acceptodds
Under review as a conference paper at ICLR 2027

SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation

21 views

Abstract

Large-scale, diverse datasets have driven the success of LLMs and VLMs. But VLAs for robotics remain limited by the cost and complexity of real-world data collection. While simulation offers a scalable alternative, its potential for zero-shot sim-to-real VLA learning in mobile manipulation remains largely underexplored. We introduce SimVLA, a framework that trains VLAs entirely from synthetic simulation data generated without teleoperation for mobile manipulation. SimVLA is first pre-trained on two complementary simulation-derived datasets: SimAction, a large-scale robot action dataset spanning 30 diverse mobile manipulation tasks, generated by composing atomic skills, and SimVQA, which leverages privileged simulator state to provide spatial, geometric, and subtask-level visual-language supervision. We further post-train SimVLA on a mixture of SimAction and SimDeploy, a dataset collected from policy rollouts across diverse simulated environments. We evaluate SimVLA on tasks including restocking, pouring, and cleaning, and show zero-shot transfer to real-world mobile manipulation, including real home environments. SimVLA outperforms policies trained on 50 in-domain real-world demonstrations, suggesting that simulation can enable scalable zero-shot sim-to-real mobile manipulation. We also demonstrate that multiple complementary forms of supervision in simulation can serve as an compelling alternative to expensive real-world data collection.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, VLA, Sim-to-Real, Mobile 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_FSQKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SimVLA, a sim-to-real framework for mobile manipulation that addresses the high cost of collecting expert demonstrations. It trains a π0.5-style VLA purely on three simulation-derived data sources. SimAction contains ~200K trajectories over 35 tasks, generated in procedurally built kitchens with skill APIs and motion planning. SimVQA contains ~1M VQA pairs derived from privileged simulator state. SimDeploy contains policy rollouts collected in diverse simulated environments. The framework also uses hybrid system identification and domain randomization. Extensive real-world experiments validate the effectiveness of the proposed method: SimVLA transfers zero-shot to a kitchen, a pantry and a real home, and outperforms π0.5 fine-tuned on 50 real demonstrations.

Critical strengths
  1. The proposed sim-to-real framework is a key approach to the difficulty of collecting mobile manipulation data, and the data synthesis pipeline is comprehensive.
  2. The paper is clearly written, and the experimental results are convincing.
Critical weaknesses
  1. Missing real-to-sim component. In my view, the key to sim-to-real is real-to-sim rather than data synthesis: building visually and physically faithful 3D assets of real scenes, then applying simple domain randomization. The paper skips this step. All scenes are procedurally generated from Objaverse assets, and the only real-to-sim step is system identification of the robot's own controller gains. Environment physics, such as the friction and PD gains of drawers and handles, is only randomized over hand-set ranges (Table 5). As a result, transfer depends on whether hand-crafted generators happen to cover the target domain. This limits generality beyond kitchen-like settings.
  2. Insufficient study of the key sim-to-real components. All ablations (Sec. 4.4–4.6) target VLA training ingredients rather than the sim-to-real gap. There is no analysis of how the visual gap between simulator and real world is reduced. For example, the author can explore whether 3D point-cloud input has a smaller sim-to-real gap than 2D images.
Questions for authors
  1. Could SimVLA be combined with or compared against a real-to-sim pipeline, e.g., reconstructed scenes with measured physical parameters? How much effort would it take to extend it beyond kitchens?
  2. Can you ablate visual vs. physical domain randomization and analyze the visual sim-to-real gap?
  3. Have you tried RGB-D or point-cloud inputs, and do they reduce the sim-to-real gap compared with RGB-only inputs?
3 comments

We thank the reviewer for the detailed and constructive feedback.

[W1] Missing real-to-sim component. … building visually and physically faithful 3D assets of real scenes, then applying simple domain randomization.

We agree that real-to-sim—building visually and physically faithful 3D assets of real scenes—is an important factor in improving sim-to-real performance. However, our core contribution is zero-shot sim-to-real transfer for mobile manipulation without environment-specific information. Following prior sim-to-real VLA works [1], we leverage domain randomization to promote generalization across diverse environments without environment-specific fine-tuning. Concurrently, MolmoB0T [2] demonstrates the effectiveness of large-scale simulation training with diverse environments and domain randomization. Accordingly, SimVLA generates SimAction data across 100 diverse kitchen scenes, incorporating multiple object instances per category (e.g., 80 different mugs from Objaverse) and 41 types of domain randomization spanning camera positions, kitchen environments, object assets, robot physical configurations, material properties, and lighting conditions.

While deployment environments beyond kitchen-like settings may introduce additional visual and physical variations, these do not fundamentally limit our approach, as SimVLA does not rely on hand-crafted generators. Therefore, we respectfully suggest that the absence of real-to-sim components is not a weakness, but is a deliberate difference in research focus.

 

[W2-1, Q2] All ablations (Sec. 4.4–4.6) target VLA training ingredients rather than the sim-to-real gap. … no analysis of how the visual gap between simulator and real world is reduced. … Can you ablate visual vs. physical domain randomization and analyze the visual sim-to-real gap?

The key component that enables sim-to-real transfer is leveraging three complementary forms of simulation supervision—SimAction, SimVQA, and SimDeploy—within a two-stage VLA training. SimVLA suggests that simulation can serve not only as a source of action trajectories (SimAction), but also as a source of visual-language supervision (SimVQA) and deployment-oriented rollout data (SimDeploy). By these, SimVLA demonstrates robust performance across five long-horizon mobile manipulation tasks in three real-world environments. In a real home, SimVLA achieves an average task progress of 58.5% and a binary success rate of 45% (Table 15), without real-world training data or environment-specific information.

We acknowledge in Section 6 (Limitations & Future Work) that our work does not measure how the sim-to-real visual gap is reduced or compare against physical domain randomization. Instead, our work places greater emphasis on providing the ingredients and instructions that enable zero-shot sim-to-real transfer in mobile manipulation.

We clarified our experimental design in Section 4 (Experiments) and made this point clearer in Section 6 (Limitations & Future Work) of the revised paper.

 

[W2-2, Q3] Have you tried RGB-D or point-cloud inputs, and do they reduce the sim-to-real gap compared with RGB-only inputs?

Thank you for suggesting an interesting idea! Now, our simulation pipeline supports generating RGB, depth, and point-cloud data, although our VLA training currently uses only RGB inputs. Our goal is to develop an RGB-camera-based VLA, which is widely adopted design choice of recent robot foundation models [1, 2, 3, 4]. We agree that these additional modalities may help reduce the sim-to-real gap. However, we respectfully claim that comparing RGB-D and point-cloud inputs is beyond the scope of our work.

We added RGB-D/Point-cloud data generation to Appendix D (SimAction) and Figure 15 of the revised paper. The supplementary materials include the updated code.

[Q1] Could SimVLA be combined with or compared against a real-to-sim pipeline, e.g., reconstructed scenes with measured physical parameters? How much effort would it take to extend it beyond kitchens?

We thank the reviewer for the constructive feedback. A real-to-sim pipeline can be integrated with SimVLA as a complementary extension to enable deployment in specific target scenes. Specifically, a single image of a target environment can be processed using ACDC [5] to segment fixtures and estimate metric depth, followed by scene reconstruction using SimVLA's procedural assets and demonstration generation. Now, our code supports this integration; however, we have not conducted fine-tuning or experiments using this real-to-sim integration in the current work. We emphasize that our core contribution remains zero-shot sim-to-real transfer for mobile manipulation without environment-specific information.

A procedurally generatable environment is the main requirement for extending SimVLA beyond kitchens, making such an extension straightforward once available. Furthermore, a trained policy can be deployed outside kitchens as long as the task remains the same. For example, we trained a "Push Chair" policy using data collected from kitchen table chairs and evaluated it zero-shot in a study room generated with SceneSmith [6], where it successfully pushed a study desk chair.

We added the real-to-sim integration in Appendix K (Real-to-Sim). The supplementary material includes updated code and corresponding generated data.

 

References

[1] Deng et al., GraspVLA: A Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data, CoRL, 2025.

[2] Deshpande et al., MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation, CoRL, 2026.

[3] Black et al., π0.5\pi_{0.5}: A Vision-Language-Action Model with Open-World Generalization, CoRL, 2025.

[4] Bjorck et al., GR00T N1.5: An Improved Open Foundation Model for Generalist Humanoid Robots, NVIDIA, 2025.

[5] Dai et al., Automated Creation of Digital Cousins for Robust Policy Learning, CoRL, 2024.

[6] Pfaff et al., SceneSmith: Agentic Generation of Simulation-Ready Indoor Scenes, ICML, 2026.

If you have any further questions or if there are remaining points you would like us to clarify, we would be happy to provide additional clarification before the discussion period ends.

Thank you again for your time and thoughtful feedback.

Reviewer_2Q7JOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SimVLA combines scripted, motion-planned demonstrations in procedural kitchens (SimAction), simulator-derived spatial and subtask question answering (SimVQA), and successful learned-policy rollouts in additional simulated scenes (SimDeploy). A PaliGemma-based VLA is described as undergoing insulated multimodal pretraining followed by joint flow-matching post-training. The contribution is a simulation data and training pipeline for mobile manipulation, rather than a new policy architecture. Evaluation covers five real-robot tasks, changes in layout and appearance, a pantry and a home, alongside simulation comparisons and supervision ablations.

Critical strengths

The strongest result is transfer to the pantry and home, not the small advantage over 50 demonstrations in the original kitchen. In Table 17, SimVLA completes 26/60 pantry trials, compared with 3/60 for π0.5 fine-tuned on 200 demonstrations per task; in the home the corresponding totals are 9/20 and 0/20. These are complete-task successes, so the advantage is not merely an artifact of partial-progress scoring. The drawer and bimanual pouring tasks also extend the evidence beyond isolated grasping. Together with the substantial degradation without pretraining, these results make the broader simulation pipeline a potentially valuable contribution to mobile manipulation.

Critical weaknesses

The SimDeploy experiment does not isolate the proposed benefit of policy-induced states. Section 3.4 changes both the source of trajectories and the scene distribution by introducing RoboCasa, MolmoSpaces, and SceneSmith. The collector exports only successful episodes, adding outcome-based selection as well. Table 13 shows a substantial effect—14/100 complete successes without SimDeploy versus 38/100 with it—but this comparison cannot determine whether the gain comes from policy-generated recovery trajectories, additional scene coverage, or successful-trajectory selection. This matters because deployment-oriented supervision is a central distinction from existing simulation data pipelines. A comparison against an equal-sized set of successful planner trajectories in the same added scenes would directly test that distinction; alternatively, the conclusion should remain at the level that this combined additional dataset helps.

Questions for authors
  1. Clarification needed to establish the central claim: Which exact initial checkpoint and saved training configurations produced the reported SimVLA results and ablations? Please reconcile the active π0.5 initialization, persistent gradient insulation, and FAST loss in the supplied code with Table 7 and Eq. (5). If the public configurations differ from the runs, the actual configurations and checkpoint lineage would resolve this concern.

  2. Evidence needed for the mechanism claim: Were the 22K SimDeploy trajectories selected exactly by the supplied success-only collector, and what were the collection success rate and SimAction/SimDeploy sampling weights? Is there a matched-scene, matched-size comparison that separates policy-rollout experience from expanded scene coverage? This would affect my assessment of the claimed benefit of deployment-oriented supervision.

5 comments

We thank the reviewer for the detailed and constructive feedback.

[W1, Q2-2] A comparison against an equal-sized set of successful planner trajectories in the same added scenes would directly test that distinction

We thank the reviewer for this insightful suggestion! To assess whether SimDeploy’s gains result from additional data volume or from its policy-rollout supervision, we generated approximately 22K additional SimAction demonstrations, matching the size of the SimDeploy dataset, and used them to post-train SimVLA with the original SimAction dataset. We evaluated the resulting checkpoint in the real-world Mock Kitchen-ID setup, conducting 20 trials per task and reporting the average task progress with ±1σ, along with binary success rates.

Adding SimAction data produced similar overall performance to the model without SimDeploy, with changes of only one or two successful trials per task. SimVLA with SimDeploy achieved higher success on all five tasks. This comparison suggests that increasing data volume alone does not account for the gains associated with SimDeploy; its policy-rollout supervision on other simulation environments appears to be beneficial.

ModelTask 1Task 2Task 3Task 4Task 5
SimVLA (w/o SimDeploy)0.45 ± 0.32 (3/20)0.38 ± 0.31 (3/20)0.45 ± 0.33 (4/20)0.33 ± 0.34 (2/20)0.33 ± 0.34 (2/20)
SimVLA (w/o SimDeploy + 22K SimAction)0.39 ± 0.31 (4/20)0.32 ± 0.26 (3/20)0.47 ± 0.33 (5/20)0.33 ± 0.28 (3/20)0.33 ± 0.29 (3/20)
SimVLA0.60 ± 0.42 (9/20)0.50 ± 0.36 (6/20)0.52 ± 0.38 (6/20)0.55 ± 0.46 (9/20)0.55 ± 0.43 (8/20)

 

[Q1] Which exact initial checkpoint and saved training configurations produced the reported SimVLA results and ablations?

The reported SimVLA results and ablations were initialized from PaliGemma base, not π0.5\pi_{0.5}; π0.5\pi_{0.5} was used only for the fine-tuned baselines. The training configurations are provided in Appendix G. The supplementary materials include code that matches the paper (Table 7 and Eq. (5)).

 

[Q2-1] Were the 22K SimDeploy trajectories selected exactly by the supplied success-only collector, and what were the collection success rate and SimAction/SimDeploy sampling weights?

We thank the reviewer for the constructive feedback. SimDeploy trajectories are collected automatically by rolling out the policy in RoboCasa, MolmoSpaces, and SceneSmith, retaining only successful episodes. The resulting dataset contains approximately 22K trajectories: 6K from MolmoSpaces, 11K from RoboCasa, and 5K from SceneSmith. For Anubis, the success rate for placing mugs, bowls, and bottles in the sink was approximately 20%, primarily due to failures in locating the sink despite successful grasping. However, success rates varied across tasks and environmental settings. Because the simulator provides ground-truth scene information, including sink positions, cabinet joint states, and robot initial configurations, successful trajectories can be identified automatically without human intervention. We mix SimDeploy with SimAction during post-training at an approximate 9:1 ratio by trajectory count.

We clarified this in Appendix G (SimDeploy) of the revised paper.

Thank you for adding the 22K SimAction control; it helps assess the effect of data volume. Could you clarify whether this new group was evaluated in Mock Kitchen-ID or OOD-Pos, given that the rebuttal labels the comparison as ID but the full SimVLA row matches the OOD-Pos results in Table 12?

We thank the Reviewer 2Q7J for the clarification. It was conducted in Mock Kitchen-ID. We apologize for the confusion. We have modified our initial rebuttal table.

If you have any further questions or if there are remaining points you would like us to clarify, we would be happy to provide additional clarification.

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

SimVLA trains a mobile-manipulation VLA using synthetic action trajectories, simulation-derived VQA supervision, and policy rollouts collected in additional simulated environments. It combines these data with system identification and domain randomization, and evaluates zero-shot deployment on a real mobile manipulator across three environments.

Critical strengths
  1. The paper addresses an important problem and includes actual real-world transfer experiments with coordinated base-arm control.

  2. Combining action, spatial-language, and policy-rollout supervision is a sensible approach to exploiting simulation.

  3. The component ablations and reported binary success counts provide useful evidence beyond qualitative demonstrations.

Critical weaknesses
  1. The mechanism enabling sim-to-real transfer is not sufficiently established. Given the substantial visual gap, I would like to understand what enables transfer with approximately 2,778 hours of SimAction and 460 hours of additional SimDeploy data. The component ablations are useful, but do not clearly separate the effects of pretrained visual representations, visual domain randomization, and spatial supervision. My concern is not that this data scale cannot support transfer, but that the paper provides limited insight into why it works and when it fails.

  2. The extent of required simulation–reality alignment remains unclear. The paper describes robot calibration and procedural scene generation, but does not sufficiently characterize how closely training assets, camera configurations, and scene layouts resemble the real test settings. Since evaluation tasks are included in training, the authors should distinguish transfer to new appearances and layouts from broader task generalization, and clarify how much target-domain knowledge is required.

  3. Language generalization is not demonstrated. SimVLA uses a VLA architecture, but the experiments do not establish whether language supports more than selecting among familiar tasks. There is no dedicated evaluation of unseen instruction formulations, language-driven target changes within the same scene, or held-out task compositions. SimVQA performance alone does not establish these capabilities.

  4. The data contribution is difficult to assess without access and a detailed breakdown. The datasets are not currently publicly available, and the reported totals do not explain how data are distributed across the three embodiments, tasks, and environments. This makes it difficult to judge the effective diversity behind the transfer results or reproduce the central claim.

Questions for authors
  1. Can the authors document the train-test separation for scenes, objects, and tasks?

  2. How exactly are SimDeploy trajectories collected, filtered, and supervised? What are the rollout success rate, human involvement, and mixing ratio with SimAction during post-training?

  3. Can the authors provide a controlled comparison that separates additional data volume from scene diversity and policy-rollout supervision, together with uncertainty estimates for real-world success rates?

  4. Will the datasets, generation pipeline, and checkpoints be released? Can representative data and per-embodiment statistics be made available during rebuttal?

7 comments

We thank the reviewer for the detailed and constructive feedback.

[W1] enabling sim-to-real transfer is not sufficiently established. … The component ablations are useful, but do not clearly separate the effects of pretrained visual representations, visual domain randomization, and spatial supervision.

The mechanism that enables sim-to-real transfer is leveraging three complementary forms of simulation supervision—SimAction, SimVQA, and SimDeploy—within a two-stage VLA training. Our experiments in Tables 11–14 investigate the roles of pretrained visual representations and spatial supervision in enabling zero-shot sim-to-real transfer.

Pretrained visual representations. In the Mock Kitchen-ID setup across five tasks (Tables 11 and 14), SimVLA without pretraining achieves only 18.8% task progress, despite using the same PaliGemma VLM backbone and SimAction (100 procedurally generated kitchens) & SimDeploy (deployment data collected in RoboCasa, MolmoSpaces, and SceneSmith) post-training. It mostly failed to position the robot near the target object or furniture. Adding pretraining that updates only the action head using SimAction increases task progress to 38.8%. This improved the robot's ability to reach the target, but most failures were caused by imprecise grasping. Full SimVLA, which additionally supervises the backbone with tokenized SimAction and SimVQA, further improves task progress to 54.4%. Moreover, SimVLA achieves 53.6% robust performance in Mock Kitchen-OOD-vis (Table 13), demonstrating that its pretraining effectively prepares the model for zero-shot sim-to-real deployment on various visual settings.

Spatial supervision. SimVQA provides spatial supervision during the pre-training stage. The ablation study evaluating its effect is detailed in Section 4.5. Across the five tasks in Mock Kitchen-ID, removing SimVQA reduces average task progress from 54.4% to 43.8%, while removing both SimVQA and tokenized SimAction reduces it to 38.8% (Table 14). Qualitatively, the model pre-trained without SimVQA often fails due to imprecise grasping, suggesting that SimVQA provides valuable spatial supervision for downstream manipulation.

Visual domain randomization. We collect photo-realistic rendering data from IsaacLab with 41 types of domain randomization, spanning camera positions, kitchen environments, object assets (20–80 assets per category), robot physical configurations, material properties, and lighting conditions. Ablation experiments on visual domain randomization would be valuable; however, we couldn’t conduct them due to the computational cost of pretraining. We acknowledged this in Section 6 (Limitations & Future Work).

We clarified our experimental design in Section 4 (Experiments) and made this point clearer in Section 6 (Limitations & Future Work) of the revised paper.

 

[W2, Q1] Not sufficiently characterize how closely training assets, camera configurations, and scene layouts resemble the real test settings

SimVLA does not aim to directly align simulation with real-world settings in terms of scene layouts, assets, or visual appearances. SimVLA enables generalization to unseen real-world environments, demonstrating robust performance in both mock kitchen, pantry and real-home. The target-domain knowledge we used was only the robot embodiment (e.g. kinematics, dynamics, camera configurations) between simulation and the real world.

Training scene details are provided in Appendix A, while test scenes are shown in Figures 4 and 5. Training object details are also provided in Appendix A; real-world test instances are excluded from training. All five tasks used in the real-world evaluation are included in the training data.

We made this point clearer in Appendix I (Experimental Setup) of the revised paper.

[W3] no dedicated evaluation of unseen instruction formulations, language-driven target changes within the same scene, or held-out task compositions.

We conducted an experiment to evaluate language-driven target selection within a fixed scene. Using the instruction “Put <object> to sink,” we varied the target between a bowl and a bottle, with the other object serving as a distractor. We ran 10 trials for each target–distractor pair. The diagonal entries correspond to trials with a single object and no distractor. These results show SimVLA successfully uses language instructions to identify the intended object within a scene, even in the presence of distractors. We expect that training and evaluation across more diverse tasks and generated environments could enable stronger language-level generalization.

Target objectBowl distractorBottle distractor
Bowl6/10 (no-distractor)5/10
Bottle7/107/10 (no-distractor)

Regarding the other two aspects, although SimVLA uses a VLA architecture, our central contribution is generalization to unseen environments for zero-shot sim-to-real mobile manipulation. We respectfully clarify that generalization to unseen instruction formulations or held-out task compositions is not necessary to establish the central contribution of this paper.

We added language-driven target change experiments in Appendix J (Additional Experiments) of the revised paper.

 

[Q2] SimDeploy details

We thank the reviewer for the constructive feedback. SimDeploy trajectories are collected automatically by rolling out the policy in RoboCasa, MolmoSpaces, and SceneSmith, retaining only successful episodes. The resulting dataset contains approximately 22K trajectories: 6K from MolmoSpaces, 11K from RoboCasa, and 5K from SceneSmith. For Anubis, the success rate for placing mugs, bowls, and bottles in the sink was approximately 20%, primarily due to failures in locating the sink despite successful grasping. However, success rates varied across tasks and environmental settings. Because the simulator provides ground-truth scene information, including sink positions, cabinet joint states, and robot initial configurations, successful trajectories can be identified automatically without human intervention. We mix SimDeploy with SimAction during post-training at an approximate 9:1 ratio by trajectory count.

We clarified this in Appendix G (SimDeploy) of the revised paper.

 

[Q3] Controlled comparison that separates additional data volume from scene diversity and policy-rollout supervision

We thank the reviewer for this insightful suggestion! To assess whether SimDeploy’s gains result from additional data volume or from its policy-rollout supervision, we generated approximately 22K additional SimAction demonstrations, matching the size of the SimDeploy dataset, and used them to post-train SimVLA with the original SimAction dataset. We evaluated the resulting checkpoint in the real-world Mock Kitchen-ID setup, conducting 20 trials per task and reporting the average task progress with ±1σ, along with binary success rates. Adding SimAction data produced similar overall performance to the model without SimDeploy, with changes of only one or two successful trials per task. SimVLA with SimDeploy achieved higher success on all five tasks. This comparison suggests that increasing data volume alone does not account for the gains associated with SimDeploy; its policy-rollout supervision on other simulation environments appears to be beneficial.

ModelTask 1Task 2Task 3Task 4Task 5
SimVLA (w/o SimDeploy)0.45 ± 0.32 (3/20)0.38 ± 0.31 (3/20)0.45 ± 0.33 (4/20)0.33 ± 0.34 (2/20)0.33 ± 0.34 (2/20)
SimVLA (w/o SimDeploy + 22K SimAction)0.39 ± 0.31 (4/20)0.32 ± 0.26 (3/20)0.47 ± 0.33 (5/20)0.33 ± 0.28 (3/20)0.33 ± 0.29 (3/20)
SimVLA0.60 ± 0.42 (9/20)0.50 ± 0.36 (6/20)0.52 ± 0.38 (6/20)0.55 ± 0.46 (9/20)0.55 ± 0.43 (8/20)

 

[W4, Q4] Will the datasets, generation pipeline, and checkpoints be released?

Yes. We will release datasets, generation pipelines, and checkpoints of SimVLA. We have included the data-generation code in the supplementary materials.

The following breakdown shows the data distribution. Since our work focuses on five sim-to-real tasks for the Anubis robot, these tasks account for the majority of the data.

RobotEpisodes
Anubis80%
RB-Y115%
ROBOTIS AI Worker5%
TasksEpisodes
Put mug in sink25%
Put bowl in drawer16%
Pour from bottle to mug17%
Put bowl in sink16%
Put bottle in sink14%
Others12%
EnvironmentsEpisodes
peninsula21%
L-shaped24%
U-shaped22%
island14%
single-wall19%

If you have any further questions or if there are remaining points you would like us to clarify, we would be happy to provide additional clarification before the discussion period ends.

Thank you again for your time and thoughtful feedback.

Comments on the paper

2 comments

Dear reviewers, the discussion period is coming to an end. Can you engage with the rebuttal submitted? Thanks, your AC

Dear Area Chair,

We sincerely thank the reviewers for their thoughtful and constructive feedback. Our rebuttal led all reviewers to state that most of their concerns were resolved, and reviewer eMBM raised their score from 2 to 3. As a result, all reviewers now lean toward acceptance. We addressed every comment from all reviewers and revised the paper accordingly, with substantial updates highlighted in yellow.

 

(1) Controlled comparison of SimDeploy data volume [2Q7J, eMBM]

  • Replacing SimDeploy with 22K additional SimAction demonstrations yielded no improvement, suggesting that SimDeploy’s gains stem from its supervision rather than data volume. These details are in Table 11.

 

(2) SimDeploy details [2Q7J, eMBM]

  • We added SimDeploy collection details, including per-source counts, success-only filtering, success rates, and the mixing ratio, to Appendix G.

 

(3) Experiment design and limitation clarification [FSQK, eMBM]

  • We clarified the experimental design and noted the lack of an experiment isolating visual domain randomization in Section 4, Appendix I and Section 6.

 

(4) Code extension [FSQK]

  • We added real-to-sim integration and depth/point-cloud generation support in Appendix K and D, Figure 15.

 

(5) Generalizability beyond kitchen-like settings [FSQK]

  • SimVLA targets zero-shot sim-to-real transfer of 35 mobile-manipulation tasks in kitchen-like environments to unseen real-world scenes, rather than the broad task generalization of generalist VLAs. We stated this limitation in Section 6, together with future work on extending to broader household settings and expanding the simulated task repertoire.

 

(6) Pretraining [eMBM]

  • We will fine-tune the publicly released π0.5\pi_{0.5} base checkpoint on the same SimAction (2,778 hours) and SimDeploy (460 hours) data used to train SimVLA, and compare it with SimVLA in the real-world Mock Kitchen-ID setup. Because both models are finetuned on identical simulation data, this comparison separates the contribution of the SimVLA pretraining. Due to the limited computing resources, we will report the results in the revised paper.

We hope this summary is helpful in assessing our submission. Thank you again for your time and consideration.

Sincerely, 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.