acceptodds
Under review as a conference paper at ICLR 2027

TrajScale-VLA:Trajectory Aligned Scale Autoregressive Action Generation for Vision-Language-Action Models

Abstract

Autoregressive vision-language-action models (VLAs) rely on action tokenization to represent continuous trajectories as discrete sequences for prediction. A key challenge is that fine-grained action tokens preserve control precision but lead to long autoregressive sequences, while hierarchical tokenization can shorten decoding but does not inherently determine how trajectory information should be organized across scales. Existing tokenizers impose an ordering or multi-scale grouping on action tokens, but do not explicitly align each token group with the trajectory structure it is expected to represent. To address this challenge, we propose TrajScale-VLA, which uses trajectory supervision to guide the learning of an action token hierarchy, with early predictions capturing coarse motion and subsequent predictions recovering local detail. We first train a multi-scale residual tokenizer with a multi-scale trajectory refinement (MTR) objective that aligns cumulative reconstructions with targets at progressively finer temporal resolutions. This supervision assigns each scale a distinct reconstruction task within the action hierarchy. We then train a policy to predict the resulting token groups autoregressively across scales and in parallel within each scale, reducing the number of sequential decoding steps. TrajScale-VLA outperforms all compared StarVLA baselines, achieving success rates of 98.4% on LIBERO and 59.8% on RoboCasa-GR1. On real-world USB Pick-and-Insert, which requires coordinated transport and fine alignment, it surpasses the best baseline in each setting by 13.3 and 26.7 percentage points under in-distribution and out-of-distribution conditions. These results highlight the importance of organizing action representations through trajectory supervision for autoregressive robot control.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Action Tokenization, Robotic Manipulation

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_7T4zOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TrajScale-VLA brings VAR-style next-scale prediction to action tokenization. The core contribution is MTR. The cumulative reconstruction after the first k scales, D(F̂_k), is supervised against a scale-matched target. That target is built by average-pooling the expert chunk to s_k steps and linearly interpolating back to H (Eq. 3). The intermediate scales are weighted uniformly with λ_mtr/(K−1), and the final scale is supervised against the full chunk. In Stage 2, a Qwen3-VL-4B policy predicts the token groups scale by scale, with all tokens inside a scale predicted in parallel (K sequential rounds).

Results:

  • 98.4% on LIBERO and 59.8% on RoboCasa-GR1, against baselines mostly copied from the StarVLA reports and other papers.
  • A same-backbone ablation (Table 3) comparing w/o MTR, OAT-style and MINT-style intermediate supervision: on RoboCasa, MTR scores +9.0 pp over w/o MTR and +6.5 pp over MINT-style.
  • A real-robot USB pick-and-insert task (30 trials per condition): 90.0 / 76.7% ID/OOD vs 76.7 / 50.0% for the best variant.
Critical strengths
  1. The within-backbone ablation shows that the choice of intermediate target affects policy learnability, even when reconstruction is nearly identical. On RoboCasa-GR1 (1,200 trials), MTR scores 59.8% vs 50.8% without MTR and 53.3% for MINT-style. Final reconstruction MSE barely changes (3.58 vs 3.63 ×10⁻⁶). The Fig. 5/6 diagnostics are consistent with a coarse-to-fine organization: the peak correction frequency progresses 1→2→4→8 under MTR but stays at 1→1→1→8 for the other variants.

  2. The idea is simple and adds no inference cost. MTR changes only Stage-1 targets and is easy to adopt in any residual-scale action tokenizer.

Critical weaknesses
  1. Lack of discussion of efficiency. The paper motivates TrajScale by the observation that fine-grained action tokens produce long autoregressive decoding steps in prior AR-VLAs, yet it never explicitly discusses the efficiency of its own design.
  • From the reported configuration (G = 16 tokens per temporal position, dyadic scales), the token count doubles at every scale. LIBERO (H = 8, scales {1,2,4,8}) uses 16+32+64+128 = 240 tokens per chunk, i.e., more than the 8×7 = 56 raw action values. RoboCasa-GR1 (H = 16, scales {1,…,16}) uses 16+32+64+128+256 = 496 tokens per chunk.
  • The paper's implicit argument is fewer sequential rounds (K = 4–5). However, each round must still emit up to 256 tokens in parallel, conditioned on a growing context. Whether this is actually faster than flat AR or compact tokenizers is unclear. I believe the authors need to discuss and quantify the efficiency advantage of TrajScale, e.g., per-chunk latency, throughput and control frequency versus flat AR with the same tokenizer and versus StarVLA-FAST/OFT.
  1. Mismatch between motivation and mechanistic analysis. MTR is motivated as a reconstruction objective, but Table 3 does not support reconstruction as the explanation for the success gain. On RoboCasa-GR1, w/o MTR has a lower final MSE (3.63 vs 5.18/5.13 ×10⁻⁶) and a lower intermediate MSE (6.75 vs 10.0 ×10⁻³) than both OAT-style and MINT-style, yet a lower success rate (50.8 vs 51.6/53.3%). The authors acknowledge that reconstruction "does not fully characterize downstream policy performance", but offer no further analysis of why MTR works. Their remaining diagnostic, intermediate MSE, is computed against MTR's own training targets and so is favourable by construction. Prior work (e.g., ActionCodec, cited by the paper) suggests that the structure of the token space matters more than reconstruction fidelity; analyses such as per-scale Stage-2 token prediction accuracy, codebook usage, or latent neighbourhood structure are needed before the +9 pp can be attributed to the proposed mechanism.
Questions for authors
  1. Latency. Is the scale-autoregressive head a separate module, or are the previous scales' tokens fed back through the VLM? What is the per-chunk latency at batch size 1, and the control frequency, compared with StarVLA-FAST/OFT/π and with flat AR using the same tokenizer?
  2. Inductive bias of the MTR target. MTR fixes in advance what each scale must represent: an average-pooled, linearly interpolated trajectory at a dyadic resolution. Does this not inject a strong inductive bias that constrains how the tokenizer allocates information across scales? How was average pooling + linear interpolation arrived at? Were alternatives tried, e.g., other downsampling/upsampling kernels, low-pass DCT/wavelet targets, keyframe-based targets, or learned/adaptive targets?
6 comments

Thank you for your thoughtful review and for recognizing MTR's simplicity and our controlled ablations.

1. Weakness 1: Efficiency.

Thank you for raising this concern. We report ms/chunk, as in FAST, at batch size 1 on an H100. Timing includes preprocessing, the VLM, action generation and postprocessing, excluding network and robot execution. See Table 5 of the revised manuscript.

MethodLIBERORoboCasa-GR1
TrajScale-VLA73.6178.58
Flat AR257.78414.08
StarVLA-π (4 steps)214.92218.94
StarVLA-GR00T (4 steps)80.4780.03
StarVLA-GR00T (10 steps)136.14133.19
StarVLA-OFT49.0648.14
StarVLA-FAST347.77918.25

TrajScale is 3.50–5.27× faster than Flat AR, 4.72–11.69× faster than FAST and 2.79–2.92× faster than StarVLA-π. It is close to 4-step GR00T and faster than 10-step GR00T; OFT remains faster with its MLP head. As in VAR, tokens within each scale are predicted in parallel: 240/496 codes need only 4/5 rounds (LIBERO/RoboCasa).

2. Weakness 2: Mechanistic analysis.

Thank you for highlighting this issue. Similar average MSE need not imply similar control success: small alignment or insertion errors may cause failure, while larger deviations in less sensitive segments may have little effect.

ActionCodec examines token structure and prefix errors beyond reconstruction. We examine coarse prediction and prefix use.

Experiment 1: Predictability.

On RoboCasa-GR1, TF@s is token accuracy (%) with expert prefixes. FR MSE uses predicted tokens throughout; Recon is from Table 3.

MethodTF@1TF@2TF@4TF@8TF@16Recon MSE (×10⁻⁶)FR MSE (×10⁻³)
w/o MTR62.24837.87041.14842.76350.4183.6322.145251
OAT-style66.83838.27739.62643.58252.1475.1782.071389
MINT-style70.21539.68538.19445.13754.7625.1272.027516
MTR76.45740.80835.45746.40856.1863.5831.903602

MTR has higher TF accuracy at the first two scales and 11.3% lower FR MSE than w/o MTR.

Experiment 2: Prefix use.

We replace the first one or two scales with expert tokens (GT1/GT2), then regenerate later scales (+ regen) or keep their original predictions (only). Normal is FR; MSE is in normalized action space (×10⁻³).

MethodNormalGT1 + regenGT2 + regenGT1 onlyGT2 only
w/o MTR2.1452511.5574280.5796742.4391032.245905
OAT-style2.0713891.3360750.2799942.2798772.044022
MINT-style2.0275161.1881720.2641112.2058682.032551
MTR1.9036020.2037960.1163621.9138111.864604

With the GT prefix fixed within each model, regeneration lowers mean MSE by 89.4%/93.8% for MTR (GT1/GT2), versus 36.1%/74.2% for w/o MTR, compared with retaining later tokens.

Together, TF supports coarse predictability, the prefix comparison shows how later scales use coarse information, and FR shows lower action error with policy-predicted tokens, consistent with the control gains. These analyses appear in revised Table 4 and Appendix C.1 (Tables 8–9).

3. Question 1: Architecture and control frequency.

Thank you for asking. The VLM encodes images and language once. Earlier-scale embeddings enter a separate action head; the tokenizer decoder runs once. We clarify this in revised Section 3.3.

W1 reports batch-one latencies. TrajScale makes 13.6/12.7 chunk queries/s (LIBERO/RoboCasa), executing 8/12 actions per query; StarVLA-π (4 steps) makes 4.65/4.57 queries/s.

4. Question 2: Target inductive bias.

Thank you for this question. MTR intentionally imposes a temporal prior: intermediate targets progress from coarse to fine, and the final target is the full trajectory.

Table 3 examines the benefit of scale-matched supervision. Here we retain this principle and vary how the targets are constructed.

We chose a local, parameter-free construction for its kinematic meaning: window means summarize motion trends, and linear interpolation connects them over time. Appendix A shows that means are L2-optimal constant summaries and interpolation minimizes first-derivative energy through fixed mean knots. W2 examines its role in policy prediction.

With both training stages otherwise unchanged, we tested nearest upsampling and low-pass DCT (retain s_k coefficients including DC, then inverse DCT). See revised Appendix C.2 (Table 10). Success (%):

TargetLIBERORoboCasa-GR1
AvgPool + linear (MTR)98.459.8
AvgPool + nearest98.057.9
Low-pass DCT97.458.2

Both alternatives exceed w/o MTR (96.2/50.8, Table 3), supporting scale-matched supervision beyond one construction. AvgPool + linear's further gains suggest that construction also matters. Nearest can introduce jumps; DCT uses a global basis. Local summaries with gradual transitions may explain our higher success.

We hope these responses address your concerns and welcome further discussion. Thank you again.

Dear Reviewer 7T4z,

Thank you again for your thoughtful review and constructive feedback. We added batch-one H100 latency comparisons against Flat AR and StarVLA baselines, and clarified the independent action head, single VLM encoding, and single decoder pass. We also supplemented reconstruction diagnostics with teacher-forced per-scale token accuracy, free-running action MSE, and prefix-intervention experiments, providing evidence for coarse-scale predictability and effective prefix use during refinement. Finally, we clarified the intentional temporal prior and the rationale for average pooling with linear interpolation, supported by comparisons with nearest upsampling and low-pass DCT targets.

We would greatly appreciate it if you could let us know whether these clarifications and additional results address your concerns. We would be happy to discuss any remaining questions.

Best regards, The Authors

Thank you for the responses and additional experiments. The clarifications on the model architecture and VAR-style decoding are very helpful for understanding TrajScale-VLA's inference pipeline, and the new target ablations in Q2 are also informative.

However, for Weakness 2, reporting additional metrics (TF@s, FR MSE, and prefix-replacement MSE) is not a mechanistic explanation; it mainly replaces Stage-1 MSE with offline metrics that correlate better with task success. Moreover, Experiment 2 reveals an anomaly: although ground-truth coarse prefixes (GT1/GT2 + regen) substantially reduce the error of later scales, comparing MTR's Normal MSE (1.904) with GT1 + regen (0.204) shows that ~89% of MTR's test-time error comes from the first-scale prediction (s = 1, 16 parallel tokens), while finer scales (s = 2 … 16) cannot correct errors made at s = 1. Thus, the dominant error source lies in the first scale rather than multi-scale refinement, leaving MTR's Normal FR MSE (1.904) still very close to MINT-style (2.028). How does this reconcile with the +6.5–9.0 pp closed-loop success gain?

This also connects to Q1. Since the paper frames TrajScale-VLA as an autoregressive VLA, I expected action tokens to be generated by the VLM itself (as in OpenVLA, FAST, or OAT), rather than delegated to a separate small action head after encoding once. Why choose a separate lightweight head over native VLM generation, and given that ~89% of MTR's error comes from Scale 1 (which captures coarse motion intent), wouldn't letting the VLM itself predict the tokens (or at least Scale 1) be more effective?

Reviewer_o5UyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TrajScale-VLA learns a hierarchy of discrete action tokens whose intermediate reconstructions are supervised at different temporal resolutions. Its main contribution is the multi-scale trajectory refinement objective, MTR. At each non-final scale, demonstrated actions are average-pooled and interpolated back to the original action horizon to create a scale-specific target. A shared decoder reconstructs actions from cumulative residual-quantized representations. After the tokenizer is trained, it is frozen, and a policy predicts token groups sequentially across scales and in parallel within each scale.

The paper reports 98.4% success on LIBERO and 59.8% on RoboCasa-GR1. Within the proposed architecture, MTR improves success over removing intermediate supervision by 2.2 and 9.0 percentage points, respectively. It also outperforms full-trajectory and frequency-domain supervision variants. A real-world USB grasp-and-insert task provides additional evidence, with 90.0% success in distribution and 76.7% under shifted initial configurations.

Critical strengths

Giving each scale its own trajectory target is a clear and sensible idea. The supervision ablation is the strongest part of the paper, since it provides reasonably direct evidence that this choice matters for downstream control.

The evaluation covers two simulation benchmarks and a real-world insertion task. The results on RoboCasa-GR1 and the physical robot are particularly encouraging and give the proposed representation some practical support beyond the high LIBERO score.

Critical weaknesses

The broader comparison is less convincing than the supervision ablation. Several baseline results come from different training setups, while the OAT-style and MINT-style variants do not represent the complete original methods. The lack of results across training seeds also makes the smaller improvements difficult to judge.

The analysis shows that the tokenizer learns to follow the proposed targets, but is less clear about why this helps the policy. Most diagnostics use tokens encoded from demonstrations. I would be more convinced by evidence that the same structure helps when tokens are predicted by the policy and earlier mistakes can affect later scales.

Questions for authors

Were the policy training settings and loss weights matched across the supervision variants?

How consistent are the gains across different training seeds?

Does the hierarchy still help when using policy-predicted tokens rather than demonstration-encoded tokens?

6 comments

Thank you for your careful review and for recognizing the value of the scale-specific targets, the supervision ablation, and the simulation and real-world evaluations.

1. Weakness 1: Baseline comparisons and consistency across training seeds.

Thank you for recognizing the value of the supervision ablation. We appreciate the opportunity to clarify the broader comparisons. Following common benchmarking practice, Tables 1–2 include results from the StarVLA reports and other original papers (sources in Appendix B.3 of the revised manuscript). These systems differ in architecture, pretraining, hyperparameters and implementation pipelines, making a single unified training setup difficult. Our implementation follows StarVLA's data and evaluation pipeline, while Table 3 provides controlled comparisons within our framework.

We agree that these variants are controlled adaptations rather than complete reproductions of the original systems. For the supervision ablation, we adapted the intermediate supervision principles of OAT and MINT to our shared StarVLA framework: reconstructing the full trajectory from partial representations for OAT-style supervision, and matching the full DCT spectrum at successive scales for MINT-style supervision. Sharing the tokenizer and policy architecture lets us assess scale-matched supervision without architectural differences. These controlled adaptations, labeled “-style,” complement the system-level comparisons.

We also evaluated the original OAT and MINT implementations. Success (%):

MethodLIBERORoboCasa-GR1
OAT97.250.8
MINT98.255.3
TrajScale-VLA98.459.8

We agree that results across training seeds are important for judging smaller improvements. We repeated the supervision ablation with three training seeds and report mean success rate ± sample standard deviation (%) below.

MethodLIBERORoboCasa-GR1
w/o MTR96.40±0.7250.07±0.64
OAT-style97.40±0.5350.97±0.65
MINT-style97.73±0.5054.33±0.91
MTR98.33±0.3160.10±0.36

2. Weakness 2: How the hierarchy helps policy prediction and handles earlier errors.

Thank you for this suggestion. ActionCodec examines token structure and prefix errors beyond reconstruction. We examine coarse prediction and prefix use.

Experiment 1: Policy-predicted tokens.

On RoboCasa-GR1, FR MSE measures action error when the policy predicts every scale using its own earlier tokens. TF@s is token accuracy (%) given expert tokens at earlier scales; Recon is from Table 3.

MethodTF@1TF@2TF@4TF@8TF@16Recon MSE (×10⁻⁶)FR MSE (×10⁻³)
w/o MTR62.24837.87041.14842.76350.4183.6322.145251
OAT-style66.83838.27739.62643.58252.1475.1782.071389
MINT-style70.21539.68538.19445.13754.7625.1272.027516
MTR76.45740.80835.45746.40856.1863.5831.903602

MTR predicts the first two scales more accurately and has 11.3% lower FR MSE than w/o MTR.

Experiment 2: Prefix use.

We replace the first one or two scales with expert tokens (GT1/GT2), then regenerate later scales (+ regen) or keep their original predictions (only). Normal is FR; MSE is in normalized action space (×10⁻³).

MethodNormalGT1 + regenGT2 + regenGT1 onlyGT2 only
w/o MTR2.1452511.5574280.5796742.4391032.245905
OAT-style2.0713891.3360750.2799942.2798772.044022
MINT-style2.0275161.1881720.2641112.2058682.032551
MTR1.9036020.2037960.1163621.9138111.864604

With the GT prefix fixed within each model, regeneration lowers mean MSE by 89.4%/93.8% for MTR (GT1/GT2), versus 36.1%/74.2% for w/o MTR, compared with retaining later tokens.

Together, TF supports coarse predictability, the prefix comparison shows how later scales use coarse information, and FR shows lower action error with policy-predicted tokens, consistent with the control gains. The revised manuscript summarizes these results in Table 4, with full results in Appendix C.1 (Tables 8–9).

3. Question 1: Matched policy training settings and loss weights.

Yes, thank you for checking. Within each benchmark, all variants share the policy backbone, data and training settings (Table 7 of the revised manuscript). Loss coefficients are matched where applicable; only intermediate supervision changes, and w/o MTR omits that term.

4. Question 2: Consistency of gains across training seeds.

As reported in W1, we repeated the ablation with three training seeds. MTR has the highest mean success on both benchmarks, with low variation across runs.

5. Question 3: Effectiveness with policy-predicted tokens.

Yes. W2 reports lower action error with policy-predicted tokens and effective use of coarse prefixes.

We hope these responses address your concerns and welcome further discussion. Thank you again.

Dear Reviewer o5Uy,

Thank you again for your careful review and constructive feedback. We clarified the purpose of the OAT-style/MINT-style comparisons and confirmed matched policy training settings and applicable loss weights. We also reported evaluations of the original OAT and MINT implementations and three-seed ablations on LIBERO and RoboCasa-GR1, showing consistent gains. To address how the hierarchy helps policy prediction, we added free-running action-error evaluation and prefix interventions, providing evidence of improved policy-generated action quality and effective use of coarse information during refinement.

We would greatly appreciate it if you could let us know whether these clarifications and additional results address your concerns. We would be happy to discuss any remaining questions.

Best regards, The Authors

Dear authors,

I noticed that the main text does not cite Figure 1. Without a citation or explanation, readers cannot understand the insight this figure offers.

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

I pasted the wrong paper. This paper proposes TrajScale-VLA, which introduces multi-scale trajectory refinement (MTR) to improve action tokenization in autoregressive VLA models. The main idea is to supervise intermediate reconstructions using trajectory targets at different temporal resolutions. The method achieves 98.4% on LIBERO and 59.8% on RoboCasa-GR1, with further improvements demonstrated in real-world USB insertion experiments.

Critical strengths

The paper addresses a relevant problem in action tokenization. The proposed scale-matched supervision is simple and well motivated.

The experimental results are strong, particularly on RoboCasa-GR1. The ablation studies provide useful evidence for the effectiveness of MTR.

Critical weaknesses
  1. The method mainly combines existing multi-scale quantization and autoregressive generation with a new supervision objective. The technical contribution is somewhat incremental.

  2. The improvement on LIBERO is relatively small, and the real-world experiments involve only 30 trials per setting. Multiple runs and confidence intervals would make the results more convincing.

Minor issue: The experiments focus on a single USB insertion task. More diverse manipulation tasks would help demonstrate generalization.

Questions for authors

Please refer to the weakness part.

3 comments

Thank you for your careful review and for recognizing the motivation behind scale-matched supervision, the RoboCasa-GR1 results, and the value of the ablation studies.

1. Weakness 1: Technical contribution.

Thank you for raising this concern. Our contribution is a supervision principle for hierarchical action tokenization: each cumulative representation is trained to reconstruct a trajectory at a specified temporal resolution. TrajScale-VLA builds on existing multi-scale quantization and scale-autoregressive generation. In our controlled OAT-style variant, intermediate reconstructions share the full-trajectory target; in the MINT-style variant, they match the full DCT spectrum. MTR assigns progressively finer trajectory targets to successive cumulative reconstructions, explicitly specifying what each should recover, from coarse motion summaries to local action variations.

Table 3 isolates the effect of this supervision while keeping the tokenizer and policy architectures fixed. MTR improves success from 96.2% to 98.4% on LIBERO and from 50.8% to 59.8% on RoboCasa-GR1, compared with removing intermediate supervision. It also outperforms the OAT-style and MINT-style supervision variants under the same architectures.

Table 10 tests whether this benefit extends beyond our default construction, with other settings unchanged. Average pooling with nearest-neighbor upsampling and low-pass DCT both outperform removing intermediate supervision on both benchmarks. On RoboCasa-GR1, they achieve 57.9% and 58.2%, respectively, versus 50.8% without intermediate supervision and 59.8% with average pooling and linear interpolation. This supports scale-matched supervision across the tested target constructions.

We also examine how the representation supports precise control. In an offline analysis of windows from RoboCasa-GR1 training demonstrations using policy-generated tokens, Table 12 shows that MTR reduces action MSE relative to MINT-style by 32.4% during task-critical intervals, compared with 6.1% overall; error outside those intervals is slightly higher. The improvement is therefore concentrated around interactions requiring precise control. This helps explain why a modest change in overall action error can accompany a larger closed-loop success gain. These results support scale-matched supervision and connect its benefits with task-critical phases.

2. Weakness 2: LIBERO gains and consistency across training seeds.

Thank you for this suggestion. We agree that the improvement on LIBERO is modest and that repeated training runs are important for assessing it. We repeated the supervision ablation with three training seeds, keeping the policy architecture, data, and training settings matched across variants within each benchmark. The table reports mean success rate ± sample standard deviation (%).

MethodLIBERORoboCasa-GR1
w/o MTR96.40±0.7250.07±0.64
OAT-style97.40±0.5350.97±0.65
MINT-style97.73±0.5054.33±0.91
MTR98.33±0.3160.10±0.36

MTR achieves the highest mean success on both benchmarks, with low variation across the three runs. Compared with MINT-style, its mean success is 0.60 percentage points higher on LIBERO and 5.77 points higher on RoboCasa-GR1. These runs support consistent gains, with a smaller margin on LIBERO.

For the real-world evaluation, we report 95% Wilson confidence intervals for all methods, with 30 trials per method in each setting. Entries are success rates [95% CI] (%).

MethodIDOOD
OAT-style66.7 [48.8, 80.8]50.0 [33.2, 66.8]
MINT-style76.7 [59.1, 88.2]46.7 [30.2, 63.9]
MTR90.0 [74.4, 96.5]76.7 [59.1, 88.2]

These intervals reflect the uncertainty from the limited trial count.

3. Minor issue: Diversity of real-world tasks.

Thank you for this suggestion. We chose USB Pick-and-Insert because it combines pick-and-place motion with precise connector alignment and insertion. The robot must grasp and transport the connector, then make small, accurate adjustments to insert it successfully. This demanding task tests both transport and fine manipulation. The success rates of 90.0% in distribution and 76.7% under shifted initial configurations are encouraging in this setting.

We agree that more diverse real-world tasks would strengthen the evaluation. The remaining discussion period does not leave enough time to complete and validate another real-robot experiment. We will add one additional real-world manipulation task in a subsequent revision to assess whether the observed benefits carry over to a different task.

We hope these responses address your concerns and welcome further discussion. Thank you again for your time and constructive feedback.

Thanks for your response. Most of my concerns have been well addressed. Therefore, I will keep my positive rating.

Thank you for your constructive feedback and continued support of our work. We appreciate your time and thoughtful consideration of our responses.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

General Response and Summary of Revisions (Submission 22700)

Dear Area Chair,

We thank the Area Chair for their time and consideration, and all reviewers for their careful evaluation and constructive feedback. We have addressed the main concerns through clarifications and additional experiments, summarized below.

1. Mechanistic explanation of MTR’s benefits

(Reviewer 7T4z: W2, Q2 and follow-up; Reviewer o5Uy: W2, Q3; Reviewer GBxr: W1)

(Table 4; Appendices C.1–C.3) We clarified the role of scale-matched supervision and added target-construction ablations, policy-prediction analyses, and prefix controls. Closed-loop RoboCasa-GR1 success rises from 53.4% to 59.8% when finer scales are included. An offline analysis of training-demonstration windows shows 32.4% lower task-critical MSE than MINT-style, versus 6.1% overall. Together, these results support the contribution of refinement to control and help explain gains beyond average reconstruction accuracy.

2. Inference speed and architecture

(Reviewer 7T4z: W1, Q1 and follow-up)

(Section 3.3; Table 5; Appendix C.3, Table 13) We clarified the single VLM encoding and separate scale-autoregressive action head, and added batch-one H100 latency comparisons. TrajScale-VLA takes 73.61/78.58 ms per chunk on LIBERO/RoboCasa-GR1, achieving 3.50–5.27× speedups over Flat AR and 4.72–11.69× over StarVLA-FAST; OFT remains faster. Compared with direct VLM generation at all scales, the separate head achieves similar success while being 2.40×/2.73× faster, supporting this architecture choice.

3. Controlled comparisons and training robustness

(Reviewer o5Uy: W1, Q1, Q2; Reviewer GBxr: W2)

(Table 3; Appendix B.3; additional results in the individual responses) We clarified the distinction between system-level baselines and OAT-style/MINT-style supervision adaptations, confirmed matched training settings within the controlled ablation, and reported evaluations of the original OAT and MINT implementations. Three-seed supervision ablations show that MTR achieves the highest mean success on both benchmarks: 98.33 ± 0.31% on LIBERO and 60.10 ± 0.36% on RoboCasa-GR1, supporting the consistency of the gains.

Reviewers o5Uy and GBxr have both maintained their positive ratings of 3 (Weak Acceptance). Reviewer 7T4z confirmed that our follow-up resolved their remaining questions and raised their rating from 2 to 3. We thank all three reviewers for the constructive discussion and hope the Area Chair will take these assessments into consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.