tried, and copying the labels yields unmapped glyphs. Please describe their content and embed the figure text as fonts or outlines.
1 comment

Q1 Prefix error relative to per-step error

The prefix-to-per-step error ratio falls relative to clean-action, and most of this reduction comes from prefix supervision.

ConfigurationPer-step MAE ↓Prefix MAE ↓Prefix / per-step MAE ↓Prefix / per-step MSE ↓
Clean-action baseline0.24181.35465.603125.3733
Clean-action + prefix0.21811.12925.178120.0332
Full TRAC0.21771.12705.177820.0776

The comparison uses 40,000 chunks per configuration with H=15H=15. Relative to clean-action, full TRAC reduces the MAE ratio by 7.59% and the MSE ratio by 20.87%.

Prefix-only already gives nearly the same MAE ratio and a slightly lower MSE ratio than full TRAC. These diagnostics therefore attribute the reduction in relative error accumulation mainly to prefix supervision.

Appendix F.1, Tables 12–13 and Figures 6–8, reports the per-step and prefix errors. Appendix F.2, Table 14 and Figures 9–10, reports the ratios.

Q2 Three-seed and clean-action controls

The same three-seed means are 93.38% for clean-action, 95.50% for prefix-only, and 96.60% for TRAC.

ConfigurationSeed 1000Seed 2000Seed 3000Mean ± std
Clean-action baseline93.1593.6593.3593.38 ± 0.21
Prefix-only95.5095.2595.7595.50 ± 0.20
Full TRAC96.5596.4596.8096.60 ± 0.15

All entries are success rates (%); standard deviations are population standard deviations across the three training seeds. Appendix E.5 reports the results obtained using the same three training seeds. The structural term improves success over prefix-only by 1.05, 1.20, and 1.05 points.

The added LIBERO-Plus control is:

ModelLongGoalObjectSpatial
Clean-action baseline49.4936.4865.1358.56
TRAC49.2755.2772.3654.04

All entries are success rates (%). TRAC improves Goal and Object, while Long and Spatial are lower.

The real-robot comparison uses 20 trials per task:

TaskClean-action baselineTRAC
Plate insertion14/2017/20
Object-to-block placement17/2018/20
Block-to-cup placement15/2017/20
Two-object bowl placement15/2019/20
Overall61/80 (76.25%)71/80 (88.75%)

The robot results show an observed gain of 12.50 points over clean-action, with higher success on all four tasks.

Q3 Parameter selection and penalty activity

The Table 7 settings were developed incrementally on LIBERO and then kept fixed across models and datasets. There was no separate hyperparameter search.

At the final checkpoint, the structural penalty is nonzero in 40.29% of valid windows. We analyze 4,000 action chunks (1,000 batches of four), yielding 6,917 valid windows using the training noise/time sampling procedure (Appendix F.4). For the constraint shown in panel (a), 40.29% are active overall (2787/6917). Near the noise endpoint (0≤t≤0.10\le t\le0.1), 93 of 1,020 valid windows are active, or 9.12% within that region. These 93 windows account for 1.34% of all valid windows (93/6917). Our one-step sampler directly uses the clean-action prediction at t=0t=0 for execution, making its quality important for task success.

Q4 Robot setup and the Table 2 baseline

We use a Galaxea R1 robot with wrist-mounted RealSense D405 cameras and a head-mounted ZED camera. Actions are executed at 15 Hz, and the action chunk length is 15.

The original SmolVLA column in Table 2 uses the velocity prediction head. The expanded comparison separately identifies velocity, clean-action, and TRAC, with 52/80, 61/80, and 71/80 successes, respectively.

Q5 Figure 2 rendering and content

Figure 2 has been converted to address the rendering issue. Loss Computation calculates the squared difference between predicted and demonstrated cumulative translations at each prefix. Aggregate combines these errors using the specified prefix weights to form the prefix-consistency loss.

Reviewer_3QixOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRAC adds two more two more objectives in addition to flow-matching. A weighted prefix loss compares cumulative translation commands against demonstrations, penalizing how individual action errors combine across the chunk. A local structural prior detrends predicted continuous-action windows and penalizes residual singular-value magnitudes, with greater tolerance for dominant modes. Both constraints operate on clean-action estimates and add no inference-time module or integration steps. The paper evaluates SmolVLA on LIBERO and LIBERO-Plus, an adapted GR00T model on LIBERO-Plus, several execution horizons, held-out prediction errors, and four physical tasks.

Critical strengths
  1. The prefix loss is a meaningful objective, without incurring extra cost during sampling. It can be applied to clean-action estimates from either output parameterization, and Table 4 reports gains for both: 86.50% to 88.75% with a velocity head and 93.15% to 95.50% with a clean-action head. This is a simple, transferable contribution with an explicit connection to the training objective.
Critical weaknesses
  1. The prefix objective modifies both temporal coupling (when summing the action errors) and error individual weighting w_h. The current ablations do not distinguish these effects. Notice that squaring accumulated errors introduces cross-time interaction terms, while the prefix weights w_h also determine how strongly each action error is penalized. This paper does not isolated the benefit of cross-time interaction.

  2. Inspired by JiT, the authors adopt the same way to predict clean-action and train with prescribed velocity field. But the original task is for image generation, I wonder if this finding can directly transfer to VLA regime.

Questions for authors
  1. Follow the notation from B.2, can you compare the prefix objective with a point-wise control that replaces Q by diag(Q)? Everything else should stay the same.
  2. Could the authors provide the experiments where the FM objective is eps-loss or x-loss and the network is parametrizing eps or v?
1 comment

Q1 Prefix supervision versus the diagonal control

The full prefix objective improves average success by 1.10 points over the diagonal control.

ObjectiveLongGoalObjectSpatialAverage
Weighted prefix: QQ89.298.099.895.095.50
Pointwise control: diag⁡(Q)\operatorname{diag}(Q)83.296.0100.098.494.40

All entries are success rates (%). Both configurations use the clean-action head and exclude the structural prior. The control preserves the diagonal weights, masking, and loss normalization, removing only the cross-time interaction terms.

The comparison therefore separates the cross-time terms from per-action reweighting and supports their contribution in this setting.

Q2 Alternative losses and output parameterizations

We evaluated x-MSE and noise-MSE with direct velocity prediction, and additionally tested x prediction with noise-MSE. All configurations retain prefix supervision and exclude the structural prior. The alternative main losses replace velocity MSE.

Network outputMain MSE targetLongGoalObjectSpatialAverage
Velocity vvVelocity vv73.6090.6099.0091.8088.75
Velocity vvClean action xx76.2088.8096.2091.6088.20
Velocity vvNoise ϵ\epsilon69.2088.4096.8088.8085.80
Clean action xxVelocity vv89.2098.0099.8095.0095.50
Clean action xxNoise ϵ\epsilon77.0096.4099.6095.4092.10

All entries are LIBERO success rates (%). For velocity prediction, x-MSE and noise-MSE decrease average success by 0.55 and 2.95 percentage points, respectively. For x prediction, noise-MSE decreases it by 3.40 percentage points. Velocity matching therefore gives the highest average success for both output heads among these tested configurations. Direct noise prediction trained with x-MSE obtains 0.00% average success in our implementation.

Reviewer_X4QqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRAC augments flow-matching training with a weighted prefix-consistency loss on cumulative translation and a local residual-structure prior on action sequences. Both operate on clean-action estimates and leave the inference procedure unchanged. The paper evaluates LIBERO, LIBERO-Plus, an adapted GR00T policy, and four real-robot tasks.

Critical strengths
  1. The prefix objective has an intuitive connection to accumulated errors in relative translation, and intermediate prefixes expose deviations that can cancel at the final step.
  2. The clean-action baseline is an important controlled comparison: on LIBERO, the paper reports 93.15% for this baseline and 96.55% for the full method.
  3. Experiments examine objective components, different chunk lengths, a second model implementation, and physical manipulation tasks.
Critical weaknesses
  1. The original velocity baseline scores 86.50%, whereas the clean-action baseline scores 93.15%. Thus the reported 10.05-point gain over the original baseline cannot be attributed entirely to the proposed trajectory objectives. The 3.40-point controlled comparison deserves greater emphasis.
  2. Key ablations use different numbers of training seeds. In the matched seed-1000 comparison, the local structure term adds 1.05 points to prefix supervision. Repeated matched-seed comparisons are needed to assess whether this incremental benefit is stable.
  3. Summing normalized translation commands depends on the action representation and coordinate system. It does not directly capture rotational composition, contact dynamics, or closed-loop state error; claims about trajectory reliability should be bounded accordingly.
  4. The real-robot evaluation has only ten trials per task and compares against original SmolVLA. A matched clean-action baseline and uncertainty estimates are needed to isolate the effect of the proposed objectives in the physical setting.
  5. The local structure prior has several design choices. Simpler smoothness or difference penalties and sensitivity studies would clarify whether its specific formulation is necessary.
Questions for authors

See the weaknesses.

2 comments

Q1 Contribution beyond the clean-action baseline

The 3.40 points gain over the clean-action baseline is the comparison that isolates the added TRAC objectives: 93.15% to 96.55% in the seed-1000 run. The separate increase from 86.50% to 93.15% comes from the clean-action reparameterization.

To make this distinction clear, the abstract and introduction emphasize the comparison with the clean-action baseline. The method description also credits the existing reparameterization to JiT, separating it from the added supervision on cumulative translation and local action structure.

The additional controls in Appendix F.1 make this separation visible in the error analysis as well. Prefix MAE decreases from 1.3546 for clean-action to 1.1292 with prefix supervision and 1.1270 with full TRAC. These comparisons measure the changes from the added objectives while keeping the clean-action head.

Q2 Results using the same training seeds

Adding the structural prior improves the success rate for each of the three training seeds., with an average paired gain of 1.10 points over prefix-only.

ConfigurationSeed 1000Seed 2000Seed 3000Mean ± std
Clean-action baseline93.1593.6593.3593.38 ± 0.21
Prefix-only95.5095.2595.7595.50 ± 0.20
Full TRAC96.5596.4596.8096.60 ± 0.15

All entries are success rates (%). Standard deviations are population standard deviations across training seeds. The paired gains over prefix-only are 1.05, 1.20, and 1.05 points, giving 1.10 ± 0.07 points.

Two further full-TRAC runs, seeds 4000 and 5000, reach 96.05% and 96.25%. Across all five full-TRAC runs, the result is 96.42 ± 0.26%. Appendix E.5 reports the comparison using the same three seeds separately from the additional full-TRAC runs.

Q3 Action shaping and the scope of reliability

We agree that the scope of reliability needs to be clearer. TRAC shapes the learned action distribution through training losses on desired action properties. Prefix supervision guides command accumulation during execution, while the structural prior guides local coordination in clean-action estimates during denoising. We therefore clarify in the introduction and Section 4.4 that reliability of predicted action chunks refers to consistency with these cumulative and local properties in the chosen action representation.

Q4 Real-robot baseline and uncertainty

The expanded robot evaluation includes the clean-action baseline and 20 trials per task. TRAC reaches 71/80 successes, compared with 61/80 for clean-action.

PolicySuccessesOverall success rate95% Wilson CI
Velocity baseline52/8065.00%[54.08%, 74.55%]
Clean-action baseline61/8076.25%[65.86%, 84.24%]
TRAC71/8088.75%[79.98%, 93.97%]

The task-level results are:

TaskVelocity baselineClean-action baselineTRAC
Plate insertion11/2014/2017/20
Object-to-block placement13/2017/2018/20
Block-to-cup placement14/2015/2017/20
Two-object bowl placement14/2015/2019/20

The intervals are descriptive Wilson intervals over the 80 pooled trials for each policy. They do not measure variation across training seeds. TRAC's observed overall gain over clean-action is 12.50 points, with higher success on each task. These are the expanded results for Table 2 and Appendix E.3.

Q5 Structural-prior design and penalty choices

The simplified smoothing penalty reaches 95.60%, compared with 96.55% for full TRAC. It simplifies the residual penalty while retaining the local-reference framework.

SettingAverage success rate (%)
Full TRAC96.55
Direct residual smoothing95.60
Without detrending95.80
Without temporal whitening95.10
SVD replaced by an isotropic residual penalty95.50

Simplified smoothing penalty. It directly penalizes departures from the fitted local trend. A time-weighted Smooth L1 loss is applied to each entry of the whitened residual. The simplification is in the penalty: SVD grouping and adaptive thresholds are omitted.

Retained steps and their roles. The local trend defines the reference for smoothing. Temporal whitening determines how departures at neighboring steps are measured, and time weighting controls the penalty at each noise level. The detrending and whitening ablations support retaining these operations in the tested implementation.

The sensitivity settings and results are:

SettingAverage success rate (%)
Window length 594.70
Window length 1395.50
r=2r=295.00
Structural-loss weight 0.594.95
Structural-loss weight 296.20
Coefficient ratio β⊥/β∥=5\beta_\perp/\beta_\parallel=595.75
Coefficient ratio β⊥/β∥=10\beta_\perp/\beta_\parallel=10, default96.55
Coefficient ratio β⊥/β∥=20\beta_\perp/\beta_\parallel=2096.00

Using the same three seeds, adding the structural term to prefix supervision gives 1.10 ± 0.07 points of improvement.

Both penalties follow the same aim of encouraging local action structure. The full version distinguishes coordinated residual variation and penalizes only the excess above adaptive thresholds. All added operations are confined to training.

Reviewer_JK9ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Post-rebuttal update: Scores unchanged. See my response to the rebuttal for details.

This paper addresses two problems that arise when vision-language-action (VLA) models execute action chunks: accumulated displacement drift and high-frequency local jitter. The authors attribute both to the standard pointwise loss, which does not model dependencies across steps within a chunk. To address them, the paper proposes TRAC (Trajectory-Reliable Alignment Training). TRAC consists of three main components:

  1. Weighted prefix consistency loss (Lexec\mathcal{L}_{\text{exec}}). For the relative translation coordinates, this loss explicitly aligns the accumulated displacement over every prefix length with that of the expert demonstration. The prefix weights follow an inverse-harmonic form, wh∝1/(H−h+1)w_h \propto 1/(H-h+1), which is designed to balance the gradients between early and late steps.
  2. Local residual structural prior (Lstruct\mathcal{L}_{\text{struct}}). Within sliding windows, a quadratic polynomial fit removes the local motion trend. The resulting residual matrix is decomposed with SVD, and the non-dominant, higher-order incoherent modes are strongly suppressed with adaptive thresholds.
  3. Zero inference overhead. Both regularizers act only on the clean-action estimate during training. At inference, the original flow-matching architecture and the number of Euler steps stay unchanged.

The method is evaluated on the simulation benchmarks LIBERO and LIBERO-Plus. The evaluation also covers longer execution horizons (H∈[10,50]H \in [10, 50]), an accumulated-error analysis over 40,000 held-out action chunks, and four real-robot manipulation tasks. TRAC consistently outperforms the baselines. The average LIBERO success rate rises from 93.15% for the clean-action baseline to 96.55%, and the real-robot success rate rises from 65% to 80%.

Critical strengths

Simple and clear core idea. The method supervises the cumulative translation over every prefix of the action chunk. It is easy to implement, adds no inference overhead, and does not depend on a particular backbone. The expansion of Eq. (6) and the derivation in the appendix explain the mechanism clearly.

Measured theoretical claims. Appendix B.2 correctly notes that the unconstrained optimum is still the conditional mean. The prefix loss therefore only reweights how residual error is distributed under finite model capacity, and the authors do not overstate this point.

Consistent gains from the prefix loss in controlled comparisons. Table 4 shows that adding the prefix term improves success rate by about 2.3 points for both the velocity head and the clean-action head.

Broad evaluation coverage. The evaluation spans:

  • base performance on the four LIBERO suites (Spatial, Object, Goal, Long);
  • robustness to perturbations on LIBERO-Plus;
  • transfer across backbones (SmolVLA and StarVLA-GR00T);
  • a sweep of execution horizon H from 10 to 50;
  • a quantitative analysis of accumulated error;
  • real-robot experiments.
Critical weaknesses

Most of the headline gain comes from a change that is not this paper's contribution. Of the 10.05-point improvement highlighted in the abstract, 6.65 points come from switching to the clean-action output head (Yang et al., 2026). The proposed losses contribute only 3.40 points. The same pattern holds on GR00T: the largest single jump (75.23 → 81.95) comes from the authors' "adapted clean" variant, not from TRAC. The abstract and introduction should foreground the improvement over the clean-action baseline.

Several key experiments do not separate the effect of the parameterization change from that of the new objectives. The real-robot experiments (Table 2), the SmolVLA results on LIBERO-Plus, and the accumulated-error diagnostics (Fig. 4, Table 3) all compare TRAC against SmolVLA with the velocity head rather than against the clean-action baseline. This critical control is missing.

Weak mechanistic evidence. If the gains came from suppressing error accumulation, the improvement over the clean-action baseline should grow with H. Table 9 shows no such trend: the gains for H = 10–50 are +3.20, +2.85, +1.75, +3.60, and +2.70. The robustness at longer execution horizons therefore appears to come mainly from the clean-action parameterization itself.

The local residual structural prior is overly complex. It involves more than ten hyperparameters (Table 7) plus hand-designed adaptive thresholds (Eqs. 17–19). Yet the paper provides no sensitivity analysis, no ablation of its components (detrending, whitening, SVD decomposition, adaptive thresholds), and no comparison with simple alternatives such as finite-difference or jerk penalties.

Questions for authors

Attribution of the clean-action baseline. The clean-action baseline appears to be essentially the same mechanism as Action Manifold Learning (AML) in ABot-M0 (Yang et al., 2026). In both, the network directly outputs the clean action, which is then converted into a velocity for velocity matching. Could the authors make this attribution explicit in the method and experiment sections? Could they also clarify whether TRAC's clean-action implementation differs substantively from ABot-M0's?

Accumulated-error diagnostics for the clean-action baseline and the +prefix variant. Could the authors report the accumulated-error results in Table 3 and Fig. 4 for the clean-action baseline and the +prefix variant as well? Appendix E.1 defines a pointwise error E_point, but Table 3 reports only the prefix error. Could both be reported? This would show:

  • how much of the reduction in accumulated error comes from the prefix loss and how much from the clean-action parameterization;
  • whether TRAC specifically suppresses error accumulation or simply lowers the per-step error.

Clean-action baseline in the remaining experiments. Could the authors add the clean-action baseline to the real-robot experiments and to the SmolVLA experiments on LIBERO-Plus?

Hyperparameter selection and sensitivity.

  • How were the hyperparameters in Table 7 selected, and on what data?
  • Could the authors provide a sensitivity analysis for W,rW, r, the ratio β⊥,β∥,λstructβ_⊥, β_∥, λ_{struct}, and λexecλ_{exec}?
  • Do the GR00T, LIBERO-Plus, and real-robot experiments use exactly the same structural-prior hyperparameters as SmolVLA on LIBERO?
  • If any hyperparameters were tuned on test success rates, please state this explicitly.

Necessity of each component of the structural prior. Which of the four components of the structural prior are necessary: detrending, temporal whitening, SVD decomposition, and adaptive thresholds? How does the prior compare with a simple smoothness regularizer, such as a second-order difference or jerk penalty, at a matched regularization strength? If a simple regularizer achieves similar performance, most of these hyperparameters could be dropped.

4 comments

Q1 Clean-action attribution and implementation

We use the same x-pred-to-velocity conversion and velocity-matching loss as AML, following JiT's reparameterization. We clarify this attribution in the method description and add clean-action comparisons to identify the gains from our action objectives.

  • Section 3.1: We cite ABot-M0 and acknowledge that AML also evaluates x-pred with velocity matching.
  • Section 4.1: We add the clean-action and prefix-only error diagnostics, together with clean-action baseline comparisons on LIBERO-Plus and the real robot.

Differences from AML:

  1. Inference setting. Our x-pred results use one denoising step, while ABot-M0 reports four by default. The velocity-head comparisons in our paper use 10 steps.
  2. Relation between x-pred and v-pred. Following JiT, we use the x–v relation in both directions: x→vx\to v for flow-matching supervision, and v→xv\to x for expressing action properties in clean-action space.

Q2 Clean-action and prefix-only error diagnostics

Prefix supervision reduces cumulative error beyond the reduction in per-step error. The requested controls are reported in Appendix F.1, Tables 12–13 and Figures 6–8, using 40,000 chunks per configuration with H=15H=15.

ConfigurationPer-step MAE ↓Prefix MAE ↓Prefix / per-step MAE ↓
Velocity prediction0.28241.66875.9087
Clean-action baseline0.24181.35465.6031
Clean-action + prefix0.21811.12925.1781
Full TRAC0.21771.12705.1778

Adding prefix supervision to the clean-action baseline reduces per-step MAE by 9.79% and prefix MAE by 16.64%. Full TRAC reduces prefix MAE by 16.81% relative to clean-action.

Most of the additional cumulative-error reduction therefore comes from prefix supervision. The structural term changes these diagnostics only slightly. Appendix F.2, Table 14 and Figures 9–10, reports both MAE and MSE ratios: the lower ratios show that the improvement is more than a uniform reduction of individual errors.

Q3 Clean-action controls on the robot and LIBERO-Plus

The added clean-action baseline reaches 76.25% on the robot, compared with 88.75% for TRAC. Each policy is evaluated in 20 trials per task, or 80 trials in total.

TaskVelocity baselineClean-action baselineTRAC
Plate insertion11/2014/2017/20
Object-to-block placement13/2017/2018/20
Block-to-cup placement14/2015/2017/20
Two-object bowl placement14/2015/2019/20
Overall52/80 (65.00%)61/80 (76.25%)71/80 (88.75%)

TRAC improves the observed overall success rate over clean-action by 12.50 points, with higher success on all four tasks.

The LIBERO-Plus comparison is:

ModelLongGoalObjectSpatial
Velocity baseline46.5335.8966.2032.85
Clean-action baseline49.4936.4865.1358.56
TRAC49.2755.2772.3654.04

All LIBERO-Plus entries are success rates (%). Relative to clean-action, TRAC gains 18.79 points on Goal and 7.23 points on Object. The gains are therefore task-dependent.

Q4 Parameter selection and sensitivity

The Table 7 settings were developed incrementally on LIBERO, without a separate hyperparameter search. The finalized settings were then used unchanged for SmolVLA, GR00T, LIBERO-Plus, and the real-robot experiments.

Default TRAC reaches 96.55% average success. The sensitivity settings and results are:

SettingDefaultAlternative setting: success rate (%)
Window length WW95: 94.70; 13: 95.50
Rank rr12: 95.00
Coefficient ratio β⊥/β∥\beta_\perp/\beta_\parallel105: 95.75; 20: 96.00
Prefix-loss weight λexec\lambda_{\mathrm{exec}}10.5: 94.65; 2: 96.30
Structural-loss weight λstruct\lambda_{\mathrm{struct}}10.5: 94.95; 2: 96.20

Q5 Component ablations and penalty choices

The simplified smoothing penalty reaches 95.60%, while full TRAC reaches 96.55%. The component results are:

SettingAverage success rate (%)
Full TRAC96.55
Direct residual smoothing95.60
Without detrending95.80
Without temporal whitening95.10
SVD replaced by an isotropic residual penalty95.50

Simplified smoothing penalty. It directly penalizes departures from the fitted local trend. A time-weighted Smooth L1 loss is applied to each entry of the whitened residual. The simplification is in the penalty: SVD grouping and adaptive thresholds are omitted.

Retained steps and their roles. The local trend defines the reference for smoothing. Temporal whitening determines how departures at neighboring steps are measured, and time weighting controls the penalty at each noise level. The detrending and whitening ablations support retaining these operations in the tested implementation.

The relative-weight settings are:

Coefficient ratio β⊥/β∥\beta_\perp/\beta_\parallelAverage success rate (%)
595.75
10, default96.55
2096.00

The implementation has several steps, but its underlying aim is simple: preserve useful local action changes while constraining excessive irregular variation. The full penalty adds control over which deviations exceed the allowed range. All these operations are confined to training.

Thank you for the detailed rebuttal and the additional experiments. The clarified attribution to AML, the added clean-action controls, and the error diagnostics for the prefix loss address several of my concerns.

However, my main concern remains. The structural prior's gain appears only at the default configuration, and its ablated variants fall to roughly the clean+prefix level. TRAC also does not improve over the clean-action baseline on LIBERO-Plus Long and underperforms it on Spatial. The method therefore effectively reduces to clean-action plus a prefix loss, which I see as a modest, incremental contribution. I will keep my overall rating unchanged. I appreciate the authors' effort and hope these comments help strengthen the paper.

Thank you for your response and for clarifying the concerns that remain.

1. SmolVLA reproduction reveals a training-resolution limitation

Our LeRobot v0.3.2 reproduction produced the following LIBERO results:

SettingLongGoalObjectSpatialAverage
Initial reproduction62.0083.0076.0092.00≈78
512×512 observations72.0089.0098.0087.0086.50
  • SmolVLA resizes images to 512×512 to match its SmolVLM backbone. Low-resolution training observations, even after upsampling, provide less visual detail. Using higher-resolution observations brought our reproduction close to the published 87.3%, with Long improving from 62% to 72%. We therefore retained this setting throughout subsequent experiments.
  • For LIBERO-Plus, the demonstration-specific initial simulator states needed for replay were unavailable to us, preventing equivalent regeneration of higher-resolution training observations.
  • Our reproduction suggests that this limitation can leave SmolVLA insufficiently trained for the available visual inputs, substantially affecting Long. We therefore believe its Long/Spatial results provide an incomplete basis for judging the method’s overall value.

2. The complete module design transfers to GR00T

Our paper reports transferring the complete SmolVLA module design to Qwen2.5-VL–based StarVLA-GR00T, which also uses a flow-matching action head. Its latest results are:

ConfigurationLongGoalObjectSpatialAverage
Clean-action75.1078.3086.9784.1781.13
TRAC83.0985.9589.4091.9787.60

TRAC achieves higher success on all four suites, with an average difference of 6.47 percentage points over clean-action. The latest result is also nearly two percentage points above the corrected result currently reported in the paper. We will release the StarVLA-GR00T code and checkpoints.

3. This work developed alongside related research

An earlier version, combining our action-generation design with adaptive action length, was submitted to a robotics conference early this year, around the same period as AML. The feedback primarily concerned presentation, motivating us to reorganize the work.

Related adaptive-length studies appeared this year, and we identified closely related work shortly before the current submission deadline, leaving limited time for substantial restructuring. The current paper represents an important stage of this continuing research effort.

If these clarifications and new results address your remaining concerns, we would appreciate your reconsideration of the contribution and overall rating.

Comments on the paper

1 comment

General Response

Summary of Contributions and Additional Evidence

We sincerely thank all reviewers for their thoughtful and constructive feedback! We are encouraged by their recognition of the prefix objective's clear motivation, transferable design, and zero inference overhead.

Our central contribution is to supervise how actions accumulate and vary together within an action chunk. Similar per-step errors can produce different cumulative deviations. TRAC directly supervises cumulative translation against demonstrations and introduces a complementary constraint on local action variation. Both objectives operate during training, preserving the inference procedure, including one-step denoising in our clean-action setting. Prefix supervision improves both velocity and clean-action heads, with evaluations covering two VLA backbones.

The added controls establish gains beyond clean-action reparameterization. We add clean-action comparisons to the error analysis, LIBERO-Plus, and real-robot evaluation, and repeat the LIBERO comparisons using the same three training seeds.

EvaluationClean-action baselineTRACImprovement
LIBERO success, three-seed mean ↑93.38%96.60%+3.22 percentage points
Robot success, 80 trials per policy ↑76.25%88.75%+12.50 percentage points
Prefix / per-step MAE ↓5.60315.1778−7.59%

The error analysis uses 40,000 action chunks per configuration. The lower error ratio shows that cumulative error improves beyond a uniform reduction in individual prediction errors. Most of this improvement comes from prefix supervision. The structural prior further increases task success over prefix-only in all three training seeds, averaging 1.10 percentage points. On the robot, TRAC achieves higher observed success on all four tasks.

We also provide component ablations, sensitivity tests, and a simpler smoothing comparison to clarify the structural design. Section 3.1 acknowledges the shared reparameterization with JiT and AML. The error analyses appear in Appendices F.1–F.2, and the comparisons using the same training seeds are reported in Appendix F.3 (Table 15).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.