acceptodds
Under review as a conference paper at ICLR 2027

GeoMerge-OPD: Balanced Skill Integration in Vision-Language-Action Models

25 views

Abstract

Vision-language-action (VLA) models excel at individual robotic skills, yet integrating heterogeneous skills into a single policy remains a key challenge for general-purpose robotic control, especially in whole-body tasks with differing manipulation and locomotion requirements. Common multi-skill integration strategies, such as joint fine-tuning and parameter merging, can induce unintended actions during closed-loop execution, severely degrading individual skills. Task-conditioned analysis of expert updates reveals response energy concentrated in a few dominant directions and opposing expert responses along high-energy directions within a shared, task-relevant response subspace. This structure co-occurs with unintended action shifts after merging. We propose GeoMerge-OPD, combining response-space regularization, parameter merging, and on-policy distillation. During independent expert fine-tuning, G-isometry constrains update-response gains within task-activation subspaces. The expert parameters are then statically merged into a student with preliminary multitask capabilities. Fine-tuned task experts serve as teachers, providing velocity-field supervision at the environment observations and intermediate action-flow states visited by the student to calibrate residual behavioral deviations. Experiments with two VLA backbones on LIBERO, LIBERO-PRO, and SIMPLE show that GeoMerge-OPD achieves more balanced performance across skills, significantly outperforms joint fine-tuning and multitask reinforcement-learning baselines, and remains robust under perturbations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Multi-Task Learning, Skill Integration, Model Merging, On-Policy Distillation, Response Regularization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hu18OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents GeoMerge-OPD, a framework for combining multiple VLA experts into a single policy. Its main contribution is a regularizer that makes experts easier to merge. The merged policy is then refined through on-policy distillation. Experiments with two VLA backbones on SIMPLE, LIBERO, and LIBERO-PRO show the promising results.

Critical strengths
  1. The paper offers an insightful approach to improving mergeability by analyzing and regularizing expert responses before merging. Ablations across multiple merging methods and two VLA backbones show that regularized experts merge more successfully than standard experts, while largely preserving individual expert performance. This provides convincing evidence for the practical value of the proposed regularizer.
  2. The paper provides substantial theoretical and empirical support, with carefully designed comparisons. TIES, WUDI, and RegMean++ also use G-isometry-regularized experts, while Joint FT is evaluated from both the pretrained base and the same regularized merge used by the proposed method. These controls make the comparisons more informative and help separate the benefits of initialization from those of subsequent training.
Critical weaknesses

The benefit of velocity-field supervision over final-action distillation remains unclear. Although the paper motivates its objective through a local KL formulation, it does not compare against a simpler baseline that learns from the teachers’ final actions under the same initialization and interaction budget. This makes it difficult to determine whether the gains come from the proposed distillation design or from access to dense teacher supervision more generally.

Questions for authors

Some tasks may benefit from stronger responses along particular directions. Have you observed cases where the regularizer hurts individual expert performance, and how sensitive is this trade-off to the regularization strength?

5 comments
AuthorsOpenReview ↗

Response to Reviewer hu18 (1/2)

【part 1/2】

We would like to thank the reviewer for appreciating our analysis of expert responses and the controlled comparisons. We provide an additional final-action distillation comparison and a regularization-strength ablation on SIMPLE, which we hope can address your concerns.

W1. What does velocity-field supervision add over final-action distillation?

A W1. Thank you for the suggestion. To clarify the benefit of velocity-field supervision, we added a final-action distillation baseline that trains the student using teacher-generated actions as labels. Both settings use the same G-regularized initialization and task-specific teachers, with matched environment-interaction and optimizer-update budgets. This comparison evaluates learning from the teachers' final actions versus teacher velocity-field supervision at the student's intermediate action-flow states.

Teacher targetStudent objectiveSIMPLE (%)LIBERO (%)LIBERO-PRO (%)
Final actionsFlow matching with teacher-generated action labels73.395.765.8
Intermediate velocity fieldsVelocity-field distillation86.797.167.3
Improvement (pp)+13.3+1.4+1.5

Under these matched conditions, velocity-field distillation achieves higher success across all three benchmarks. This shows that supervising the student's intermediate action-generation states with teacher velocity-field targets improves the shared policy's closed-loop performance relative to final-action distillation.

AuthorsOpenReview ↗

Response to Reviewer hu18 (2/2)

【part 2/2】

Q1. Does regularization hurt individual expert performance, and how sensitive is this to its strength?

A Q1. Thank you for the question. To address your concern about regularization strength, we added a regularization-strength ablation on SIMPLE. We test β∈{0,1,2,4}\beta \in \{0,1,2,4\}, with the default strength in the submitted paper set to β0=2\beta_0=2. All groups use the same remaining expert-training configuration and RegMean++ merging setup. For each nonzero strength, we initialize the student with the corresponding merged model and use the task-specific experts trained at that strength as teachers for OPD. The results are as follows:

β/β0\beta/\beta_0BendPick expert (%)CloseDoor expert (%)OpenOven expert (%)Expert avg. (%)Static merge avg. (%)GeoMerge-OPD avg. (%)
0 (Standard FT)90.0100.0100.096.73.323.3
0.5100.0100.0100.0100.020.043.3
1.0 (default)100.0100.0100.0100.040.086.7
2.090.0100.090.093.346.786.7

Note: The β=0\beta=0 OPD result is taken from the fixed-teacher initialization comparison in Table 8 of the submitted paper and is included as a reference.

The following per-task results further show skill retention after OPD at different regularization strengths:

β/β0\beta/\beta_0BendPickCloseDoorOpenOvenGeoMerge-OPD avg. (%)
0 (Standard FT)0/106/101/1023.3
0.53/105/105/1043.3
1.0 (default)8/1010/108/1086.7
2.07/1010/109/1086.7

Within the tested range of regularization strengths, the task-specific experts maintain high overall success rates, with small decreases on individual tasks under stronger regularization. Weaker regularization already improves skill integration, but its performance remains below the default setting. Further increasing the strength yields similar average performance after OPD to the default. The default regularization strength in the submitted paper therefore provides a useful balance between preserving individual expert capabilities and integrating multiple skills.

AuthorsOpenReview ↗

Looking Forward to Your Further Feedback

Thank you for your thoughtful comments and for raising clear and important questions. Your feedback is helping us strengthen the paper, and we sincerely appreciate your time and effort.

In our rebuttal, we have responded to the points you raised. We would appreciate hearing whether our responses have addressed your concerns and whether you have any remaining questions. We look forward to further discussion and suggestions, as they will help us improve the quality of the paper.

If your concerns have been satisfactorily addressed, we would be grateful if you would consider updating your score accordingly.

Reviewer_5L4mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi-skill integration in vision-language-action models. The authors proposes GeoMerge-OPD, a method that combines G-isometry regularization, parameter merging, and multi-teacher on-policy distillation. The method constrains task-conditioned update responses before merging and uses teacher velocity-field supervision along student-visited states and action-generation trajectories. Provided experiments on SIMPLE, LIBERO, LIBERO-PRO, and a Unitree G1 robot show the effectiveness of the proposed skill integration method.

Critical strengths

The paper addresses an important problem in VLA skill integration, where merging specialized experts can degrade previously learned skills. The proposed combination of response-space regularization and on-policy distillation is evaluated across multiple simulation benchmarks and on a real robot, providing useful empirical evidence for its applicability.

Critical weaknesses
  1. The proposed method combines existing parameter-merging methods with on-policy distillation, while the main new component is G-isometry regularization. The paper does not sufficiently establish how this combination advances beyond prior VLA skill-merging work and multi-teacher distillation methods. A clearer distinction from these approaches is needed.

  2. G-isometry regularizes each expert's response geometry, but does not directly resolve conflicts between different experts. The observed reduction in response concentration and improvement in merging performance do not fully establish that the proposed regularizer addresses cross-task interference itself. Stronger comparison with existing regularization methods would help clarify its specific contribution.

Questions for authors
  1. How does G-isometry compare with existing orthogonality or response-space regularization methods under matched training settings?

  2. How much of the improvement comes from on-policy distillation itself, compared with simply fine-tuning the G-isometry-regularized merged model on the task data?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 5L4m (1/2)

【part 1/2】

We would like to thank the reviewer for appreciating the importance of VLA skill integration and our empirical evaluation. We further clarify our contribution and provide additional comparisons with orthogonal regularization and joint fine-tuning, which we hope can address your concerns.

W1. How does GeoMerge-OPD advance VLA merging and multi-teacher distillation?

A W1. Compared with existing VLA skill-merging methods, the main contribution of GeoMerge-OPD is to mitigate capability collapse through response-space regularization before merging and behavioral calibration afterward, enabling stronger and more balanced skill retention. During expert training, G-isometry constrains activation-weighted update responses to improve the mergeability of task-specific experts. We then use the merged shared policy as the student and the task-specific experts participating in the merge as teachers for distillation. Compared with VLA distillation methods such as VLA-OPD and VLA-AD, our refinement focuses on the flow-matching generation mechanism of the action expert. Conditioned on student-visited observations, the corresponding task-specific teacher provides velocity-field targets at intermediate states along the student's own action-flow trajectories, directly supervising action generation and calibrating residual post-merging behavioral deviations.

W2 / Q1. Does G-isometry reduce cross-task interference more effectively than existing regularizers?

A W2 / Q1. Regularizing response geometry aims to control update-response gains and energy concentration within task-activation subspaces, reducing the risk that a few directions dominate merging and thereby improving skill retention. To clarify its contribution, we added the requested comparison with OrthoReg, a parameter-space orthogonal regularizer. We use its official configuration and match the expert training budget and RegMean++ merging setup.

RegularizationEvaluated modelSIMPLE (%)LIBERO (%)LIBERO-PRO (%)
NoneUnmerged task-specific experts96.799.553.1
NoneRegMean++ merge3.352.321.4
NoneRegMean++ merge + OPD23.395.463.3
OrthoRegUnmerged task-specific experts100.099.052.6
OrthoRegRegMean++ merge23.368.232.4
OrthoRegRegMean++ merge + OPD43.396.665.7
G-isometryUnmerged task-specific experts100.099.353.7
G-isometryRegMean++ merge40.073.236.9
G-isometryRegMean++ merge + OPD86.797.167.3

G-isometry achieves better multi-skill merging results across all three benchmarks. In addition, we added OrthoReg to the BendPick–CloseDoor two-expert merging experiment in Appendix D for a diagnostic of response geometry and post-merging behavior. Here, C5C_5 measures the energy share of the top five principal response directions, cBc_B measures the cross-expert response cosine under the common BendPick reference, and VxyV_{xy} measures the mean magnitude of the denormalized BendPick translation command.

RegularizationBendPick C5C_5 (%)CloseDoor C5C_5 (%)Response cosine cBc_BBendPick VxyV_{xy}BendPick successesCloseDoor successes
None78.193.3−0.730.450/105/10
OrthoReg69.775.1−0.480.191/105/10
G-isometry17.614.3−0.310.0343/107/10

Compared with no regularization and OrthoReg, G-isometry yields lower response energy concentration, a cross-expert response cosine closer to zero, and weaker unnecessary translation commands in BendPick. These results provide evidence from response geometry and execution behavior that G-isometry mitigates merging interference in this two-task setting.

AuthorsOpenReview ↗

Response to Reviewer 5L4m (2/2)

【part 2/2】

Q2. How much does OPD improve over fine-tuning the same G-isometry-regularized merged model?

A Q2. Thank you for the suggestion. In Tables 1 and 2 of the submitted manuscript, we compare GeoMerge-OPD with simply fine-tuning the G-isometry-regularized merged model on the task data, denoted Joint FT(G). The results show higher average success for GeoMerge-OPD on all three benchmarks and more balanced skill retention on SIMPLE. To further assess the stability of these gains, we report additional results from multiple training seeds:

BenchmarkJoint FT(G), mean ± SD (%)GeoMerge-OPD(G), mean ± SD (%)Mean improvement (pp)
SIMPLE67.8 ± 5.185.6 ± 1.9+17.8
LIBERO95.8 ± 0.996.9 ± 0.9+1.1
LIBERO-PRO65.0 ± 0.367.4 ± 0.9+2.4

GeoMerge-OPD outperforms Joint FT(G) on every benchmark at each of the three training seeds. Per-seed and per-task results are provided in part 3/3 (A W2 / Q3) of our response to Reviewer vwVa.

AuthorsOpenReview ↗

Looking Forward to Your Further Feedback

Thank you for your thoughtful comments and for raising clear and important questions. Your feedback is helping us strengthen the paper, and we sincerely appreciate your time and effort.

In our rebuttal, we have responded to the points you raised. We would appreciate hearing whether our responses have addressed your concerns and whether you have any remaining questions. We look forward to further discussion and suggestions, as they will help us improve the quality of the paper.

If your concerns have been satisfactorily addressed, we would be grateful if you would consider updating your score accordingly.

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

This paper studies how to consolidate independently fine-tuned vision-language-action experts into a single shared policy. GeoMerge-OPD combines three stages: expert fine-tuning with G-isometry regularization on activation-weighted update responses, static parameter merging, and multi-teacher on-policy distillation. The distillation stage supervises the student’s velocity field at both student-visited environment observations and intermediate states along student-generated action-flow trajectories. Evaluation covers two VLA backbones on SIMPLE, LIBERO, and LIBERO-PRO, together with two real-world manipulation tasks. The results show that G-isometry improves static merging across several merging operators, while subsequent distillation further improves the consolidated policy.

Critical strengths
  1. The evidence that regularizing experts before merging improves mergeability is consistent and practically useful. Tables 6–7 show improvements across all three evaluated merging operators and both backbones, while the individual experts largely retain their task performance. For example, G-isometry improves RegMean++ from 3.3% to 40.0% on SIMPLE and from 52.3% to 73.2% on LIBERO. These substantial improvements across different merging operators are the paper’s strongest empirical finding. They support the usefulness of modifying expert training to facilitate subsequent consolidation.

  2. The framework addresses a relevant robotics problem and includes several thoughtful experimental controls. The resulting policy shares parameters across skills without requiring expert routing at deployment. The experiments include Joint FT initialized from the same G-regularized merge, matched interaction and update budgets for OPD and GRPO-mix, and an initialization ablation using the same teachers. In Table 8, OPD reaches 86.7% on SIMPLE from the G-regularized merge, compared with 30.0% from the pretrained base and 23.3% from an unregularized merge. These controls provide useful evidence that the quality of the merged initialization matters for subsequent distillation under the evaluated training setup.

Critical weaknesses
  1. The distinctive contributions of task-conditioned regularization and two-level on-policy distillation are not sufficiently isolated. The regularization ablation compares G-isometry with no regularization. It therefore leaves unclear how much of the improvement depends specifically on task-conditioned activation weighting, versus more general control of update scale and spectral imbalance. A matched activation-agnostic orthogonality regularizer, such as the approach discussed in Appendix E.2, would make this distinction more informative.

Similarly, comparisons with Joint FT and GRPO-mix change the available supervision: OPD additionally receives dense teacher velocity-field targets. The initialization ablation establishes the importance of initialization, but does not isolate the benefit of the proposed distillation sampling procedure. A comparison with distillation using the same initialization, teachers, and loss but fixed demonstration observations would help establish the contribution of environment-level on-policy sampling. These attribution gaps matter because the response geometry and two-level sampling procedure are central to the claimed methodological contribution.

  1. The reliability of the final gains over strong baselines remains uncertain. Appendix A states that all training uses one training seed. On SIMPLE, the improvement over Joint FT with the same initialization is 26/30 versus 22/30 successes, based on only ten trials per task. On LIBERO and LIBERO-PRO, the improvements over the strongest Joint FT baseline are 0.9 and 2.4 percentage points, respectively. Repeated evaluation seeds do not characterize variation across independent training runs.

These results are promising, but the current evidence provides limited support for the consistency of the final advantage over Joint FT. The larger improvements in the regularization ablations are more convincing than the smaller differences between the strongest final policies. Additional evidence on training variability and evaluation uncertainty could materially change my assessment.

Questions for authors
  1. G-isometry attribution — request for targeted experimental evidence. How does G-isometry compare with an activation-agnostic update regularizer that also controls scale or spectral imbalance, under the same training budget and comparable hyperparameter tuning? This comparison would clarify the contribution of task-conditioned activation weighting.

  2. On-policy sampling — request for a targeted ablation. What happens if student-collected environment observations are replaced with fixed demonstration observations, while retaining the same G-regularized initialization, teachers, flow-distillation loss, and a comparable teacher-supervision budget? This would isolate the environment-level on-policy component from the benefit of teacher supervision.

  3. Reliability of the reported gains — request for additional evidence. Can the authors report results across additional independent training runs for GeoMerge-OPD and the strongest Joint FT baseline, together with uncertainty estimates for the SIMPLE evaluation? In particular, how stable are the improvements over Joint FT across training runs?

4 comments
AuthorsOpenReview ↗

Response to Reviewer vwVa (1/3)

【part 1/3】

We would like to thank the reviewer for appreciating the practical value of regularizing experts before merging and our experimental controls. We provide additional regularization and distillation comparisons, together with results from different training seeds, which we hope can address your concerns.

W1 / Q1. How does G-isometry compare with activation-agnostic regularization?

A W1 / Q1. Thank you for the suggestion. We added the requested comparison with OrthoReg, the parameter-space orthogonal regularizer cited in Appendix E.2, using its official configuration and matching the expert training budget and RegMean++ merging setup. G-isometry uses the configuration in Appendix A. The results are as follows:

RegularizationEvaluated modelSIMPLE (%)LIBERO (%)LIBERO-PRO (%)
NoneUnmerged task-specific experts96.799.553.1
NoneRegMean++ merge3.352.321.4
NoneRegMean++ merge + OPD23.395.463.3
OrthoRegUnmerged task-specific experts100.099.052.6
OrthoRegRegMean++ merge23.368.232.4
OrthoRegRegMean++ merge + OPD43.396.665.7
G-isometryUnmerged task-specific experts100.099.353.7
G-isometryRegMean++ merge40.073.236.9
G-isometryRegMean++ merge + OPD86.797.167.3

The following per-task SIMPLE results further show skill performance after three-expert static merging and after subsequent OPD:

RegularizationEvaluated modelBendPickCloseDoorOpenOvenAverage success rate (%)
NoneRegMean++ merge0/100/101/103.3
NoneRegMean++ merge + OPD0/106/101/1023.3
OrthoRegRegMean++ merge0/103/104/1023.3
OrthoRegRegMean++ merge + OPD1/107/105/1043.3
G-isometryRegMean++ merge3/105/104/1040.0
G-isometryRegMean++ merge + OPD8/1010/108/1086.7

G-isometry achieves better merging results across all benchmarks. On SIMPLE, unregularized and OrthoReg merges still exhibit substantial skill imbalance, whereas G-isometry achieves relatively balanced skill integration. Under matched training budgets and comparable hyperparameter tuning, these results support the added benefit of task-conditioned activation weighting over activation-agnostic parameter-space orthogonal regularization in improving expert mergeability and mitigating post-merging degradation of individual skills, yielding a more balanced shared policy.

AuthorsOpenReview ↗

Response to Reviewer vwVa (2/3)

【part 2/3】

W1 / Q2. What if fixed demonstration observations replace student-collected environment observations?

A W1 / Q2. Thank you for the suggestion. We added the requested ablation of environment-level on-policy sampling by replacing student-collected observations with fixed demonstration observations. Both settings use the same G-regularized initialization, task-specific teachers, and velocity-field loss. Both sample student-generated action-flow trajectories and use the same gradient computation, teacher-supervision budget, and optimization budget.

Environment observationsTeacher supervisionSIMPLE (%)LIBERO (%)LIBERO-PRO (%)
Fixed demonstrationsIntermediate velocity fields66.794.762.8
Student-collectedIntermediate velocity fields86.797.167.3
Improvement (pp)+20.0+2.4+4.5

With teacher supervision held fixed, student-collected observations improve success across all benchmarks. This shows that distillation on environment states visited by the student improves the closed-loop performance of the shared policy.

AuthorsOpenReview ↗

Response to Reviewer vwVa (3/3)

【part 3/3】

W2 / Q3. Are gains over Joint FT stable across training runs, and how uncertain is the SIMPLE evaluation?

A W2 / Q3. Thank you for the suggestion. We added results from two additional independent training seeds. For methods initialized from the G-regularized merge, we fixed the experts and merged initialization and repeated post-training with different seeds.

Seed 2 and Seed 3 are the additional training seeds. We report the mean and sample standard deviation across all three seeds. B and G denote the pretrained base and the merge of G-isometry-regularized experts, respectively.

BenchmarkMethod / initializationSeed 1 (%)Seed 2 (%)Seed 3 (%)Mean ± SD (%)
SIMPLEJoint FT(G)73.366.763.367.8 ± 5.1
SIMPLEGeoMerge-OPD(G)86.786.783.385.6 ± 1.9
LIBEROJoint FT(B)96.294.296.595.6 ± 1.3
LIBEROJoint FT(G)95.295.496.995.8 ± 0.9
LIBEROGeoMerge-OPD(G)97.195.997.796.9 ± 0.9
LIBERO-PROJoint FT(B)64.563.763.663.9 ± 0.5
LIBERO-PROJoint FT(G)64.964.865.365.0 ± 0.3
LIBERO-PROGeoMerge-OPD(G)67.366.568.367.4 ± 0.9

The per-task SIMPLE results for each policy are as follows:

Training seedMethodBendPickCloseDoorOpenOvenTask-macro success rate (%)Improvement over Joint FT(G) (pp)
Seed 1Joint FT(G)5/109/108/1073.3—
Seed 1GeoMerge-OPD(G)8/1010/108/1086.7+13.3
Seed 2Joint FT(G)4/108/108/1066.7—
Seed 2GeoMerge-OPD(G)8/109/109/1086.7+20.0
Seed 3Joint FT(G)4/108/107/1063.3—
Seed 3GeoMerge-OPD(G)7/1010/108/1083.3+20.0

GeoMerge-OPD maintains its performance advantage over the strong Joint FT baselines across all three training seeds on all three benchmarks, demonstrating consistent and reproducible gains across the evaluated training seeds.

AuthorsOpenReview ↗

Looking Forward to Your Further Feedback

Thank you for your thoughtful comments and for raising clear and important questions. Your feedback is helping us strengthen the paper, and we sincerely appreciate your time and effort.

In our rebuttal, we have responded to the points you raised. We would appreciate hearing whether our responses have addressed your concerns and whether you have any remaining questions. We look forward to further discussion and suggestions, as they will help us improve the quality of the paper.

If your concerns have been satisfactorily addressed, we would be grateful if you would consider updating your score accordingly.

Comments on the paper

1 comment

Dear Reviewers,

Please read the authors’ response and let us know whether it addresses your concerns. If any concerns remain, please briefly indicate which ones.

Thank you for your time and for your thoughtful reviews.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.