acceptodds
Under review as a conference paper at ICLR 2027

AdaMoE-VLA: Adaptive Mixture of Action Experts for Vision-Language-Action Learning

19 views

Abstract

Vision-Language-Action (VLA) models show strong potential as generalist robot policies, yet enhancing their action-generation capability remains challenging due to the diversity of robot behaviors and the need for fine-grained action modeling. While Mixture-of-Experts (MoE) offers an efficient way to increase model capacity through sparse activation, its role in VLA is not simply generic model scaling. Long-horizon manipulation trajectories often contain heterogeneous and frequently switching action modes, such as reaching, grasping, aligning, inserting, and recovery behaviors, making multi-expert action modeling especially suitable for VLA policies. However, directly applying conventional MoE to VLA action heads introduces an optimization conflict: the router is required to both select experts under a load-balancing objective and provide precise state-dependent expert weighting for continuous action generation. This coupling can interfere with fine-grained action learning, particularly when behavior modes change frequently. Therefore, we propose AdaMoE-VLA, an adaptive mixture of action experts for VLA learning. AdaMoE-VLA decouples expert selection from expert weighting by preserving the standard router for trainable sparse selection while introducing a lightweight Scale Adapter for task-conditioned expert reweighting. This design mitigates gradient interference between load balancing and action prediction, enabling task-adaptive expert reweighting while preserving real-time inference. Across simulation and real-world benchmarks, AdaMoE-VLA consistently improves -series VLA baselines, and its gains become more pronounced as action-mode transitions become more frequent. It reaches 98.3% on the LIBERO benchmark, while delivering larger improvements in more behaviorally diverse settings, with a 9.3% gain on RoboTwin 2.0 and a 20% – 21.5% absolute gain in real-world experiments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, Mixture of experts, Load Balancing, Robotics

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_jh4GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AdaMoE-VLA, a mixture-of-experts architecture for vision-language-action models. It introduces a Scale Adapter to adjust the contributions of selected action experts, aiming to reduce interference between load-balancing regularization and task-dependent expert weighting. Experiments on LIBERO, RoboTwin 2.0, and real-world manipulation tasks show improvements over dense VLA baselines and vallina MoE.

Critical strengths
  1. Simple and practical architectural design. The proposed Scale Adapter provides additional flexibility in expert weighting while retaining standard sparse expert selection. The method can be integrated into existing VLA architectures with relatively limited modifications.
  2. Detailed experimental results. The method achieves improvements across simulation and real-world manipulation tasks. Comparisons with Vanilla MoE, architectural ablations, and supplementary analyses provide useful empirical support for its effectiveness.
Critical weaknesses
  1. Insufficient validation of the proposed expert-weighting design. AdaMoE-VLA only marginally outperforms CSMoE on LIBERO (96.0% vs. 95.5%), and CSMoE is not evaluated on RoboTwin 2.0. Since RoboTwin involves more diverse action patterns and frequent mode transitions, this missing comparison makes it difficult to establish whether additive residual weighting provides a meaningful advantage over alternative expert-weighting designs.

  2. Room for more comprehensive architectural evaluation. While the paper demonstrates improvements with both π0 and π0.5, the key comparisons among Vanilla MoE, CSMoE, and AdaMoE-VLA are primarily conducted with π0. Extending these comparisons to π0.5, along with including a Vanilla MoE baseline in the real-world experiments, would provide stronger evidence for the generalizability of the proposed design and help further isolate the contribution of the Scale Adapter.

Questions for authors
  1. Could the authors evaluate CSMoE on RoboTwin 2.0 under the same configuration as AdaMoE-VLA? Given the small performance gap on LIBERO, this comparison would help establish whether additive residual weighting provides a meaningful advantage on more complex manipulation tasks.

  2. Could the authors provide the key architectural comparison among Vanilla MoE, CSMoE, and AdaMoE-VLA using π0.5? Additionally, a Vanilla MoE baseline in the real-world experiments would help isolate the specific contribution of the Scale Adapter.

  3. Could the authors clarify how the combined router–Scale Adapter gradient alignment is computed, particularly how gradients from different parameter groups are aggregated and normalized? Since adding task-only adapter gradients can reduce the angle in the joint parameter space without necessarily reducing interference in the router itself, what evidence supports attributing the performance gains specifically to gradient-interference mitigation rather than increased expert-weighting flexibility?

1 comment

Thank you for the responses and additional experiments. The updated results address my first two concerns, particularly regarding the comparison with CSMoE and the validation across different backbones and evaluation settings. Regarding the third concern, the additional gradient analysis provides useful insights, although further validation could strengthen the evidence for the proposed gradient interference mitigation mechanism. Overall, I lean towards weak accept.

Reviewer_wbT4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies action generation in vision-language-action models using a mixture of experts. In standard MoE, the router determines both expert selection and contribution weights while also receiving a load-balancing loss. The authors argue that this coupling can interfere with learning the weights needed for accurate action prediction.

AdaMoE-VLA adds a Scale Adapter that predicts input-dependent corrections to the selected experts’ router weights. The adapter is not directly affected by the load-balancing loss at the same layer. The method modifies the action generator of π0\pi_0 and π0.5\pi_{0.5}, using a shared expert and sparsely activated routed experts initialized from pretrained weights.

The experiments cover LIBERO, RoboTwin 2.0, and real-world manipulation. AdaMoE-VLA reaches 98.3% average success on LIBERO with π0.5\pi_{0.5} and improves RoboTwin success from 40.4% to 49.7% with π0\pi_0. Comparisons with vanilla MoE show higher average success on both benchmarks, and the real-world experiments show improvements on four short tasks and the stages of a longer manipulation task.

Critical strengths
  • The experiments cover two simulation benchmarks and real-world robot tasks, with LIBERO results using both π0\pi_0 and π0.5\pi_{0.5}. The improvements across these settings suggest that the method works beyond one particular training setup. The ablations also examine the number of experts, top-k selection, and load-balancing weight.
  • The comparisons with vanilla MoE support the benefit of the Scale Adapter beyond adding experts. With four routed experts and top-1 selection, AdaMoE improves average LIBERO success from 94.8% to 96.0%, including an increase from 87.4% to 92.0% on LIBERO-Long. It also improves average RoboTwin success from 45.9% to 49.7%. These results show that the proposed weighting mechanism improves on standard routing in the tested settings.
Critical weaknesses

The experiments definitely show that the adapter helps, but the justification of the mechanism is not as convincing. Equation 7 shows that the local balancing loss does not directly update the adapter, and Figure 3 shows better alignment for the combined router-adapter gradient. However, the smaller angle does not by itself show that reduced balancing interference explains the performance improvement. The adapter also gives the model more flexibility to weight expert outputs, and the current comparisons do not clearly separate these two explanations.

Questions for authors

Do the authors have evidence that separates the benefit of reducing balancing interference from the benefit of more flexible expert weighting? This would help explain whether the results establish the proposed mechanism in addition to showing that the adapter works.

Reviewer_HRD8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AdaMoE-VLA, which applies a sparse mixture of experts to the action expert of π-series VLAs. A standard router selects the top-k experts and is trained with a load-balancing loss. A separate Scale Adapter adds task-conditioned residual weights to the selected experts. The authors argue that this partially separates balanced expert selection from state-dependent expert weighting, reducing interference between load balancing and continuous-action learning.

Critical strengths
  1. The paper targets a relevant problem: increasing action-model capacity while keeping inference sparse and addressing action-mode diversity in manipulation.
  2. The empirical evaluation spans two simulation benchmarks and real-robot experiments. The authors include comparisons against Vanilla MoE and a concatenated adapter variant, as well as sweeps over expert count, top-k, and load-balancing weight.
  3. The paper provides useful implementation and diagnostic details. The loss-free balancing comparison is a meaningful additional control, although it does not resolve the gradient-analysis concern below.
Critical weaknesses
  1. The decoupling claim is overstated. Equation 6 retains router probabilities in expert weights, while Equation 7 establishes only local gradient independence. Neither establishes full decoupling or compensation for balancing interference.

  2. The gradient analysis is inconclusive. Appendix H does not clearly define how router and adapter gradients are combined. Adding task-only gradient coordinates can mechanically reduce the combined angle without correcting router interference. Direct task–balance gradient comparisons are needed.

  3. The mechanism is not sufficiently isolated. Existing ablations support empirical effectiveness, but a zero-balancing-loss ablation and a matched scaling control would help distinguish interference mitigation from additional task-conditioned capacity.

  4. Statistical robustness remains unclear. LIBERO gains are modest, RoboTwin improvements vary across tasks, and long-horizon real-robot evaluation uses only 10 trials per stage. Repeated runs and confidence intervals would strengthen the evidence.

  5. The action-mode explanation remains correlational. Cross-dataset comparisons do not isolate whether greater mode diversity or switching frequency causes larger gains.

Questions for authors

Please see weaknesses.

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

This paper proposes AdaMoE-VLA, a sparse mixture-of-experts architecture for the action module of π-series vision-language-action policies. The method combines an always-active shared expert with sparsely selected routed experts and introduces a Scale Adapter that adds input-conditioned residual adjustments to the selected experts’ routing weights. The motivation is to reduce interference between load balancing and task-dependent expert weighting. The evaluation covers LIBERO, 19 RoboTwin 2.0 tasks, and real-world dual-arm manipulation. The method achieves 98.3% average success on LIBERO with π₀.₅ and improves RoboTwin success from 40.4% for π₀ and 45.9% for vanilla MoE to 49.7%. The paper also reports improvements on real-world tasks and provides architectural ablations, gradient analysis, and computational-efficiency measurements.

Critical strengths
  1. A focused architectural modification compatible with pretrained VLA policies. The method expands the action module while retaining the vision-language backbone and action-generation formulation. Copying pretrained feed-forward weights into experts makes the approach practical for adapting existing policies. Improvements with both π₀ and π₀.₅ on LIBERO support its applicability beyond a single base model.
  2. Evaluation across simulation and physical robots, with relevant architectural comparisons. The experiments extend beyond LIBERO to RoboTwin and real-world manipulation. Comparisons with vanilla MoE and auxiliary-loss-free balancing provide useful evidence that the additional weighting pathway can help under the evaluated settings.
Critical weaknesses
  1. The incremental gains are limited, and the comparisons do not establish an advantage over further baseline training. AdaMoE improves over vanilla MoE by 3.8 percentage points on RoboTwin (45.9% → 49.7%) and by 1.2 points in the controlled four-expert, top-1 LIBERO setting (94.8% → 96.0%). Although the methods use the same number of training steps, the paper does not provide success-rate curves against training steps or cumulative compute, nor establish that the baselines have converged. Consequently, it remains possible that allocating additional compute to training π₀ or vanilla MoE longer could achieve comparable or better performance. Given the additional architectural complexity and inference overhead, the current results do not sufficiently establish that these incremental gains offer a compelling practical advantage.
  2. The gradient analysis does not sufficiently establish that the improvements result from mitigating harmful optimization conflict. Appendix H compares the total gradient with the pure task gradient. A nonzero angle demonstrates that the auxiliary balancing loss changes the update direction, but does not by itself establish that the objectives have opposing gradients or that this change harms task learning. Likewise, a smaller combined-gradient angle after adding an adapter with a task-supervised pathway does not alone demonstrate that the original router’s conflict has been reduced. The analysis is a useful diagnostic, but it does not sufficiently establish the proposed causal connection between load-balancing interference, its mitigation by the adapter, and improved policy performance.
Questions for authors
  1. Have π₀ and vanilla MoE converged under the reported training schedules? Please provide success-rate curves against training steps and cumulative compute, including longer-trained baselines, to clarify whether AdaMoE retains its advantage as the training budget increases.
  2. How consistent are the improvements over vanilla MoE across independent training seeds? Please report uncertainty for the controlled comparisons, particularly the 1.2-point LIBERO and 3.8-point RoboTwin gains.
  3. Please precisely define the combined gradient and reference task gradient in Appendix H, including how router and adapter parameter spaces are combined. What evidence distinguishes reduced harmful interference from a smaller aggregate angle caused by adding task-supervised parameters?

Comments on the paper

2 comments

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

AuthorsOpenReview ↗

Closing Summary for Area Chair and Reviewers

Dear Area Chair and Reviewers,

As the discussion period concludes, we sincerely thank all reviewers for their constructive feedback, and especially thank Reviewer jh4G for engaging in the discussion, confirming that our new experiments resolved the key empirical concerns, and leaning towards acceptance. Since Reviewers jPcU, HRD8, and wbT4 were unable to respond before the deadline, we briefly summarize how our revised manuscript (Appendices H, P, and Q, in blue) has addressed every point raised across all four reviews:

  • Uncontested Core Design & 100% Experimental Fulfillment: None of the reviewers questioned the motivation, problem formulation, or sparse efficiency of AdaMoE-VLA, and we have completed every additional experiment requested across all four reviews.
  • CSMoE & Cross-Backbone Generalization (Appendix Q — jh4G, wbT4, HRD8): Evaluations on RoboTwin 2.0 across both π0\pi_0 and π0.5\pi_ {0.5} establish Base<Vanilla MoE<CSMoE<AdaMoE-VLA\mathrm{Base} < \mathrm{Vanilla\ MoE} < \mathrm{CSMoE} < \mathrm{AdaMoE\text{-}VLA}, isolating decoupling gains from adapter capacity (confirmed resolved by jh4G).
  • 300-Rollout Real-World Control (Appendix P — jh4G, HRD8, jPcU): Across 100 trials/model on the 3-stage Organize Table task (π0.5\pi_ {0.5} base), AdaMoE-VLA reaches 83.0% full-task success vs. 74.0% (Vanilla MoE) and 67.0% (π0.5\pi_ {0.5}), with stage-wise gains widening monotonically (confirmed resolved by jh4G).
  • Multi-Seed Robustness & Convergence (Appendix Q — jPcU, HRD8): 3-seed runs on LIBERO and RoboTwin 2.0 confirm consistent gains and >4×>4\times lower variance on LIBERO-Long, while 120k-step checkpoints verify that baselines have fully plateaued by 90k steps.
  • Exact Gradient Formulation (Appendix H — all reviewers): Clarified our exact gradient-angle definition in the shared [B,S,K][B, S, K] gating output space without parameter concatenation (acknowledged as insightful by jh4G).

We hope this self-contained summary and the revised PDF assist the Area Chair and reviewers during final deliberations. Thank you again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.