acceptodds
Under review as a conference paper at ICLR 2027

Revealing State Aliasing in Visual Representations of Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models are a promising framework toward general-purpose robot manipulation that leverages the rich prior knowledge of pretrained vision-language models for diverse tasks. However, they often struggle with tasks that require high-precision control. Our analysis across various visual encoders in VLA shows that a key source of the limitation is state aliasing, where the visual encoder fails to distinguish fine-grained states that require different actions even though the states are visually similar. Motivated by this observation, we propose a simple yet effective auxiliary inverse dynamics objective for VLA training. By predicting intervening actions between two frames in a demonstration trajectory, our auxiliary task encourages the visual encoder to preserve action-relevant state distinctions. Our approach requires no additional manual annotations and the inverse dynamics prediction is used only during training, introducing no additional computational overhead at inference. Our experiments on LIBERO, CALVIN ABCD, and SimplerEnv demonstrate consistent improvements in manipulation performance over standard VLA training across diverse VLA architectures. Representation-level analyses further show reduced state aliasing and more accurate action prediction in the learned visual representations, consistent with the trend observed across visual encoders.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
vision-language-action models, robot manipulation, visual representation learning, state aliasing, inverse dynamics, robot learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KayMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates state aliasing in VLA visual representations, where fine-grained state differences may not be sufficiently preserved for accurate robot control. The authors analyze the relationship between state-difference prediction and action prediction across pretrained visual encoders, and propose an auxiliary inverse dynamics loss to improve state-sensitive representations. Experiments on LIBERO, CALVIN, and SimplerEnv demonstrate improvements across several VLA architectures.

Critical strengths
  1. The paper identifies an interesting and important problem: whether pretrained VLM representations preserve sufficiently fine-grained information for robot control. The motivation is clear and worth investigating.

  2. The proposed method is simple and practical, requiring no additional annotations or inference-time overhead. The substantial improvement on LIBERO is encouraging, and the representation-level analyses provide additional supporting evidence.

Critical weaknesses
  1. Limited probing methodology. Using a simple MLP on frozen feature differences may underestimate Qwen3-VL's capabilities. The authors should evaluate stronger probes, including the full VLM with its LLM decoder, before concluding that the visual encoder lacks fine-grained state information.

  2. Weak baseline. The LIBERO baseline achieves only 53.9%, far below recent VLAs such as StarVLA, VLAct, and is trained for only one epoch per stage. The authors should validate their method on a stronger, well-trained baseline to demonstrate its effectiveness.

Questions for authors
  1. Have the authors considered evaluating state discrimination using the full Qwen3-VL model with its LLM decoder? Could the relatively poor probing performance reflect limitations of the probe architecture rather than missing information in the visual representation? I believe VLM models can show the difference between the two images.

  2. Why is the LIBERO baseline substantially weaker than recent VLA methods? Can the authors provide training curves and evaluate the proposed method on a stronger baseline, such as StarVLA, VLAct, under matched training conditions?

  3. Why should inverse dynamics supervision be particularly effective at reducing state aliasing? How does it compare with direct state-difference prediction or other auxiliary representation-learning objectives?

6 comments
AuthorsOpenReview ↗

Response to Reviewer KayM

Thank you for your thoughtful comments and suggestions. We provide clarifications and additional results addressing your concerns below.


W1 / Q1. Probe limitations and full-VLM evaluation

You ask, “Have the authors considered evaluating state discrimination using the full Qwen3-VL model with its LLM decoder?” Following your suggestion, we have now extended our state-difference analysis to representations obtained after the LLM decoder. We appreciate this suggestion. All results below use Qwen3-VL 2B, with the vision encoder and LLM frozen and only the probe trained.

We compare vision-encoder features with image-token hidden states after the LLM decoder. We additionally present both images in temporal order within a single prompt, asking “What is the difference in the robot's state between these two images?”, and probe an equal number of final-token hidden states.

Feature used for probingState-difference reconstruction loss ↓
Vision-encoder output0.1905
Image-token hidden states after the LLM decoder0.2086
Joint two-image prompt: final-token hidden states0.3072

Vision-encoder features yield the lowest reconstruction loss in this comparison. Under our probing protocol, neither LLM-processed image-token features nor the joint-prompt representations improve the recovery of fine-grained robot-state differences over vision-encoder features.


W2 / Q2. Baseline strength and training curves

You ask, “Why is the LIBERO baseline substantially weaker than recent VLA methods?” Our baseline is VLM4VLA, a VLA built with minimal changes to a pretrained VLM, trained jointly on all four LIBERO suites, a more demanding setting than single-suite training.

You also request “training curves” and evaluation on “a stronger baseline” under matched training conditions. We provide the training curves in Appendix D.3, Figure 3, and a stronger-baseline comparison below. The curves show that the decline in loss had substantially slowed by the end of training.

We evaluate FLOWER [1] on LIBERO-10, following the original paper’s LIBERO training and evaluation setup and comparing vanilla and +Ours under the same conditions.

ModelLIBERO-10 success rate ↑
FLOWER vanilla88.4%
FLOWER +Ours94.4% (+6.0)

The improvement from 88.4% to 94.4% shows that our method remains effective with a stronger baseline.


The experiments for your remaining points are still running; we will post the results in a follow-up comment as soon as they are ready.


[1] FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies, CoRL 2025.

AuthorsOpenReview ↗

Additional results: inverse dynamics vs. state-difference supervision

Q3. Why inverse dynamics and comparison with alternative objectives

Thank you for this suggestion. You ask, “How does it compare with direct state-difference prediction or other auxiliary representation-learning objectives?” Direct state-difference supervision yields a lower state-difference probe loss, while inverse dynamics achieves better policy performance.

On CALVIN ABC→D, we compare vanilla FLOWER with three auxiliary objectives: behavior cloning (BC)—actions from current visual features; state-difference prediction (SD)—state differences from current and future visual features; and Ours—intervening actions from current and future visual features.

Model1 task ↑2 tasks ↑3 tasks ↑4 tasks ↑5 tasks ↑Avg. successful sequence length ↑
Vanilla96.9%91.4%85.5%78.9%73.1%4.258
+BC96.6% (−0.3)91.9% (+0.5)86.2% (+0.7)80.0% (+1.1)73.0% (−0.1)4.277 (+0.019)
+SD99.2% (+2.3)94.9% (+3.5)88.0% (+2.5)80.8% (+1.9)73.2% (+0.1)4.361 (+0.103)
+Ours99.5% (+2.6)96.6% (+5.2)91.2% (+5.7)86.9% (+8.0)81.6% (+8.5)4.558 (+0.300)

We also freeze the trained vision encoders and apply the same state-difference probe to features from the CALVIN validation set.

ModelState-difference probe loss ↓
Vanilla0.033
+BC0.024
+SD0.007
+Ours0.021

SD yields a lower probe loss, but Ours achieves better policy performance. One possible explanation is that predicting intervening actions encourages sensitivity to state changes useful for the policy's own action prediction, beyond supervising state differences alone.

Dear Reviewer KayM,

With one day remaining in the discussion period, we would appreciate your feedback on our responses above, including the probing of representations after the LLM decoder (W1/Q1), the stronger FLOWER baseline on LIBERO-10 (W2/Q2), and the comparison with direct state-difference prediction and auxiliary BC (Q3), as well as our clarification of the training curves (W2/Q2). Could you let us know which concerns, if any, remain unresolved? We would be happy to provide further clarification.

Thank you for your time and consideration.

Reviewer_xVFaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper defines state aliasing as the failure of visual encoders to preserve fine-grained distinctions between robot states. Across encoders, they observe a consistent trend: representations with less state aliasing also support more accurate action prediction. This association suggests that state discrimination is closely related to the utility of visual representations for low-level control. Motivated by the relationship observed in their diagnostic, they introduce inverse dynamics learning as an auxiliary objective that directly supervises the VLA vision encoder during training. Given current and future observations from a demonstration trajectory, a lightweight inverse dynamics head predicts the corresponding robot action from their visual features. This objective encourages the encoder to retain visual distinctions that determine low-level robot actions. Gradients from this auxiliary objective are propagated through the inverse dynamics head into the vision encoder, directly shaping the visual representation.

Critical strengths
  • The diagnostic is the most valuable part of the submission. Freezing each encoder and fitting a state-difference probe is a clean and cheap protocol, and running it over seven pretrained encoders lifts the observation above a single-model anecdote.

  • The method is genuinely simple and architecture-agnostic. It needs no annotation beyond the demonstrations themselves, and it adds nothing to the inference graph.

  • Empirical breadth is above average for an auxiliary-objective paper. Five VLA configurations spanning autoregressive, flow-matching, future-frame-prediction, and spatially grounded architectures, across three benchmarks, is a serious evaluation effort.

  • The appendix is detailed, with probe architectures, sampling protocols, and per-model hyperparameters spelled out.

Critical weaknesses
  • The FLOWER comparison is not like-for-like. The vanilla FLOWER number is taken from the original paper because their own reproduction underperformed, while the +Ours number comes from their own run with the released configuration. A baseline and its modified variant have to come from the same training environment; otherwise the delta is meaningless.

  • Novelty is incremental. Inverse dynamics as a representation objective has a long prior line, and the paper cites it fairly. Porting it to VLA encoders is a useful application, but the conceptual delta is small, and the closest concurrent work on representation regularization for VLA is cited and never compared against.

  • The abstract overclaims. It states that the analysis shows state aliasing to be a key source of the limitation, while the analysis only establishes an association. The contribution list in the introduction is correctly hedged, and the abstract should be aligned with it.

Questions for authors
  • Can the LIBERO comparison be repeated against a stronger, well-converged baseline, so that the large average gain can be separated from baseline weakness?

  • What happens with a single-frame auxiliary action head or a contrastive objective on nearby frames; that is, does the two-view inverse structure matter, or is any extra action supervision enough?

  • Why does GR-1 use a different gradient clipping value and an auxiliary weight two orders of magnitude larger, and does the gain survive when those are matched?

  • How much additional training compute does the second encoder forward pass cost?

4 comments
AuthorsOpenReview ↗

Response to Reviewer xVFa

Thank you for your careful review and constructive feedback. We address your concerns and provide additional evidence below.


W2. Novelty and comparison with related representation regularization

Regarding your assessment that “Novelty is incremental,” our contribution lies in connecting a representation problem, its analysis, and a method for addressing it, rather than introducing inverse dynamics itself. We diagnose state aliasing—the failure of visual encoders to preserve fine-grained robot-state distinctions—and establish its empirical relationship with action prediction through frozen-encoder probing (Section 3). Motivated by this analysis, we apply auxiliary inverse dynamics directly to the VLA vision encoder during policy training to improve state discriminativity (Section 4).

You also point out that the closest concurrent work is “cited and never compared against.” We now provide a direct comparison with RS-CL [1], following its SigLIP2 policy architecture and training recipe, including its LIBERO training settings, on LIBERO-10. We also compare their state discriminativity using the same frozen-encoder state-difference probe.

ModelLIBERO-10 success rate ↑State-difference probe loss ↓
Vanilla93.2%0.04741
+RS-CL93.8% (+0.6)0.04609
+Ours94.6% (+1.4)0.04257

Q1. Stronger, well-converged LIBERO baseline

Thank you for asking whether “the large average gain can be separated from baseline weakness.” Yes—the improvement persists with a stronger baseline: our method increases FLOWER’s LIBERO-10 success rate from 88.4% to 94.4%. We follow the original FLOWER [2] paper’s LIBERO training and evaluation setup, comparing vanilla and +Ours under the same conditions.

ModelLIBERO-10 success rate ↑
FLOWER vanilla88.4%
FLOWER +Ours94.4% (+6.0)

W1. Fair comparison with FLOWER

Thank you for carefully checking the comparison details and pointing out that “The FLOWER comparison is not like-for-like.” We now compare +Ours against our own reproduced vanilla FLOWER baseline, improving the fairness of this evaluation. We trained the baseline on CALVIN ABC→D following FLOWER’s official code, released configuration, and training setup, and evaluated its 40K-step checkpoint.

Model1 task ↑2 tasks ↑3 tasks ↑4 tasks ↑5 tasks ↑Average sequence length ↑
FLOWER vanilla (our reproduction)96.9%91.4%85.5%78.9%73.1%4.258
FLOWER +Ours99.5% (+2.6)96.6% (+5.2)91.2% (+5.7)86.9% (+8.0)81.6% (+8.5)4.558 (+0.300)

Our method increases the success rate for completing five consecutive tasks from 73.1% to 81.6% (+8.5 pp), and the average successful sequence length from 4.258 to 4.558 (+0.300).


Q4. Additional training compute

Thank you for asking “How much additional training compute does the second encoder forward pass cost?”, which helps us clarify the practical trade-off between performance gains and additional training cost. In FLOWER, forward/backward computation increases from 0.466 to 0.834 TFLOPs per step (+79%). FLOWER uses a compact, pruned language backbone, so the additional vision-encoder pass represents a relatively large fraction of its baseline computation. Measured training time per epoch, however, increases from 52m 38s to 53m 27s (approximately +1.6%). The smaller wall-clock overhead is consistent with efficient batched vision-encoder computation on the GPU: additional FLOPs need not translate into a proportional increase in training time.


W3. Association versus causation in the abstract

Thank you for pointing out that “the analysis only establishes an association,” which helps us state our conclusions more precisely. We agree and have revised the abstract to align its claims with the main text, replacing “a key source of the limitation is” with “less accurate action prediction is associated with.”


The experiments for your remaining points are still running; we will post the results in a follow-up comment as soon as they are ready.


[1] Contrastive Representation Regularization for Vision-Language-Action Models, ICML 2026

[2] FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies, CoRL 2025.

AuthorsOpenReview ↗

Additional results: two-view inverse dynamics vs. auxiliary BC

Q2. Necessity of the two-view inverse dynamics structure

Thank you for this suggestion. You ask, “does the two-view inverse structure matter, or is any extra action supervision enough?” Adding single-frame auxiliary action prediction yields only a small policy improvement, whereas two-view inverse dynamics yields a larger gain.

On CALVIN ABC→D, we compare vanilla FLOWER with two auxiliary objectives: behavior cloning (BC)—actions from current visual features; and Ours—intervening actions from current and future visual features.

Model1 task ↑2 tasks ↑3 tasks ↑4 tasks ↑5 tasks ↑Avg. successful sequence length ↑
Vanilla96.9%91.4%85.5%78.9%73.1%4.258
+BC96.6% (−0.3)91.9% (+0.5)86.2% (+0.7)80.0% (+1.1)73.0% (−0.1)4.277 (+0.019)
+Ours99.5% (+2.6)96.6% (+5.2)91.2% (+5.7)86.9% (+8.0)81.6% (+8.5)4.558 (+0.300)

We also freeze the trained vision encoders and apply the same state-difference probe to features from the CALVIN validation set.

ModelState-difference probe loss ↓
Vanilla0.033
+BC0.024
+Ours0.021

BC improves state discriminativity, but this improvement translates into only a small increase in policy performance. One possible explanation is that single-frame action prediction can exploit visual–action shortcuts, limiting the policy gains from improved state discriminativity. Our two-view inverse dynamics objective, combined with pseudo time reversal, is designed to reduce reliance on such shortcuts.

AuthorsOpenReview ↗

Clarification on GR-1 hyperparameter settings

Q3. Gradient clipping and auxiliary loss weight in GR-1

Thank you for your careful reading. You ask, “Why does GR-1 use a different gradient clipping value and an auxiliary weight two orders of magnitude larger”. We heuristically increased the clipping threshold from 1.0 to 2.0 to accommodate the additional auxiliary loss. For the auxiliary weight, GR-1 scales its action loss by 100, so our weight of 10 gives an auxiliary-to-action loss coefficient ratio of 0.1, as in our VLM4VLA 8B setting.

The matched-clipping GR-1 run could not be completed within the discussion period; we will report it in the final version.

Dear Reviewer xVFa,

With one day remaining in the discussion period, we would appreciate your feedback on our responses above, including the reproduced FLOWER baseline (W1), the direct comparison with RS-CL (W2), the stronger LIBERO-10 baseline (Q1), and the comparison of auxiliary objectives (Q2), as well as our clarification of training compute (Q4) and the GR-1 settings (Q3). Could you let us know which concerns, if any, remain unresolved? We would be happy to provide further clarification.

Thank you for your time and consideration.

Reviewer_2C4AOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to achieve state-discriminative representations that enables distinguishing fine-grained states.

Critical strengths

Using inverse dynamics for learning discriminativity among states is reasonable

Critical weaknesses
  • Novelty claim
    • The authors' claim is very similar to the motivation in RS-CL [1]
    • Moreover, the proposed method seems very similar to the proposed method. The main difference is where to pick the tokens.
    • However, there's no direct quantitative comparison w/ these previous methods.
  • Experiments
    • No real-world experiments
    • Comparison baselines are not balanced across benchmarks. Only its base framework (i.e., VLM4VLA) are compared on libero (table 2, 3) while some other baselines appear on results table on the CALVIN ABC and SimplerEnv benchmarks.
      • Currently, the generalizability of the proposed method cannot be confirmed across benchmarks and action models
    • hyperparameter sensiivity on lambda_inv is not investigated
  • There's no deep analysis on why the inverse dynamics approach enhance discriminativity. Simply providing quantitative performance does not clarify or explain these underneath rationale.

[1] Contrastive Representation Regularization for Vision-Language-Action Models, ICML 2026

Questions for authors

will be updated

3 comments
AuthorsOpenReview ↗

Response to Reviewer 2C4A

Thank you for your valuable feedback. We address your concerns and questions below.


W1. Novelty and direct comparison with RS-CL

Regarding your concern that “The main difference is where to pick the tokens”: Not quite. The distinction lies in the problems targeted and the resulting methods, not merely in token extraction.

RS-CL [1] emphasizes aligning similar robot states despite different visual appearances, using a contrastive objective in which “embeddings associated with similar proprioceptive states receive higher weights, are pulled closer in the representation space” (RS-CL, Section 2.2). We instead address state aliasing by distinguishing subtly different robot states despite similar visual appearances. Our inverse dynamics objective predicts the intervening action chunk from two temporally nearby observations, encouraging the vision encoder to retain the state differences underlying even small visual changes (our paper, Section 4.2).

AspectRS-CLOurs
ProblemState–representation misalignmentState aliasing
GoalDifferent visuals → similar features for similar statesSimilar visuals → distinct features for distinct states
SupervisionProprioceptive-state distancesIntervening action prediction

You also note that “there's no direct quantitative comparison w/ these previous methods.” We now provide a direct comparison with RS-CL, following its SigLIP2 policy architecture and training recipe, including its LIBERO training settings, on LIBERO-10. RS-CL improves success over vanilla by 0.6 percentage points, while our method improves it by 1.4 percentage points.

ModelLIBERO-10 success rate ↑
Vanilla93.2%
+RS-CL93.8% (+0.6)
+Ours94.6% (+1.4)

To examine how the methods address their respective target problems, we apply an analysis from each paper. Following RS-CL’s CKNNA protocol (Figure 8; Appendix B.3), we measure whether nearby robot states also have nearby representations. Our state-difference probe measures how accurately fine-grained robot-state differences can be recovered from frozen visual features.

ModelCKNNA ↑State-difference probe loss ↓
Vanilla0.40890.04741
+RS-CL0.45390.04609
+Ours0.41870.04257

These results support the distinction between the problems targeted by the two methods. RS-CL achieves stronger state–representation neighborhood alignment, consistent with its focus on aligning similar robot states. Our method achieves more accurate recovery of fine-grained state differences, consistent with our focus on distinguishing different robot states despite similar visual appearances.


W3. Generalizability across benchmarks and action models

Regarding your concern that “the generalizability of the proposed method cannot be confirmed across benchmarks and action models,” we have extended the evaluation of FLOWER [2] from CALVIN to LIBERO-10, in addition to the six model–benchmark settings in our submission.

ModelLIBERO-10 success rate ↑
FLOWER vanilla88.4%
FLOWER +Ours94.4% (+6.0)

W5. Beyond task performance: evidence of improved state discriminativity

Thank you for pointing out that “Simply providing quantitative performance does not clarify or explain these underneath rationale.” Our submission goes beyond task success rates: two representation-level analyses show that adding auxiliary inverse dynamics improves state discriminativity over vanilla VLA training.

State-difference probing (Section 5.2, Table 2). After LIBERO-90 training, we freeze the vision encoders and fit the same probe to recover robot-state differences from visual-feature differences. Adding our objective reduces the probe loss from 0.163 to 0.118.

State–feature alignment (Section 5.4, Table 6). On CALVIN, the partial Spearman correlation between pose distances and cosine feature distances increases from 0.489 to 0.600 for VLM4VLA and 0.303 to 0.600 for GR-1, after controlling for pixel differences. Thus, robot-pose changes are more strongly reflected in the learned features.

Together, these analyses support our rationale: predicting actions between temporally nearby observations encourages the encoder to preserve the subtle state differences that explain the transition.


W2. Lack of real-world experiments

Thank you for the suggestion, but limitations in our research environment prevent us from conducting physical robot experiments within the rebuttal period.


The experiments for your remaining points are still running; we will post the results in a follow-up comment as soon as they are ready.


[1] Contrastive Representation Regularization for Vision-Language-Action Models, ICML 2026

[2] FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies, CoRL 2025.

Dear Reviewer,

Thank you again for your careful review. We would particularly appreciate any further feedback on whether our clarification of the distinction from RS-CL and the additional comparison results adequately address your novelty concern. If you have a chance to review our response, we would be grateful for any additional thoughts.

Thank you for your time and consideration.

AuthorsOpenReview ↗

Additional results: sensitivity to the inverse dynamics loss weight

W4. Sensitivity to the inverse-dynamics loss weight

You point out that “hyperparameter sensiivity on lambda_inv is not investigated.” We evaluated this sensitivity using FLOWER on CALVIN ABC→D. The improvement over vanilla persists at twice the default weight, and performance stays close to vanilla when the weight is lowered.

Let λ=0.30\lambda = 0.30 denote the default inverse-dynamics loss weight. Under the same training and evaluation setup, we compare vanilla FLOWER with policies trained at 0.5λ0.5\lambda, λ\lambda, and 2λ2\lambda.

Weight1 task ↑2 tasks ↑3 tasks ↑4 tasks ↑5 tasks ↑Average sequence length ↑
0 (vanilla)96.9%91.4%85.5%78.9%73.1%4.258
0.5λ0.5\lambda96.6% (-0.3)92.1% (+0.7)86.9% (+1.4)80.4% (+1.5)73.8% (+0.7)4.298 (+0.040)
λ\lambda (default)99.5% (+2.6)96.6% (+5.2)91.2% (+5.7)86.9% (+8.0)81.6% (+8.5)4.558 (+0.300)
2λ2\lambda98.5% (+1.6)94.7% (+3.3)90.0% (+4.5)85.3% (+6.4)78.3% (+5.2)4.468 (+0.210)

At 2λ2\lambda, five-task success improves by 5.2 percentage points over vanilla. At 0.5λ0.5\lambda, the gain is 0.7 percentage points, consistent with a smaller contribution from the auxiliary objective at the lower weight.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank Reviewers 2C4A (R1), xVFa (R2), and KayM (R3) for their careful reading and constructive feedback. We are encouraged that R3 finds studying state aliasing for robot control important and our motivation “clear and worth investigating.” We particularly appreciate R2’s assessment of our diagnostic as “the most valuable part of the submission,” describing it as a “clean and cheap protocol” and noting that evaluation across seven pretrained encoders goes beyond a single-model observation. We are glad that reviewers find inverse dynamics reasonable for learning state-discriminative representations (R1), and our method simple and practical (R2, R3), architecture-agnostic (R2), and free of additional annotation requirements or inference-time overhead (R2, R3). R2 further highlights our empirical breadth as “above average for an auxiliary-objective paper,” with five VLA configurations across three benchmarks. We respond to each reviewer’s concerns in the individual threads.

Summary of Responses to the Reviewers’ Main Concerns

  • Relation to RS-CL (2C4A, xVFa). RS-CL aligns representations of similar robot states despite different visual appearances. Our inverse-dynamics objective predicts the actions between two observations, encouraging the encoder to distinguish subtly different robot states despite similar visual appearances. Using the RS-CL SigLIP2 recipe on LIBERO-10, RS-CL achieved 93.8% (+0.6 pp) and our method achieved 94.6% (+1.4 pp), both relative to the same 93.2% vanilla baseline.
  • Evaluation with stronger baselines (xVFa, KayM). We evaluated our method on LIBERO-10 with a SigLIP2 baseline at 93.2% and observed a gain of +1.4 pp. Adding our objective to FLOWER also increased success from 88.4% to 94.4% (+6.0 pp).
  • Auxiliary objective ablation (xVFa, KayM). On CALVIN ABC→D, vanilla FLOWER achieved a mean sequence length of 4.258. Adding single-frame action prediction, direct state-difference prediction, or two-frame inverse dynamics (ours) yielded 4.277, 4.361, and 4.558, respectively. Direct state-difference supervision gave the lowest state-difference probe loss. Inverse dynamics gave the highest policy performance among the objectives tested.

We appreciate the thoughtful discussion and thank you for considering our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.