acceptodds
Under review as a conference paper at ICLR 2027

A Data-Centric Alignment Study for Cross-Embodiment Transfer

Abstract

Cross-embodiment learning offers a compelling way to scale robot learning by transferring knowledge from other source embodiments to a target embodiment. However, a shared end-effector (EEF) format does not guarantee data compatibility, as embodiment-specific differences can exist even for similar manipulation behaviors. We take a data-centric perspective, studying how explicit data alignment affects this transfer with an otherwise unchanged policy. Specifically, we consider four aspects, including frame coordinate, action mapping across arm configurations, Cartesian states, and observation viewpoints. We evaluate the cross-embodiment transfer through a carefully designed setting around source-only tasks tested on the target embodiment. In real-world skill transfer, data alignment alone achieves 90.5% success, compared with 34.8% for naive data mixing. Further analyses show the benefits of shared tasks across embodiment, indicate that alignment benefits extend beyond a particular embodiment, and provide insights into the roles of different alignment components, as well as the choice of strategies within each component. These findings identify cross-embodiment data compatibility as an important consideration in combining heterogeneous robot data and provide a practical data-centric baseline for cross-embodiment transfer. Real-world videos are in supplementary materials.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot manipulation, Imitation learning, Data-Centric AI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_trEgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a data-level approach to cross-embodiment transfer by evaluating the performance gains achieved through aligning source-embodiment data with target-embodiment data. Specifically, the authors examine four key aspects of data alignment: end-effector frames, arm configurations, proprioceptive states, and camera viewpoints. The first two focus on matching the source embodiment to the target, whereas the latter two aim to mitigate discrepancies in observations between embodiments. Results from both real-world and simulation experiments demonstrate that such data alignment facilitates effective transfer of manipulation skills and object recognition capabilities. Furthermore, this work conducts controlled ablation studies to systematically assess the contribution of each alignment factor to cross-embodiment transfer between two distinct robotic platforms (Moz-1 and Franka).

Critical strengths
  • While many existing approaches to cross-embodiment learning focus on model-level adaptations, this work explores a complementary data-centric perspective. By applying relatively simple modifications to the source data, the proposed strategy substantially improves the cross-embodiment transfer performance, as shown in Figure 5 and Section 4.2.
  • The experimental results in Figure 5 also provide a comparison with the HPT-style baseline, and its relatively low performance suggests that model-level adaptations alone may not sufficiently address the inherent problem of data compatibility.
  • The experiment design incorporates diverse embodiments (Section 4.1.1) for the cross-embodiment transfer, pointing to the potential scalability of the proposed approach to a larger number of embodiments.
Critical weaknesses
  • The arm-configuration alignment strategy based on arm activity (Section 3.3) seems heuristic and raises questions about its applicability to more diverse task setups (e.g., what if both arms exhibit similar levels of activity, or the less active arm plays an important role in manipulation?). Although the authors acknowledge that the less active arm may provide essential support in Section 3.3, it remains unclear how well this strategy generalizes to more complex bimanual behaviors. Furthermore, Appendix E discusses the different effects of alignment on complex manipulation tasks, but does not clearly isolate the limitations of this particular strategy.
  • The object-recognition transfer evaluation (Section 4.1.2) seems somewhat indirect in demonstrating the improvement of object-recognition capabilities. Since success is measured by whether the robot moves toward the correct object and attempts to grasp it, it is unclear whether the performance gain comes from improved object recognition or better action prediction due to data alignment. Further analysis of how the proposed alignment affects the object representations learned by the VLM would help clarify whether the improvement is associated with improved visual understanding rather than primarily better action prediction.
Questions for authors
  • How does the proposed arm-configuration alignment strategy handle tasks where both arms exhibit similar levels of activity or where the less active arm plays an important role in manipulation? Could the authors provide further analysis of the limitations of this heuristic on more complex bimanual tasks?
  • The results in Figure 5 show a substantial performance gap between the proposed alignment strategy and the HPT-style baseline. Have the authors considered combining data-level alignment with the HPT-style approach to investigate whether these two strategies provide complementary benefits?
  • Could the authors provide further analysis of whether the improvement in object-recognition transfer comes from improved visual representations or better action prediction? For example, would it be possible to examine how the learned object representations differ between naive data mixing and aligned data mixing?
2 comments
AuthorsOpenReview ↗

Response to Reviewer trEg

Thank you for recognizing the value of our data-centric approach, the substantial improvements in cross-embodiment transfer, and the diversity of our experimental embodiments. We appreciate your constructive suggestions regarding arm-configuration alignment, object-recognition transfer, and the complementarity with model-centric methods, and address them point by point below.

  • Weakness 1 & Question 1: Scope and limitations of activity-based arm selection. We clarified the scope and limitations of activity-based arm selection and explained why this strategy does not affect our main experimental results.

    We clarify that the activity-based arm selection is only applied when transferring bimanual demonstrations to a single-arm robot. Since a single-arm robot executes actions through a fixed working-arm slot, we adopt this heuristic during training to retain the more active arm from each action chunk. Tasks that inherently require simultaneous coordination of two arms are beyond the capabilities of a single-arm robot, so this transfer necessarily involves a trade-off. Table 3 validates this design, showing that masking the less active arm outperforms random masking and no arm-configuration alignment.

    Importantly, our main results in Figure 5 and Appendix E do not rely on this heuristic, as Moz-1 is a bimanual target. Bimanual-to-bimanual transfer preserves both arms, while single-arm-to-bimanual transfer simply pads an idle arm. Therefore, the concern regarding activity-based arm selection does not affect our main cross-embodiment transfer results.

  • Weakness 2& Question 3: Visual representations. We conducted additional linear probing experiments.

    To distinguish improvements in visual representations from those in action prediction, we conduct an additional linear probing experiment. We use the five objects exclusive to the DROID dataset as listed in Figure 4 and collect 40 visual observations per object in the Franka robot setup, yielding 200 samples. Using "Pick up the object" as the text instruction, we extract the VLM output embeddings from the Naive Data Mixing and Aligned Data Mixing policies. We train linear classifiers on 150 samples and evaluate them on the remaining 50.

    The classification accuracies are 60% for Naive Data Mixing and 64% for Aligned Data Mixing, showing no substantial difference in the linear separability of their learned object representations. This suggests that the gains in object-recognition transfer are more likely driven by better action learning in the action expert rather than differences in visual representations.

    To illustrate, given the same spatial relationship between the robot and an object, the robot should move in the same physical direction to approach it. However, without alignment, this movement may correspond to a positive action value in one embodiment but a negative value in another. Such conflicting supervision can confuse the action expert, even when the VLM correctly identifies the object.

  • Question 2: Complementarity between data-level alignment and HPT modeling. We conducted additional experiments to evaluate the complementarity of data-centric alignment and the model-centric approach.

    As discussed in Section 4.1.3, data-level alignment and model-centric methods, represented by HPT, are not in conflict. To investigate their complementarity, we conduct an additional experiment combining HPT with our aligned data mixing strategy. On the 21 source-only skills, HPT + Aligned Data achieves a success rate of 94.3%, compared with 90.5% for Aligned Data Mixing alone in Figure 5. This further improvement demonstrates that data-level alignment can complement model-centric approaches to achieve stronger cross-embodiment transfer.

Thank you again for your feedback. We hope these clarifications address your concerns and encourage you to reconsider the overall contribution of this paper. We will update the manuscript after the discussion, incorporating all of the feedback received.

AuthorsOpenReview ↗

Invitation for Discussion

Dear Reviewer trEg,

We would greatly appreciate it if you could take a moment to review our response and let us know whether our responses have addressed your concerns. We would be happy to provide any further clarification.

Thank you for your time and consideration.

Reviewer_EXnuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies cross-embodiment robot transfer through four data-alignment steps: coordinate frames, arm configurations, proprioceptive states, and camera views, with an unchanged policy architecture. On source-only skills transferred from nine embodiments to Moz-1, alignment raises success from 34.8% to 90.5%. Bidirectional Moz-1<>Franka experiments, RoboTwin simulation, and ablations further examine transfer and alignment choices.

Critical strengths

Empirically, the paper reports significant benefits of transfer between a source and target embodiment. This increase in source-only skill success from 34.8% to 90.5%, improving over naive mixing on 20 of 21 skills, points to a compelling recipe.

I like how the manuscript examines concrete alternatives, across tables 3-5 (e.g., arm copying versus idle padding, camera masking versus reassignment, and state recentering versus position masking). These comparisons are useful to practitioners.

Critical weaknesses

I am concerned that the experiments do not sufficiently disentangle the effects of the proposed alignment from other beneficial changes to policy inputs and action space design. The experiments demonstrate the complete recipe’s effectiveness (e.g., cartesian state masking) more clearly than they explain which gains specifically arise from cross-embodiment compatibility.

The breath of the evaluation is still somewhat limited: generality beyond the evaluated policy and target settings is unclear. All experiments use Spirit-v1.5 (lines 317–323), and real-world target evaluations cover only transfer between Moz-1 and Franka. The results do not establish how strongly the recipe depends on this specific backbone and its interfaces.

Questions for authors

Section 3.2 establishes frame alignment using corresponding physical end-effector orientations (lines 150-156); Section 3.5 specifies camera mappings for each configuration (lines 208-209). Could you clarify how these reference correspondences, frame offsets, and camera mappings were obtained for each embodiment? Specifically, what information came from existing dataset metadata, and what required manual inspection, access to the robots, additional calibration measurements, or new demos? Were these choices fixed before all subsequent target-task evaluation, or were any revised using evaluation outcomes? Clarifying these requirements would help me assess the external information and the manual effort needed to reproduce the method and extend it to additional embodiments.

To what extent does the real-world performance depend on shared source/target tasks? The ablation removing shared tasks (Section 4.3) shows reduced success, but is only conducted in simulation. Is there evidence that comparable real-world gains persist without task overlap? This matters to my assessment of the method’s generality and practical applicability; I am trying assess the method's dependence on there being shared tasks between the source and target embodiments.

5 comments
AuthorsOpenReview ↗

Response to Reviewer EXnu, Part 1

Thank you for recognizing the substantial improvements in cross-embodiment transfer and the practical value of our alignment strategy and detailed ablations. We appreciate your constructive suggestions regarding alignment effectiveness, embodiment and backbone diversity, reproducibility, and shared task overlap, and address them point by point below.

  • Weakness 1: Attribution of alignment gains. We clarified the experimental controls and further explained how existing ablation results isolate the contributions of individual alignment components.

    Thank you for raising this point. Our existing ablations separately examine action compatibility and observation conditioning on source-only skills absent from the target training data.

    Policy and action representation remain fixed. Naive and Aligned Data Mixing use the same policy architecture, training protocol, and relative end-effector action encoding. Frame and arm alignment map source demonstrations into the target’s conventions and control interface (Sections 3.2–3.3; Appendix C).

    Action alignment improves transfer before visual alignment or state masking. Table 2 shows that frame and arm alignment outperform naive mixing in both transfer directions. Table 3 further shows that activity-based arm selection outperforms random masking, although both retain one arm. This supports the importance of meaningful source–target action correspondence beyond dimensionality matching.

    Visual alignment tests how existing observations are assigned. With state masking disabled, Table 4 shows that role-consistent camera mapping improves transfer over the unaligned setting. Assigning the third-person view to the idle-arm slot also outperforms copying the working-arm view, supporting the importance of observation correspondence.

    State masking is evaluated separately. Holding the other three alignment components fixed, Table 5 shows that Cartesian masking outperforms both retaining the original states and aligning their origins in skill transfer. Its additional contribution is therefore evaluated separately from the preceding action and visual alignment gains.

    These comparisons provide component-level evidence beyond the complete recipe’s effectiveness. We will make these experimental controls and the distinct contributions more explicit in the revised manuscript.

  • Weakness 2: Embodiment and pre-trained backbone diversity. We explained the selection of source and target embodiments, provided further details on interface compatibility, and conducted additional experiments with a non-pretrained policy backbone.

    Source embodiment. Our real-world evaluation covers a broad range of source embodiments. In Figure 5, we report transfer results from 9 diverse source embodiments to Moz-1. In Section 4.4, we deliberately select Moz-1 and Franka, which exhibit substantial differences in arm configuration, workspace, and observation interfaces, for detailed analysis and controlled ablations.

    Target embodiment. We acknowledge that evaluating more real-world target embodiments would further strengthen our conclusions, but such experiments are constrained by the substantial hardware. Given the limited budget, we selected Franka as an additional target embodiment because it differs substantially from Moz-1 in arm configuration, workspace, and observation interfaces, making it a particularly informative setting for evaluating cross-embodiment transfer (Section 4.4).

    Pre-trained backbone. We agree with the importance of verifying that data-centric alignment is not limited to a specific pre-trained policy backbone. We originally chose Spirit-v1.5 because its pretrained weights enable efficient convergence. During this discussion, we additionally conducted experiments using Qwen3-VL-4B as the VLM component and an action expert from scratch. This model follows the same architecture as Spirit-v1.5 but has not undergone any robotics pretraining, requiring 300k training steps. The results are comparable to those in Table 2. For Moz-1-to-Franka transfer, the new policy achieves 81.7% skill-transfer success and 90.0% object-recognition transfer, compared with 76.7% and 92.0% in Table 2. For Franka-to-Moz-1 transfer, it achieves 83.3% and 80.0%, compared with 85.0% and 80.0%, respectively. These results provide additional evidence that the benefits of our data-centric alignment do not depend on the pre-trained Spirit-v1.5 backbone.

    Regarding interface dependence, the pose/frame transformation in Section 3.2 is performed before applying the policy’s original action encoding, and the deployed policy retains the target robot’s native control interface. The alignment rules are therefore defined by the robots’ physical coordinate and action conventions, rather than by Spirit-specific network modules.

AuthorsOpenReview ↗

Response to Reviewer EXnu, Part 2

  • Question 1: Frame alignment details. We provided detailed explanations of the frame and visual alignment procedures, their implementation requirements, and their independence from target-task evaluation outcomes.

    For frame alignment, the coordinate-axis conventions RaxisR^{axis}(Section 3.2, Equation 2) can be directly obtained from the metadata of the publicly available datasets. The additional step is determining the end-effector zero-orientation offset RzeroR^{zero}. Since the robot embodiments used in these datasets are publicly documented, we can obtain their robot descriptions and forward-kinematics models. In their respective simulation environments, we set the source robot and Moz-1 to the same physical end-effector orientation, record the corresponding rotation vectors under their respective coordinate conventions, and compute RzeroR^{zero} using Equation 3. Importantly, the physical orientation is defined jointly by the gripper and wrist camera, since the gripper alone may exhibit rotational symmetry.

    This process requires no access to physical robots, additional calibration measurements, or new demonstrations. To ensure correctness, we additionally replay one source trajectory for each source embodiment after frame alignment in the Moz-1 simulation and manually inspect the resulting motion. For visual alignment, camera mappings are determined by the observation roles of the available views (e.g., wrist and third-person views), as described in Section 3.5, without requiring camera calibration.

    All frame offsets and camera mappings were determined and fixed before the target-task evaluations, and none were revised based on evaluation outcomes.

  • Question 2: Real-world transfer without shared tasks. We conducted additional real-robot experiments to evaluate cross-embodiment transfer without shared tasks.

    During this discussion, we conducted an additional real-world experiment to investigate the dependence of cross-embodiment transfer on shared tasks. We select 3 source-only skills from RoboCOIN: "Unscrew a bottle cap," "Pass a bowl between the two arms," and "Clink two cups together." We choose RoboCOIN here because its data are organized by task, allowing us to isolate individual skills more easily than in DROID. Considering the training cost, we focus on these 3 representative skills rather than all 21 source-only skills. As shown in Table 11, these three skills achieve an average success rate of 30.0% with Naive Data Mixing and 96.7% with Aligned Data Mixing.

    We train a new policy using Moz-1 data combined with only the demonstrations of these 3 RoboCOIN skills, excluding all other data. The resulting policy achieves success rates of 70%, 10%, and 60% on the three skills, respectively, averaging 46.7%. Consistent with the simulation results in Section 4.3, this demonstrates that meaningful real-world transfer remains possible without shared source-target tasks, although performance is lower than the 96.7% achieved with the full source data.

Thank you again for your feedback. We hope these clarifications address your concerns and encourage you to reconsider the overall contribution of this paper. We will update the manuscript after the discussion, incorporating all of the feedback received.

Thank you for the clarifications and additional experiments. The explanation of alignment requirements and confirmation that choices were fixed before evaluation address my reproducibility concern; this information should be made explicit in main content. Tables 2-4 also demonstrate alignment gains before state masking, reducing my attribution concern.

Two points remain:

  • The 46.7% success without shared tasks demonstrates some real-world transfer. However, the comparison with 96.7% also changes source-data quantity and diversity. Is a naive-mixing baseline available for the same three-skill subset?

  • The Qwen-based experiment reduces concern about dependence on Spirit’s robotics pretraining, but retains the same architecture. Is a matched naive-mixing baseline available to establish alignment’s benefit in this setting?

I maintain my weak-accept recommendation and encourage incorporating these clarifications, new results, and remaining limitations into the manuscript.

Reviewer_SCw2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper attempts to achieve cross-embodiment transfer when training on datasets that do not necessarily have aligned observations or actions. To achieve this, it matches observation slots using intuition, masks out position information from proprioceptive state, and aligns actions to a target embodiment's reference frame. The results show that alignment alone can result in 90.8% success in improvements over naive data mixing that achieves 34.8% and the necessity of having shared tasks across different embodiments.

Critical strengths
  1. Comprehensive real robot and simulation study. The paper does a principled study of generalization to objects and tasks unseen in the target embodiment. Further, controlled studies on the importance of sharing tasks across embodiments is valuable.
  2. Use of raw released datasets. The paper combines datasets from multiple sources on the internet, and uses them to demonstrate transfer.
  3. The method is intuitive, simple and easy to follow. For example, the use of wrist camera of resting bimanual arm as third person for single arm and vice versa is an effective workaround.
Critical weaknesses
  1. Lack of technical novelty. The paper does not propose a methodological contribution. The paper proposes a data alignment strategy that is to some extent already used in prior cross-embodiment works [1].
  2. Need for manual alignment affects scalability. The paper relies on manual effort for aligning action conventions and on paired samples across embodiments. Further, it relies on manual tricks for aligning observation images.
  3. Per-embodiment model. The paper targets all source data to target embodiment, requiring specialized training of a new model for each embodiment. In contrast, prior work [1, HPT, Octo] typically trains a single backbone, and demonstrates how it can transfer to multiple embodiments.
  4. Missing comparisons to other cross-embodiment works, example latent actions (LAPA, Ye et al. 2024) that are more scalable.
  5. The method may not cleanly extend to egocentric human data with moving cameras. Assumption of the other hand not moving in bimanual data or not affecting motion may not always be true across datasets.

[1] Pushing the Limits of Cross-Embodiment Learning for Manipulation and Navigation. Yang et al. RSS 2024.

Questions for authors
  1. Typo: L131 Even if

Please refer to weaknesses. Request authors to clarify novelty over prior works on alignment, need for manual alignment, scalability and generalizability limits, and missing comparisons to other approaches.

6 comments
AuthorsOpenReview ↗

Response to Reviewer SCw2, Part 1

Thank you for recognizing the value of our comprehensive real-world and simulation experiments, the use of publicly available datasets, and the simplicity of our alignment strategy. We appreciate your constructive feedback regarding technical novelty, alignment requirements, scalability, comparisons with prior work, and generalizability, and address these concerns point by point below.

  • Weakness 1: Our novelty. We clarified our contributions relative to prior work and further explained the novelty and scope of our study.

    Yang et al. primarily align action spaces through manual axis permutations and sign adjustments, without explicitly addressing inconsistencies in rotation reference conventions, including zero orientations. Their manipulation evaluations mainly focus on relatively simple reaching and grasping behaviors, leaving transfer across more diverse and complex manipulation skills largely unexplored. Moreover, although they demonstrate positive transfer between navigation and manipulation, they do not explicitly investigate whether knowledge acquired from one dataset can transfer to another embodiment, such as whether a robotic arm can recognize and grasp an object that appears only in navigation data.

    As also recognized by reviewers EXnu and trEg, our work provides a systematic investigation of cross-embodiment transfer. Through carefully designed transfer metrics and evaluations across a broader range of complex manipulation skills, we demonstrate that our contribution lies not in individual alignment components, but in systematically studying multiple interconnected sources of data incompatibility and identifying effective alignment pipelines. We aim to provide researchers with a systematic understanding and a simple, practical alignment pipeline for leveraging diverse open-source datasets to learn complex manipulation skills when target-embodiment data are limited.

  • Weakness 2: Manual effort. We provided additional details on the alignment procedures and explained their little manual requirements.

    For frame alignment, the coordinate-axis conventions (RaxisR^{axis}) (Section 3.2, Equation 2) can be directly obtained from publicly available dataset metadata. To determine the zero-orientation offset (RzeroR^{zero}), we use publicly available robot descriptions and forward-kinematics models to set the source and target robots to the same physical end-effector orientation in simulation, then compute the offset using Equation 3. This process only requires visually checking the robots' orientations in simulation and inspecting the replay of a single aligned trajectory. It requires no specialized expertise, physical robots, additional calibration, or new demonstrations, and is performed only once per source embodiment.

    For visual alignment, we clarify that the observation mapping is primarily necessary for single-arm embodiments. Meanwhile, many existing robot platforms already adopt a common camera configuration consisting of two wrist cameras and an overhead camera. For these platforms, observations can be directly mapped according to their camera roles, without additional calibration or manual image manipulation.

  • Weakness 3: Per-embodiment model. We provided further explanations of the unified policy design and conducted additional experiments validating cross-embodiment transfer with a single model.

    We agree that our original experiments train separate models for the two transfer directions, as acknowledged in Section 5, and we will continue exploring more flexible unified policies. We found that our alignment framework is inherently compatible with a single model controlling multiple embodiments. Specifically, as described for Franka-to-Moz-1 in Section 3.3, we randomly assign Franka's single-arm actions to either the left- or right-arm action slots during training. At inference time, we just need simply fix one slot (e.g., the left-arm slot) to control a single-arm robot, e.g. Franka. Note that more sufficient training is required to achieve strong performance on Franka, since its single-arm actions are randomly assigned to either the left- or right-arm slot during training, rather than consistently using a fixed slot. The predicted actions are then transformed back to Franka's original coordinate conventions before being passed to its controller.

    To validate this capability, we conducted an additional experiment training a single model jointly on Franka and Moz-1 data and evaluating the same checkpoint on both embodiments. On Franka, the unified model achieves 81.7% skill-transfer success and 90.0% object-recognition transfer, compared with 76.7% and 92.0% in Table 2. On Moz-1, it achieves 83.3% and 80.0%, compared with 85.0% and 80.0%, respectively. These comparable results demonstrate that our alignment framework does not inherently require per-embodiment models and can support bidirectional transfer within a single policy.

AuthorsOpenReview ↗

Response to Reviewer SCw2, Part 2

  • Weakness 4: More comparisons with model-centric methods. We explained our choice of existing baselines and conducted additional experiments to evaluate the complementarity of data-centric and model-centric approaches.

    As discussed in Section 4.1.3, we did not extensively compare with model-centric approaches because they are complementary to, rather than competing with, our data-centric alignment framework. Therefore, we selected HPT as a representative model-centric baseline in our paper. During this discussion period, to further demonstrate this complementarity, we add an additional experiment combining HPT with our data alignment. On the 21 source-only skills, HPT + Aligned Data achieves a success of 94.3%, compared with 90.5% for Aligned Data Mixing alone and 38.6% for HPT alone (Figure 5). These results demonstrate that our alignment framework can further improve model-centric approaches.

    Considering the substantial training costs, we initially limited our comparisons to HPT. We will include comparisons with additional representative model-centric methods in the camera-ready version.

  • Weakness 5: Egocentric data and arm configuration. We clarified the scope and applicability of our visual and arm-configuration alignment strategies and discussed their limitations.

    Egocentric data. We thank the reviewer for encouraging us to explore human egocentric data. We clarify that our visual alignment does not assume stationary cameras, but only requires consistent camera roles, such as distinguishing wrist-mounted views from global views. Meanwhile, human egocentric data often lack the precision required for fine-tuning robot policies and are more commonly used for pre-training, making them less suitable for our direct real-robot evaluation setting.

    Nevertheless, we acknowledge that human egocentric videos with independently moving head-mounted cameras may introduce additional challenges. Although our current work primarily focuses on teleoperated robot demonstrations, extending our framework to such data is an interesting direction for future work.

    Arm configuration. We clarify that activity-based arm selection is only applied when transferring bimanual demonstrations to a single-arm robot. Under a limited training budget, when the goal is to deploy a policy on a single-arm robot, converting bimanual demonstrations to its action interface necessarily involves a trade-off. We therefore retain the more active arm within each action chunk, without assuming that the other arm is stationary or irrelevant. We acknowledge that tasks requiring coordinated bimanual manipulation cannot be fully transferred to a single-arm robot. Importantly, Table 3 validates our design choice, showing that masking the less active arm achieves the best performance among the evaluated strategies, outperforming random masking and no arm-configuration alignment.

  • Question 1: Typo. We will correct the typo in the revised manuscript.

    Thank you for pointing out this typo. We will correct it in the revised manuscript.

Thank you again for your feedback. We hope these clarifications address your concerns and encourage you to reconsider the overall contribution of this paper. We will update the manuscript after the discussion, incorporating all of the feedback received.

I thank authors for the detailed rebuttal, which provides clarifications and additional results.

However, I still believe the methodological contributions of the work are limited. The work provides analysis that shows aligning actions (including rotation frames) and observation slots across different datasets results in empirical gains. The findings are mostly along expected lines. Moreover, as mentioned in the review it is a standard practice to align reference frames for grippers (Casas et al. [1] aligns palms for better cross embodiment transfer, [2] performs explicit alignment of wrist coordinate systems in human data with collected robot data to fixed reference). Further, the RoboCOIN dataset [3] that provides a subset of embodiments used in the paper also ensures that the rotation matrices for end effector orientations are expressed as 6D rotation matrices.

My other remaining concern (Weakness 5) was that proposed alignment tricks may not provide additional benefits when embodiments significantly differ from the target embodiment (eg. single first-person camera, multifinger data, multi-arm coordination data), and is thus not scalable compared to other works developed to scale to diverse data (HPT, LAPA). The authors acknowledge these limitations in the rebuttal.

Thank authors for clarifying the manual effort. The alignment process still seems to involve manual checks, demonstrating unified model can be trained to control multiple embodiments and improvements over HPT. However, it is unclear to me where the complementary gains over HPT are coming from. Since HPT uses separate action heads and observation encoding stems for each embodiment, additional explanation of how the proposed data alignment strategies are helping here would be useful. Further, remaining comparisons to prior cross-embodiment works are not entirely model-centric comparisons. For example, LAPA (Ye et al.) predicts learned latent actions in place of end-effector actions used in paper, and can scale to diverse data. The authors promise to include additional comparisons in camera ready version. I believe these comparisons should study how different approaches scale to more diverse datasets.

[1] MultiGripperGrasp: A Dataset for Robotic Grasping from Parallel Jaw Grippers to Dexterous Hands. Casas et al. IROS 2024.

[2] METIS: Multi-Source Egocentric Training for Integrated Dexterous Vision-Language-Action Model. Fu et al. arxiv 2025. NeurIPS 2026.

[3] RoboCOIN: An Open-Sourced Bimanual Robotic Data COllection for INtegrated Manipulation. Wu et al. arxiv 2025.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Author Response Summary

We sincerely thank all reviewers for their thoughtful and constructive feedback, as well as their recognition of the value of our work. We are particularly encouraged that all three reviewers acknowledge the substantial improvements in cross-embodiment transfer achieved through our data-alignment approach. The reviewers also highlighted the simplicity and practical value of our approach, the carefully designed transfer evaluations on tasks and objects unseen in the target embodiment, the diversity of embodiments and datasets, and the comprehensive real-world and simulation experiments with controlled ablation studies.

The reviewers also provided valuable constructive feedback, primarily regarding:

  • The novelty of our data-centric approach and its relationship to prior cross-embodiment methods.
  • The scalability and generalizability of the alignment process, including manual effort, embodiment-specific training, and diverse manipulation settings.
  • The attribution of performance improvements to individual alignment components and their compatibility with existing approaches.
  • The breadth of experimental validation, including backbone diversity and the dependence on shared tasks across embodiments.
  • The effectiveness of arm-configuration alignment in complex bimanual manipulation and the underlying mechanisms of object-recognition transfer.

We have carefully addressed this feedback through further clarifications, explanations, and detailed analyses. Beyond these clarifications, we have made additional efforts to strengthen the paper, including:

  • Continued investigation of our proposed future direction: Building on the future work outlined in our original submission, we trained a unified cross-embodiment policy using Qwen3-VL-4B and an action expert from scratch, without robotics pretraining, and evaluated the same checkpoint on both Franka and Moz-1.
  • Combination with model-centric methods: Conducted additional experiments combining HPT with aligned data mixing to evaluate their complementarity.
  • Real-world transfer without shared tasks: Conducted additional real-robot experiments using three source-only manipulation skills without shared source-target tasks.
  • Visual representation analysis: Conducted additional linear probing experiments to investigate the contributions of visual representations and action learning to object-recognition transfer.

We sincerely thank the reviewers again for their valuable feedback and constructive suggestions, which have helped us further strengthen our work. We greatly appreciate the time and effort devoted to reviewing our paper. We hope our responses adequately address the reviewers' concerns and encourage them to reconsider the overall contribution of our paper. We will carefully revise the manuscript after receiving further feedback from the reviewers.


Summary of Revision Before the Discussion Deadline

Following our discussions with the reviewers, we have revised the manuscript to incorporate their feedback. We sincerely thank all reviewers again for their constructive comments and suggestions. The main updates are summarized below.

  1. Revised Section 4.2: Added HPT-Style + Aligned Data Mixing experiments to demonstrate the complementarity of data-centric and model-centric approaches.
  2. Revised Section 4.3: Added real-world experiments without shared skills to validate cross-embodiment transfer in this setting.
  3. Revised Section 5 Paragraph "Limitations": Clarified that our study focuses on teleoperated robot demonstrations with explicitly recorded actions.
  4. Added Section 5 Paragraph "Future work": Presented results from a single model controlling both Franka and Moz-1 to explore unified cross-embodiment policies.
  5. Revised Appendix C: Added training-from-scratch experiments to examine the dependence on robotics-specific pretraining.
  6. Added Appendix F: Detailed the implementation of frame alignment, highlighting its minimal manual effort and lack of additional real-robot calibration.
  7. Added Appendix G: Conducted linear probing experiments to investigate whether object-recognition transfer gains arise from visual representations or action learning.

As we near the end of the discussion period for this submission, please check whether the authors have answered questions or responded to critiques. If they have, we encourage the reviewers to leave a comment for the authors. If reviewers have raised new points, a gentle nudge to the authors to respond would also be appreciated.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.