acceptodds
Under review as a conference paper at ICLR 2027

CroMoGen: Cross-Morphology Demonstration Retargeting for Mobile Manipulation

Abstract

Automated demonstration synthesis for imitation learning avoids costly data collection, but transferring demonstrations across morphologies often produces trajectories that the target robot cannot execute or that collide with obstacles. Mobile manipulation amplifies these challenges as the planner must coordinate base placement and arm motion simultaneously. We present CroMoGen, a learning-free, planning-based framework that synthesizes mobile manipulation demonstrations from different robot and human sources. CroMoGen formulates synthesis as a constrained optimization problem that combines kinematic feasibility, whole-body collision avoidance, and a trajectory alignment objective that preserves task-relevant object interactions. The approach adapts demonstrated motions to the target morphology and environment. Further, unlike methods that decouple base placement from arm motion, CroMoGen jointly optimizes both for greater adherence to the source demonstration. We evaluate CroMoGen on four mobile manipulation task families in RoboCasa and ParaHome. In the challenging human-to-robot setting, CroMoGen reaches a 56% demonstration-generation success rate, exceeding the strongest baseline by 37 percentage points.

Primary area:
applications to robotics, autonomy, planning
Keywords:
imitation learning, demonstation synthesis, planning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_r67nOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CroMoGen, a training-free framework for retargeting cross-embodiment mobile manipulation demonstrations. Using the object motion from the demonstration as a reference, the method generates executable robot trajectories through grasp pose adaptation for the target robot, joint optimization of the base and arm trajectories, and collision detection and repair. The paper evaluates human-to-robot and robot-to-robot cross-embodiment transfer on ParaHome and RoboCasa, and validates the executability of the generated trajectories via real-world experiments with a Spot robot.

Critical strengths
  1. Cross-embodiment experimental validation is extensive. The paper evaluates human-to-robot and robot-to-robot demonstration transfer on ParaHome and RoboCasa across three morphologically distinct mobile manipulators, and provides real-world execution results.

  2. The demonstration generation success rate shows a clear advantage. By jointly planning base and arm motions, CroMoGen significantly outperforms baselines across multiple tasks, proving the practical value of full kinematic and collision constraints for cross-embodiment transfer.

Critical weaknesses
  1. The methodological novelty is clearly insufficient. CroMoGen primarily integrates existing techniques for grasp generation, inverse kinematics, dynamic programming, and collision detection, and the concept of joint base-arm planning is not new. Lacking sufficient ablation studies and direct comparisons with existing joint planning methods, the paper struggles to demonstrate substantial algorithmic contributions beyond mere engineering integration.

  2. Experimental validation remains inadequate. The paper only compares against a few demonstration generation methods, lacking thorough comparisons with recent joint planning approaches. Furthermore, the method relies on known scene geometry and task-specific rules, and the real-world experiments cover limited tasks, which is insufficient to fully validate its generalization and practical applicability.

Questions for authors
  1. The methodological novelty is weak. What is CroMoGen's core algorithmic innovation compared to existing joint base-arm planning methods? Could direct comparisons with these related methods be provided?

  2. Can you provide ablation results that independently remove the contact pose adaptation, joint trajectory planning, and collision repair modules to quantify their respective contributions to the generation success rate?

  3. Can you provide planning time distributions and timeout rates across different tasks and robots to evaluate the method's efficiency for large-scale demonstration data generation?

  4. The current real-world experiments only cover a cup relocation task using an earlier planner version. Can you provide execution success rates for the final method across a broader range of real-world tasks?

4 comments
AuthorsOpenReview ↗

Response to reviewer r67n [1/3]

We thank the reviewer for their detailed feedback and for recognizing the breadth of our cross-embodiment evaluation and the performance gains over baselines. CroMoGen focuses on cross-morphology demonstration synthesis for mobile manipulation: generating training data for a target robot from human or other-robot demonstrations, without collecting new demonstrations on that robot. Below, we clarify our contributions and baseline choices, including comparisons with demonstration-based generators and AutoMoMa. We present new component ablations (Sec. 5.3), VLA policy-learning experiments (Sec. 5.5), and controlled comparisons with cuRobo-based whole-body planners (Appendix E.1) to substantiate our design choices and evaluate the usefulness of the generated training data. Please see our responses to the weaknesses and questions below.

AuthorsOpenReview ↗

Response to reviewer r67n [2/3]

W1, Q1, W2

We thank the reviewer for their constructive feedback. We want to clarify that we are focusing on cross-morphology demonstration synthesis to generate training data for imitation learning models (such as VLA models). To address W1, W2, and Q1, we provide answers from the following perspectives:

I. Research contributions

1) CroMoGen generates a training dataset for a target robot from a human or another robot’s demonstration, without re-collecting data on the target robot. Prior work on demonstration synthesis mostly focuses on cases where the target robot and the demonstration motion source have the same or similar morphologies.

2) Our central question is what to preserve and what to recompute when the source and target morphologies differ. The MimicGen-based baselines and MoMaGen, our closest demonstration-synthesis baselines, reuse demonstrated end-effector motion relative to the object during contact. CroMoGen instead uses demonstrated object motion as the reference and recomputes both the contacts and the coordinated base–arm motion for the target robot, without requiring correspondence between the source and target bodies.

3) Many mobile manipulation methods separate base repositioning from manipulation. CroMoGen jointly replans base and arm motion during object interaction which can be especially important when transferring demonstrations across morphologies.

4) In Appendix A, Table 1, we provide a detailed comparison table for clarifying CroMoGen’s contributions.

CroMoGen advances human-to-robot and robot-to-robot synthesis without requiring joint-to-joint correspondence. We do not claim joint planning, IK, or dynamic programming themselves as new techniques. To clarify contributions, we have updated the introduction.

II. Baseline choices

Because our research problem is to generate a mobile-manipulation dataset for a target robot from a human or another robot’s demonstration, we compare our method with the established demonstration-synthesis methods: MimicGen, WBCMimicGen, and MoMaGen. We are aware that there are methods for humanoid demonstration synthesis, but those methods are specifically designed for humanoids and do not address our full setting, which includes non-humanoid targets.

We also compared CroMoGen with AutoMoMa (CVPR 2026), which generates whole-body mobile manipulation trajectories from task goals without source demonstrations. This compares demonstration retargeting with goal-based planning. On the same 1,587 cases, CroMoGen achieves 62.7% execution success rate versus 38.6% for AutoMoMa, an improvement of 24.1 percentage points (Appendix J.2).

We would also like to receive feedback from the reviewer indicating which baseline methods or planning optimizer alternatives we should consider.

III. Ablation Study of CroMoGen

To motivate each design choice, we added component ablations in Sec. 5.3. We separately disabled the following 4 components of CroMoGen to analyze the performance:

  • Repair and refinement: without any repair or frame insertion, 26.2% of all cases succeed with the initial route, compared with 66.6% success rate for the full CroMoGen pipeline.
  • Adaptive planning-frame insertion: Disabling frame insertion while retaining collision repair reduces execution success rate from 66.6% to 48.1%.
  • Coupled base-arm planning: Decoupling base-arm planning by first planning the base route and then planning the arm motion reduces the success rate from 66.6% to 35.3%.
  • Target-specific grasp adaptation: Using the demonstrator’s grasp instead of an adapted grasp reduces the success rate from 56.2% to 10.6% on the cup-task cases.

These results show that each component is important.

IV. Base-arm planner choice of CroMoGen

Our main comparisons evaluate complete demonstration-synthesis methods. We additionally examine our search design using cuRobo, a fast, GPU-accelerated motion-planning framework also used by MoMaGen. We replace parts of our route search with cuRobo-based planners that jointly plan the base and arm (Appendix E.1, Table 5). Averaged over three runs on 80 Spot cup-task cases, CroMoGen achieves 41.2% execution success rate, compared with 15.8%, 23.8%, and 29.6% for cuRobo’s pose planner, cuRobo candidates with greedy selection, and cuRobo candidates with our DP, respectively.

V. Known scene geometry and task-specific rules

Known scene geometry and task-specific contact rules are explicit assumptions of our method, stated in Section 6. These assumptions are consistent with established demonstration-synthesis approaches: MoMaGen also uses ground-truth object poses and geometry, while both MoMaGen and MimicGen use object-centric subtask annotations. These requirements concern data generation. The VLA policies trained on our demonstrations act from camera images and robot state.

AuthorsOpenReview ↗

Response to reviewer r67n [3/3]

Q2

Can you provide ablation results that independently remove the contact pose adaptation, joint trajectory planning, and collision repair modules to quantify their respective contributions to the generation success rate?

We added component ablations Sec. 5.3 and Appendix E.1, Fig. 9. In the ablation study, we separately disabled the following 4 components:

  • Repair and refinement: without any repair or frame insertion, 26.2% of all cases succeed with the initial route, compared to the full CroMoGen pipeline success rate of 66.6%.
  • Adaptive planning-frame insertion: Disabling frame insertion while retaining collision repair reduces execution success rate from 66.6% to 48.1%.
  • Coupled base–arm planning: We decoupled the base-arm planning by first planning the base route and then planning arm motion. This reduces success rate from 66.6% to 35.3%.
  • Target-specific grasp adaptation: Using the demonstrator’s grasp instead of an adapted grasp reduces success rate from 56.2% to 10.6% on the cup-task cases.

These results show that each component is important.


Q3

Can you provide planning time distributions and timeout rates across different tasks and robots to evaluate the method's efficiency for large-scale demonstration data generation?

We have added these details to the paper: Figure 10 in Appendix F reports success rates against planning time for every method, and Table 6 (Appendix F) reports CroMoGen’s median and 90th-percentile planning times and timeout rates for each task and robot. We also compare CroMoGen and the baseline methods’ planning times in Appendix F, showing the tradeoff between planning time and successful demonstration generation.


Q4

The current real-world experiments only cover a cup relocation task using an earlier planner version. Can you provide execution success rates for the final method across a broader range of real-world tasks?

In our original submission the text “On a physical Spot, 12 of 16 selected cup-transport plans from an earlier planner revision succeed” is misleading: the results in the original submission and now also in the revised version are valid. This was an error by us due to rushed writing. We apologize.

Our comparative performance claims rest on the simulation results which are quite comprehensive. We compare with three demonstration generators in five variants: MimicGen, WBCMimicGen and MoMaGen. The last two were published at ICLR 2026. They are designed for robot demonstrations. All methods are compared across 1,587 human-to-robot cases, 720 robot-to-robot cases and 1,191 RoboCasa cases. Between robots, the best retargeting baseline succeeds in 22.5% of the 480 cross-robot cases, and CroMoGen in 70.8%. CroMoGen also runs on all 529 human demonstrations and three robots, 1,587 cases, with 62.7% success rate. For the comparison with joint planners, see our answer to Q1.

Furthermore, the comparisons between our method and baselines in physics simulation utilize friction grasps instead of commonly utilized assisted grasping (Appendix D, including D.1). The purpose of the method is to generate training data, and the revised paper adds policy learning in closed-loop simulation to show the utility of the generated data (Sec. 5.5). For VLA training and evaluation details, see Appendix I.

Thank you for the new ablations and planner comparisons. They address several of my main concerns, so I am raising my rating from 1 to 2. I am not yet confident enough to recommend acceptance.

Reviewer_KE7kOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces CroMoGen, which retargets human mobile manipulation demos to different embodiments. It formulates a constrained optimization problem that combines kinematic feasibility, whole-body collision avoidance, and trajectory alignment. The object trajectory guides the EEF reference. The proposed solution outperforms previous methods such as MimicGen, WBCMimicGen, and MoMaGen variants.

Critical strengths
  • Both simulation and real-world experiments/demos.
  • The formulation is nice --- fits the problem, joint optimization of base and arm movements, which is important for MM.
  • Breadth of the evaluation, across demo types, different embodiments, tasks, benchmarks, etc.
  • Failure taxonomy that separates planning from execution failures, transparent report.
Critical weaknesses
  • I like a lot of things about this work, however, the ultimate judge of data synthesis quality is whether the data can be used to successfully train policies, either in sim or in real. The paper will be much stronger if such analysis is included and compared to alternative baselines.
  • No within-framework ablation that helps us understand the benefits of designs: coupled vs. base-then-arm, generated grasp vs. the demonstrator's hand pose, etc.
  • Please finish MoMaGen (strongest baseline) experiment runs.
  • What are the reasons for selecting the data sources (e.g., ParaHome)?
Questions for authors

see above

3 comments
AuthorsOpenReview ↗

Response to reviewer KE7k [1/1]

We thank the reviewer for the constructive feedback and the positive assessment of our formulation, evaluation breadth, and failure analysis. In response, we added VLA policy-learning experiments (Sec. 5.5 and Appendix I) to show that the generated demonstration data can be used to successfully train policies. We also added component ablations (Sec. 5.3 and Appendix E.1) and completed the MoMaGen baseline experiments (Fig. 3 and Fig. 4). Below, we address each concern, explain our choice of ParaHome and RoboCasa, and provide responses to raised weaknesses and questions.

W1

I like a lot of things about this work, however, the ultimate judge of data synthesis quality is whether the data can be used to successfully train policies, either in sim or in real. The paper will be much stronger if such analysis is included and compared to alternative baselines.

We agree with the reviewer and have added VLA policy-learning experiments to Sec. 5.5:

  • Policy learning from generated demonstrations. We train VLA policies on demonstrations generated by CroMoGen, MoMaGen, and MimicGen, and on trajectories generated by the goal-based planner AutoMoMa. For comparison with a retargeting baseline, we evaluate the CroMoGen- and MoMaGen-trained policies on scenes where both methods generate successful demonstrations, perturbing the object and robot poses at each policy’s start. The policy trained on CroMoGen data achieves 46.7% success rate versus 18.9% with MoMaGen data, an improvement of 27.8 percentage points. Full results are reported in Appendix I.

These results show that CroMoGen data supports policy learning.


W2

No within-framework ablation that helps us understand the benefits of designs: coupled vs. base-then-arm, generated grasp vs. the demonstrator's hand pose, etc.

We added component ablations Sec. 5.3 and Appendix E.1, Fig. 9. In the ablation study, we separately disabled the following 4 components:

  • Repair and refinement: without any repair or frame insertion, 26.2% of all cases succeed with the initial route, compared to the full CroMoGen pipeline success rate of 66.6%.
  • Adaptive planning-frame insertion: Disabling frame insertion while retaining collision repair reduces execution success rate from 66.6% to 48.1%.
  • Coupled base–arm planning: We decoupled the base-arm planning by first planning the base route and then planning arm motion. This reduces success rate from 66.6% to 35.3%.
  • Target-specific grasp adaptation: Using the demonstrator’s grasp instead of an adapted grasp reduces success rate from 56.2% to 10.6% on the cup-task cases.

These results show that each component is important.


W3

Please finish MoMaGen (strongest baseline) experiment runs.

We completed the MoMaGen experiments and re-evaluated all demonstration-synthesis baselines, using the same cases for all methods within each setting. The revised paper reports MoMaGen on 1,587 human-to-robot and 720 robot-to-robot cases in ParaHome, and 1,191 cases in RoboCasa (Figs. 3 and 4). In ParaHome, MoMaGen reaches a success rate of 20.7% from human motion and 22.5% across robots. CroMoGen reaches 62.7% and 70.8% on the same cases.


W4

What are the reasons for selecting the data sources (e.g., ParaHome)?

We selected the data sources for the following reasons:

  • ParaHome provides synchronized metric 3D body, hand, and object trajectories together with household scene geometry, which supports collision-aware human-to-robot retargeting. Among the alternatives we considered, OMOMO does not provide the same combination of hand and scene information, while access to the full HUMOTO dataset requires a separate research-license application.
  • RoboCasa is a peer-reviewed mobile manipulation benchmark with diverse teleoperated demonstrations and an official MimicGen integration, supporting robot-to-robot evaluation and comparison with established baselines.

Appendix J.1 further analyzes the motion characteristics of the two data sources.

Thanks for the rebuttal, your points are well taken.

We appreciate the time investment in going through the rebuttal very much. Please, let us know if there are specific questions or anything holding back raising the score. We are happy to provide any additional information needed.

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

The paper proposes CroMoGen, a training-free framework that generates robot action sequences from a human demonstration or a demonstration from a different robot. The method contains two key innovations: 1. an object-interaction-centric motion reference; 2. joint optimization of manipulation and mobile-base motion. The results show clear improvements on human-to-robot and robot-to-robot demonstration retargeting tasks over baseline methods.

Critical strengths
  1. The proposed method shows strong cross-embodiment success over baseline methods, most of which fail to perform human-to-robot or robot-to-robot demonstration retargeting tasks.
Critical weaknesses
  1. There is a lack of method ablations to show the benefits of each component. Although the performance gains on cross-embodiment tasks are impressive, there are no results without coupled base–arm planning, without collision repair or adaptive planning-frame insertion, or without target-specific contact adaptation, which makes it difficult to attribute the gains to the core contributions.
Questions for authors
  1. I am curious why, for the baseline methods, human-to-robot performance is much better than robot-to-robot performance, whereas for the authors’ proposed method, human-to-robot performance is worse than robot-to-robot performance. Can the authors explain this?
2 comments
AuthorsOpenReview ↗

Response to reviewer Rrsz [1/1]

We thank the reviewer for the constructive feedback and for recognizing CroMoGen’s strong cross-embodiment retargeting performance. In response, we added component ablations and an analysis of trajectory repair (Sec. 5.3, Fig. 9, and Appendix E.1), and updated the baseline comparisons using the same cases for all methods within each setting (Figs. 3 and 4). Below, we explain these results and clarify the preprocessing and source-selection differences between H2R and R2R, and provide responses to weaknesses and questions.


W1

There is a lack of method ablations to show the benefits of each component. Although the performance gains on cross-embodiment tasks are impressive, there are no results without coupled base–arm planning, without collision repair or adaptive planning-frame insertion, or without target-specific contact adaptation, which makes it difficult to attribute the gains to the core contributions.

We added component ablations Sec. 5.3 and Appendix E.1, Fig. 9. In the ablation study, we separately disabled the following 4 components:

  • Repair and refinement: without any repair or frame insertion, 26.2% of all cases succeed with the initial route, compared to the full CroMoGen pipeline success rate of 66.6%.
  • Adaptive planning-frame insertion: Disabling frame insertion while retaining collision repair reduces execution success rate from 66.6% to 48.1%.
  • Coupled base–arm planning: We decoupled the base-arm planning by first planning the base route and then planning arm motion. This reduces success rate from 66.6% to 35.3%.
  • Target-specific grasp adaptation: Using the demonstrator’s grasp instead of an adapted grasp reduces success rate from 56.2% to 10.6% on the cup-task cases.

These results show that each component is important.


Q1

I am curious why, for the baseline methods, human-to-robot performance is much better than robot-to-robot performance, whereas for the authors' proposed method, human-to-robot performance is worse than robot-to-robot performance. Can the authors explain this?

We thank the reviewer for this observation. We have identified and corrected an inconsistency in the selection of evaluation cases. In the revised paper, all methods run on the same fixed cases in each setting, with revised baseline implementations (Appendix D). The baselines now perform similarly in both settings: 7.6% – 20.7% from human motion and 8.1% – 22.5% from another robot (Fig. 3 and Fig. 4). We have also updated the evaluation code to ensure reproducibility. The revision did not alter the overall conclusion: CroMoGen outperforms strongest retargeting baseline by 42.0 and 48.3 percentage points, respectively.

AuthorsOpenReview ↗

Follow-up on Reviewer Engagement with Rebuttal

We sincerely appreciate the opportunity to address your comments. We understand that you may have been busy, but we would be grateful if you could consider returning to the discussion. We believe some of the concerns raised could benefit from further clarification, and we are happy to provide any additional information you need.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overview to all reviewers and AC

We thank reviewers for the constructive feedback. We have uploaded a revised version of the paper that addresses reviewer concerns. Below we summarize what changed, details are in reviewer responses. Section, appendix, and figure numbers refer to the revised version:

  • New VLA training comparisons to show the utility of the generated data (Sec. 5.5) (Reviewers KE7k, r67n): 
    • VLAs trained using data generated by CroMoGen succeed with perturbations of initial robot and object poses for approximately 40% of valid human-source cases, compared to 5% for the strongest retargeting baseline, MoMaGen, and 23% for AutoMoMa, which plans from task goals rather than retargeting demonstrations. Success requires both successful demonstration generation and policy execution.
    • For VLA training and evaluation details see Appendix I.
  • Ablation study of CroMoGen’s components (Sec. 5.3) (Rrsz, KE7k, r67n):
    • We analyzed CroMoGen’s performance by removing individual components.
    • Outcome: each component is important for generating successful demonstrations.
  • Detailed clarification of our contribution and differences to related work (Appendix A) (r67n):
    • Appendix A discusses properties of recent demonstration synthesis approaches and the novelty of CroMoGen compared to the approaches.
  • Detailed failure case analysis (Appendix J.2, Figure 16) (KE7k, r67n):
    • We conducted a detailed analysis of three mobile demonstration synthesis methods: CroMoGen, MoMaGen, and AutoMoMa. For details, see Appendix J.2, Figure 16.
  • Compute costs (wall-clock time) analysis (Appendix F, Figure 10, Table 6) (r67n):
    • We compared demonstration-generation success rate against planning time for each demonstration synthesis method.
  • The demonstration synthesis baseline comparisons (Sec. 5.2, Figure 3 and Figure 4) (Rrsz, r67n): 
    • We re-ran experiments to fix inaccuracies.
    • Conclusions remain the same: CroMoGen outperforms comparison methods by achieving 62.7% success rate versus 20.7% for the strongest retargeting baseline on 1,587 human-to-robot cases. On 480 ParaHome cross-robot cases, CroMoGen achieves 70.8% versus 22.5%. CroMoGen also achieves the highest aggregate success rate for each of the three RoboCasa target robots.
  • Source demonstration behavior analysis (Appendix J):
    • Analysis of source demonstrations show that demonstrations include overlapping body and hand motion in ParaHome, whereas RoboCasa demonstrations separate navigation and manipulation. See Figures 12-14.
    • This motivates CroMoGen’s design choice: jointly optimize the base motion with manipulation motion.
  • We have improved the presentation of the method and our contributions

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.