acceptodds
Under review as a conference paper at ICLR 2027

What to Teach Before Acting: Revisiting Embodied VQA for VLA Models

Abstract

Vision-Language-Action (VLA) models build their policies on pretrained Vision-Language Models (VLMs), and embodied visual question answering (VQA) is widely used to prepare these VLMs for robot control. This paper revisits a practical yet seldom systematically studied question: which embodied VQA supervision actually transfers to downstream VLA policies? We introduce an automatic episode-to-VQA pipeline that turns simulated and real-world robot episodes into VQA, separating what a question teaches, which episodes it comes from, and how its answer is presented. Using this pipeline, we adapt a VLM with 16 generalist, capability-specific, and task-specific VQA recipes and train every adapted backbone under the same VLA protocol. Through closed-loop evaluation under diverse perturbations, we find that VLM benchmark averages are poor predictors of policy success. For instance, cross-view consistency VQA yields the strongest policy despite a lower VLM average than single-view spatial VQA. This challenges the common practice of selecting embodied supervision by VLM benchmarks. We further find that the gains are conditional. Cross-view supervision improves robustness to camera and especially robot-state changes, and much of the net improvement comes from preserving the policy's existing successes. Contrary to intuition, changing only the answer interface from multiple choice to direct generation raises the VLM average but lowers policy success. Finally, VQA built from real-world episodes also improves the policy, and mixing it with simulated VQA outperforms the simulation-only recipe at the same budget. These results indicate that embodied supervision should be chosen by the control conditions it improves and the successes it preserves.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Embodied VQA, Transferability

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_8DKzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks which embodied VQA supervision actually transfers to downstream VLA policies, a practical question that has received surprisingly little systematic attention despite being implicit in nearly every VLA training recipe. The authors build an automatic episode-to-VQA pipeline that separates what a question teaches (task), which episodes it comes from (source), and how the answer is presented (interface), and that runs on both simulated RoboCasa and real-world DROID episodes.

Critical strengths
  • The studied problem is an important and well-posed question, and the paper addresses this with a simple and effective decomposition, separating supervision into what is taught, where episodes come from, and how the question is asked. This decomposition both simple and well motivated.
  • The paper provided an automated, modular pipeline that is not tied to a particular model or task, which is reusable beyond the specific recipes studied here.
  • Strong experimental control, which makes the comparisons credible. All recipes start from the same checkpoint and use the same GRPO configuration; every adapted VLM is converted into a policy with one shared training recipe (30,000 steps, two seeds); and every policy is evaluated on the full LIBERO-plus benchmark. Recipe pairs are matched in example count and in the distribution of difficulty, and comparisons are made at the level of individual variants, separating gained from lost successes.
Critical weaknesses
  • The experimental scope is too narrow to support the general conclusions drawn. Although the findings are stated in general terms ("embodied supervision should be selected by the control conditions it improves and the successes it retains"), they rest on a single VLM backbone (Qwen3.5-4B) and a single policy framework (StarVLA/QwenGR00T). The paper would be considerably stronger if the headline conclusions were scoped to this setting, or if the key comparisons were replicated with a second backbone or policy framework.
  • The interface study is not a clean manipulation of the answer interface. The abstract states that the paper changes only the answer interface from multiple choice to direct generation, but the body notes that this conversion "also changes the answer target and its verifier." Because the target and the reward function change together with the format, the observed dissociation that the VLM average rising from 54.85 to 55.53 while policy success falls from 83.16 to 80.87 cannot be attributed to the answer format alone. As written, this comparison does not isolate the effect of how the question is asked.
Questions for authors
  • Can you replicate the central comparisons with a second VLM backbone and/or a different policy framework? Even a partial replication would substantially strengthen the claim.
  • The interface comparison changes the answer target and the verifier in addition to the format. Can you isolate the effect of the interface itself?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8DKz

We sincerely thank you for your time and constructive feedback. We appreciate your recognition of the research question and modular pipeline. To address your concerns, we conducted additional experiments with another backbone and policy architecture, together with a controlled study of answer options. We summarize the results below. All revisions in the updated PDF are highlighted in blue.

1: Another backbone and policy architecture (W1 & Q1)

We appreciate this suggestion. To examine whether the findings extend beyond the original setting, we conducted two additional experiments:

  • Different VLA action head architecture: retain Qwen3.5-4B and replace StarVLA-GR00T with StarVLA-π.
  • Different VLM backbone: retain StarVLA-GR00T and replace Qwen3.5-4B with Gemma 4 E2B.

For each configuration, we compared the Cross-view-adapted backbone with its unadapted baseline under the same downstream training and evaluation settings. The Qwen3.5-4B VLMs are the same models as in the paper, so their VLM scores are taken from the paper. The first two rows are the results reported in the paper.

VLM backboneVLA action headSupervisionVLM Avg.OverallCameraRobotLanguageLightNoise
Qwen3.5-4BStarVLA-GR00TBaseline56.5680.2060.7359.5592.2694.4087.95
Qwen3.5-4BStarVLA-GR00TCross-view54.8583.1667.7968.7792.4597.3786.51
Qwen3.5-4BStarVLA-πBaseline56.5680.7662.4862.9089.3996.4186.45
Qwen3.5-4BStarVLA-πCross-view54.8582.2265.1065.2392.6597.1185.20
Gemma 4 E2BStarVLA-GR00TBaseline32.2070.5946.9048.7785.1091.9464.96
Gemma 4 E2BStarVLA-GR00TCross-view32.6170.7250.4148.5889.7995.1858.03

With StarVLA-π, Cross-view improves policy success by 1.46% despite a lower Embodied VLM Avg., showing that the VLM–policy mismatch also holds for a different policy architecture. In all three settings, Cross-view shows gains under Camera, Language, and Light perturbations. On Gemma, most of these gains are lost to a drop under Noise (−6.93%), which also falls in the other two settings. Therefore, the Camera, Language, and Light gains hold in both new settings, while the overall and robot-state gains vary with the backbone and action head. We have added these results to the main text as Sec. 5.7 (Table 4).

2: Controlled answer-option comparison (W2 & Q2)

Thank you for raising this point. In the original comparison (Sec. 5.6, Table 3), removing the options also changes the answer target and verifier. To isolate the options, we fix the target and reward: both variants use the same 2,000 examples, images, prompt, and training settings, and generate the answer text inside <answer>...</answer>. They differ only in whether the four candidates are shown.

Original interface experiment

With optionsAnswer interfaceRewardEmbodied VLM Avg.LIBERO-plus (%)
YesChoiceExact option-label match54.8583.16
NoDirect gen.Short-text verifier55.5380.87

Controlled option experiment

With optionsAnswer interfaceRewardEmbodied VLM Avg.LIBERO-plus (%)
YesDirect gen.Exact success + bounded contrastive shaping55.2082.41
NoDirect gen.Exact success + bounded contrastive shaping55.3679.77

The shared reward gives 1 to an exact correct answer, 0 to an exact incorrect candidate, and bounded partial credit to close paraphrases of the target.

With the answer target and reward being fixed, adding the candidate options increases policy success by 2.64% and by 8.39% under Robot perturbations while the Embodied VLM Avg. remains nearly unchanged. Combined with the original comparison, this result attributes the performance gap primarily to the presence of candidate options. When the options are shown, changing the answer target and reward only has a marginal effect (changing from 83.16% to 82.41%), but removing them reduces success to 79.77%.

We have added this controlled experiment to Sec. 5.6 and extended Table 3 with the two new variants. We have also revised the abstract to accurately reflect these findings.

Thank you again for your constructive suggestions. We hope these additional experiments and explanations address your concerns.

Thanks to the authors for their detailed response, which addressed my main concerns. I will keep my positive assessment.

Thank you for your thoughtful feedback! We are happy to see that our rebuttal helped address your concerns and we hope our clarifications will also be helpful to the broader discussion.

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

The paper investigates what types of embodied VQA supervision can improve downstream VLA policies. For this, the authors introduce an automatic episode-to-VQA construction pipeline that converts simulated RoboCasa and real-world DROID robot episodes into training questions. The authors evaluate 16 VQA adaptation recipes based on Qwen3.5-4B and also a common VLA training procedure and closed-loop evaluation on 10,030 LIBERO-plus variants. The results show that improvements in VLM benchmark scores do not necessarily translate into better robot-policy performance.

Critical strengths
  • It addresses an important question in VLA research and investigates what types of embodied VQA supervision actually imp downstream robot control.
  • The proposed pipeline systematically separates supervision targets, data sources and also answer interfaces. This enables controlled comparisons across different training recipes.
  • I feel that the evaluation is extensive. It covers 16 VQA recipes and 10,030 LIBERO-plus variants and has consistent training settings and detailed analyses across pertubation categories which strengthen the paper.
  • The findings are also informative and reveal that VLM benchmark scores do not necessarily predict VLA performance.
  • The analysis of gained and lost successes also provides useful insights into effective supervision choices.
Critical weaknesses
  • Some supervision comparisons vary multiple factors simultaneously like the input views or question targets or data sources and this makes it difficult to isolate which factors contribute to the improvements and their share.
  • The exps use a single VLM backbone and one downstream VLA framework and has simulation-only evaluation so this I feel limits generalization to other models and real-world robots a bit.
  • Although policies are trained with two random seeds there is no confidence interval or seed-level variability in the paper. This is an imp information when evaluating some small performance gains.
Questions for authors
  • Can you try to compare single-view and cross-view supervision using the same question targets and source episodes? This will help in isolating the effect of multi-view information.
  • Please also evaluate whether the findings generalize to another VLM backbone or VLA framework. Though this is a minor point and is encouraged to be done if time permits.
  • Can you also please report the per-seed variability of the CIs?
  • Also can the authors please quantify the accuracy of DROID-derived annotations and clarify how much of the MCQ versus direct-generation difference comes from changes in the answer target and verifier?
  • Please also try to address the weaknesses which aren't in these questions.
5 comments
AuthorsOpenReview ↗

Response to Reviewer mDXg

We sincerely thank you for your time and constructive feedback. We appreciate your positive assessment of the research question and evaluation. Your suggestions have helped us strengthen our experimental analysis. Below, we present the new controlled experiments, evaluations with another backbone and policy architecture, and analyses of seed variability and annotation quality. All revisions in the updated PDF are highlighted in blue.

1: Controlled view comparison (W1 & Q1)

We appreciate this suggestion. To isolate the contribution of multiple views, we constructed two input variants from the same 2,000 Cross-view examples: synchronized left/right images (L+R) and the left-camera image duplicated (L+L). Both variants use the same adapted prompts, options, correct answers, reward, and training settings.

Compared with the original Cross-view prompts in the paper, we adapted the wording to accommodate either input. We changed the introduction to “Answer using the supplied camera observation(s) of one RoboCasa episode frame.” For phase questions, we replaced “Which single action phase explains both?” with “Which single action phase is shown?” Where present, we removed “Which option best answers the question?” and used “Answer with A, B, C, or D only.” Both variants use these same adapted prompts, preserving the question targets, option order, and correct answers. The results below therefore compare L+L and L+R under a shared adapted prompt.

InputEmbodied VLM Avg.LIBERO-plus success rate (%)
L+L56.2780.51
L+R55.1981.16

As shown in the table, L+R improves policy success by 0.65% despite a lower Embodied VLM Avg. Because the two variants share the same questions, answers, reward, and training settings, this result is consistent with a benefit of synchronized views for policy transfer.

Both variants score below the original Cross-view recipe (83.16%), which uses the same synchronized images but retains the original two-view wording. This suggests that the transfer gain arises mainly from questions that require relating the two views, rather than from the additional view alone. We have added this experiment to Appendix P (Table 31).

2: Another backbone and policy architecture (W2 & Q2)

We appreciate this suggestion. To examine whether the findings extend beyond the original setting, we conducted two additional experiments:

  • Different VLA action head architecture: retain Qwen3.5-4B and replace StarVLA-GR00T with StarVLA-π.
  • Different VLM backbone: retain StarVLA-GR00T and replace Qwen3.5-4B with Gemma 4 E2B.

For each configuration, we compared the adapted Cross-view backbone with its unadapted baseline under the same downstream training and evaluation settings. The first two rows are the results reported in the paper.

VLM backboneVLA action headSupervisionVLM Avg.OverallCameraRobotLanguageLightNoise
Qwen3.5-4BStarVLA-GR00TBaseline56.5680.2060.7359.5592.2694.4087.95
Qwen3.5-4BStarVLA-GR00TCross-view54.8583.1667.7968.7792.4597.3786.51
Qwen3.5-4BStarVLA-πBaseline56.5680.7662.4862.9089.3996.4186.45
Qwen3.5-4BStarVLA-πCross-view54.8582.2265.1065.2392.6597.1185.20
Gemma 4 E2BStarVLA-GR00TBaseline32.2070.5946.9048.7785.1091.9464.96
Gemma 4 E2BStarVLA-GR00TCross-view32.6170.7250.4148.5889.7995.1858.03

With StarVLA-π, Cross-view improves policy success by 1.46% despite a lower Embodied VLM Avg., showing that the VLM–policy mismatch also holds for a different policy architecture. In all three settings, Cross-view shows gains under Camera, Language, and Light perturbations. On Gemma, most of these gains are lost to a drop under Noise (−6.93%), which also falls in the other two settings. Therefore, the Camera, Language, and Light gains hold in both new settings, while the overall and robot-state gains vary with the backbone and action head. We have added these results to the main text as Sec. 5.7 (Table 4).

AuthorsOpenReview ↗

Response to Reviewer mDXg

3: Seed variability and confidence intervals (W3 & Q3)

Thank you for requesting these details. To provide a complete per-seed breakdown, we re-evaluated our saved policy checkpoints from policy-training (GRPO + VLA) seeds 42 and 43 across two VLA evaluation seeds: the original evaluation seed (7) and an additional evaluation seed (8).

All results below come from these new evaluations. We report the mean and sample standard deviation within each group.

Training seedEvaluation seedBaselineSingle-view spatialCross-view
42780.1080.4283.41
42881.1181.8282.88
43780.0780.0982.67
43880.2081.1782.90
Mean ± SD—80.37 ± 0.4980.87 ± 0.7882.97 ± 0.31

Cross-view outperforms both Single-view spatial and the baseline in all four training and evaluation seed combinations, reaching a mean success rate of 82.97% compared with 80.87% and 80.37%. It also varies least across runs with a SD 0.31. These results support the transfer benefit of Cross-view supervision.

To assess uncertainty across seeds, we report the improvements and paired 95% confidence intervals for each checkpoint and evaluation below. All differences are absolute changes in success rate (%).

To compute the intervals, we use 100,000 paired bootstrap resamples of the 40 base tasks. In each resample, we draw ten tasks with replacement within each suite and retain all variants of each sampled task. For each training and evaluation seed combination, we use the same sampled tasks for both policies, recompute their count-weighted success rates, and take the paired difference. Baseline comparisons use the baseline with the same training and evaluation seeds. The 2.5th and 97.5th percentiles give the interval endpoints. These intervals measure task-sampling uncertainty for each evaluated checkpoint and evaluation seed, while the SD values describe the observed variation across the four runs.

Training seedEvaluation seedCross-view − BaselineCross-view − Single-view spatial
427+3.31 [1.24, 5.85]+2.99 [1.50, 4.56]
428+1.77 [0.35, 3.21]+1.06 [-0.36, 2.39]
437+2.60 [-0.05, 5.81]+2.58 [0.50, 4.66]
438+2.70 [1.22, 4.34]+1.73 [-0.27, 3.90]

Cross-view outperforms both the baseline and Single-view spatial in every training and evaluation seed combination, with gains of 1.77–3.31% and 1.06–2.99%, respectively. The paired 95% intervals are predominantly positive, supporting a consistent advantage across seeds. We have added the per-seed results and the paired bootstrap intervals, together with the full procedure, to Appendix O (Tables 28 and 30).

4: Accuracy of DROID-derived annotations (Q4)

Thank you for raising this question. We manually checked 300 DROID-derived VQA examples spanning all 29 task families, including 100 Cross-view examples. Two annotators each reviewed 150 examples, checking every question and answer against the input images and the source video.

Each example was scored on a five-point scale, and we count an example as accurate if it scores 4 or higher. This requires the answer to be correct and supported by the visual evidence. Minor imprecision in wording or localization is allowed only if it does not change the answer.

SubsetSamplesMean scoreAccuracy (%)
All sampled DROID examples3004.58/588.33
Cross-view1004.62/588.00

The DROID-derived annotations reach 88.33% accuracy overall and 88.00% on the Cross-view subset. We have added this evaluation, including the sampling and scoring protocol, to Appendix B.4 (Table 6).

5: Controlled answer-option comparison (Q4)

Thank you for pointing this out. In the original comparison (Sec. 5.6, Table 3), the options, answer target, and verifier change together. To isolate the options, we fix the target and reward: both variants use the same 2,000 examples, images, prompt, and training settings, and generate the answer text inside <answer>...</answer>. They differ only in whether the four candidates are shown.

Original interface experiment

With optionsAnswer interfaceRewardEmbodied VLM Avg.LIBERO-plus (%)
YesChoiceExact option-label match54.8583.16
NoDirect gen.Short-text verifier55.5380.87

Controlled option experiment

With optionsAnswer interfaceRewardEmbodied VLM Avg.LIBERO-plus (%)
YesDirect gen.Exact success + bounded contrastive shaping55.2082.41
NoDirect gen.Exact success + bounded contrastive shaping55.3679.77

The shared reward gives 1 to an exact correct answer, 0 to an exact incorrect candidate, and bounded partial credit to close paraphrases of the target.

With the answer target and reward being fixed, adding the candidate options increases policy success by 2.64% and by 8.39% under Robot perturbations while the Embodied VLM Avg. remains nearly unchanged. Combined with the original comparison, this result attributes the performance gap primarily to the presence of candidate options. When the options are shown, changing the answer target and reward only has a marginal effect (changing from 83.16% to 82.41%), but removing them reduces success to 79.77%.

We have added this controlled experiment to Sec. 5.6 and extended Table 3 with the two new variants. We have also revised the abstract to accurately reflect these findings.

Thank you again for your constructive suggestions. We hope these additional experiments and explanations address your concerns.

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

The paper asks which embodied VQA supervision is useful before training a VLA policy. It proposes a pipeline that converts robot episodes into VQA data and separates reusable evidence, task targets and answer interfaces. The evidence comes from the RoboCasa simulator and from perception on DROID. The authors build sixteen VQA-adapted versions of Qwen3.5-4B, which are compared with the original model under the same downstream policy-training setup on LIBERO-plus. There are three main findings. VLM benchmark averages can reverse the ranking of the resulting policies. Transfer depends on the perturbation type. Net improvement depends on keeping previously solved cases as well as solving previous failures. Comparisons of data sources and answer interfaces give further examples of these effects.

Critical strengths
  1. The common downstream architecture, demonstrations and training budget make the recipe comparisons useful. Within-family pairs also match sample counts, difficulty and GRPO settings, providing practical evidence about supervision choices (Tables 5, 7, 11).

  2. Condition-level and gained/lost analyses show changes hidden by average success rates. Cross-view reaches 83.16% versus 80.20% for the base policy, and benchmark-set sensitivity checks support the reported ranking reversals (Figs. 2–4; Appendix G).

Critical weaknesses
  1. Cross-view and Single-view differ in question content and data source as well as views. Matching sample count and difficulty supports the recipe comparison, but does not isolate a multi-view effect (Table 5; Appendix D.3, pp. 18, 29).

  2. The paper averages two policy-training seeds but does not show individual results or clearly specify paired-outcome aggregation. This leaves the smaller recipe differences and gained/lost decomposition uncertain across seeds (Fig. 4; Table 24, pp. 8, 41).

Questions for authors
  1. Is there a comparison using the same source and question type that isolates the effect of multiple views?

  2. Could you report the per-seed results for Cross-view and Single-view?

3 comments
AuthorsOpenReview ↗

Response to Reviewer zami

We sincerely thank you for your time and constructive feedback. We appreciate your recognition of the shared training protocol and condition-level analysis. Your questions have helped us strengthen the experimental evidence and clarify our findings. In response, we conducted additional evaluations and controlled experiments, which are summarized below. All revisions in the updated PDF are highlighted in blue.

1: Controlled view comparison (W1 & Q1)

We appreciate this suggestion. To isolate the contribution of multiple views, we constructed two input variants from the same 2,000 Cross-view examples: synchronized left/right images (L+R) and the left-camera image duplicated (L+L). Both variants use the same adapted prompts, options, correct answers, reward, and training settings.

Compared with the original Cross-view prompts in the paper, we adapted the wording to accommodate either input. We changed the introduction to “Answer using the supplied camera observation(s) of one RoboCasa episode frame.” For phase questions, we replaced “Which single action phase explains both?” with “Which single action phase is shown?” Where present, we removed “Which option best answers the question?” and used “Answer with A, B, C, or D only.” Both variants use these same adapted prompts, preserving the question targets, option order, and correct answers. The results below therefore compare L+L and L+R under a shared adapted prompt.

InputEmbodied VLM Avg.LIBERO-plus success rate (%)
L+L56.2780.51
L+R55.1981.16

As shown in the table, L+R improves policy success by 0.65% despite a lower Embodied VLM Avg. Because the two variants share the same questions, answers, reward, and training settings, this result is consistent with a benefit of synchronized views for policy transfer.

Both variants score below the original Cross-view recipe (83.16%), which uses the same synchronized images but retains the original two-view wording. This suggests that the transfer gain arises mainly from questions that require relating the two views, rather than from the additional view alone. We have added this experiment to Appendix P.

2: Per-seed results and gained/lost aggregation (W2 & Q2)

Thank you for requesting these details. To provide a complete per-seed breakdown alongside trajectory-level gained/lost comparisons, we re-evaluated our saved policy checkpoints from policy-training (GRPO + VLA) seeds 42 and 43 across two VLA evaluation seeds: the original evaluation seed (7) and an additional evaluation seed (8).

All results below come from these new evaluations. We report the mean and sample standard deviation within each group.

Training seedEvaluation seedBaselineSingle-view spatialCross-view
42780.1080.4283.41
42881.1181.8282.88
43780.0780.0982.67
43880.2081.1782.90
Mean ± SD—80.37 ± 0.4980.87 ± 0.7882.97 ± 0.31

Cross-view outperforms both Single-view spatial and the baseline in all four training and evaluation seed combinations, reaching a mean success rate of 82.97% compared with 80.87% and 80.37%. It also varies least across runs with a SD of 0.31. These results support the transfer benefit of Cross-view supervision.

To clarify the gained/lost analysis, we compare each policy with the baseline from the corresponding evaluation on the same variants. A gained success is a baseline failure solved by the adapted policy. A lost success is a baseline success that the adapted policy fails. We calculate these counts separately for each training and evaluation seed combination and then average them over the four runs.

Training seedEvaluation seedCross-view gainedLostNetSingle-view gainedLostNet
427781449+332597565+32
428654476+178552480+72
437761500+261723721+2
438759488+271713616+97
Mean—738.75478.25+260.5646.25595.5+50.75

Averaging the four runs, Cross-view solves 92.5 more baseline failures and loses 117.25 fewer baseline successes than Single-view spatial, with mean net gains of 260.5 and 50.75, respectively. In every training and evaluation seed combination, Cross-view recovers more baseline failures and retains more baseline successes than Single-view spatial.

We have added the per-seed success rates, the gained/lost table, and the definition of the gained/lost aggregation to Appendix O (Tables 28–29).

Thank you again for your constructive suggestions. We hope these additional experiments and explanations address your concerns.

Thank you for your time and thoughtful review. As the discussion period will end soon, we would appreciate your thoughts on whether it addresses your concerns. Please let us know if there are any remaining points we can clarify.

Thank you for the additional controlled comparison and per-seed analyses. These address my main concerns, and I am happy to keep my current positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.