acceptodds
Withdrawn from ICLR 2027

AIR-VLA+: Decoupling Movement and Manipulation via Cascaded Dual-Action Decoders with Asymmetric MoE for Aerial Robots

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Aerial manipulation systems have long suffered from representation coupling in end-to-end control, as platform-level Unmanned Aerial Vehicle (UAV) movement and end-effector-level arm manipulation differ substantially in action scale, dynamics, and control objectives. In this paper, we propose AIR-VLA+, a flow matching action generation architecture specifically designed for aerial manipulation, featuring cascaded dual-action decoders and an asymmetric feature-level Mixture of Experts (MoE). We construct cascaded manipulation and movement decoders, allowing the UAV to unidirectionally observe the manipulator's intent during movement to achieve workflow coordination, while isolating the impact of UAV movement information backpropagation on arm manipulation stability. Addressing the characteristic that UAV movement is highly dependent on high-level semantics and responsible for task state transitions in aerial manipulation, we design an input feature enhancement module for the UAV movement decoder. This module introduces an implicit visual grasp projector to perceive the interaction state between the gripper and the object, and injects compressed global semantic features. Within the UAV movement decoder, we deploy an implicit MoE architecture, enabling different movement experts to spontaneously exhibit capacity inclinations for various task stages during training. Through dense soft blending computation on the feature manifold, the UAV movement is endowed with stronger task-stage adaptability. Experiments on the standardized AIR-VLA benchmark demonstrate that our method comprehensively surpasses all baselines with an overall average score of 48.0. The overall task completion score improves by 80.2% compared to the single-head policy, effectively mitigating the heterogeneous coordinated control conflicts of composite robots.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Aerial Manipulation, Vision-Language-Action Models

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_gL3rOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper builds on π0.5\pi_{0.5} and proposes AIR-VLA+ for aerial manipulation with a quadrotor equipped with a 7-DoF manipulator. The authors argue that jointly predicting UAV and manipulator actions through a single output projection can introduce optimization interference, manifesting as drift, jitter, and poorly coordinated task-stage transitions. The method retains a shared backbone but splits the output into two branches: a linear manipulator head and a separate UAV decoding branch. In addition to the shared features, the UAV branch receives pooled vision–language features, implicit grasp features extracted by a small MLP, and the manipulator’s predicted vector field with stop-gradient applied, forming a unidirectional Arm→Move\mathrm{Arm}\rightarrow\mathrm{Move} cascade. Within the UAV branch, a feature-level MoE softly combines intermediate features from three linear experts, followed by a shared projection that produces the UAV component of the vector field. On the simulated AIR-VLA benchmark, the method reports an overall weighted score of 48.0 versus 42.0 for single-head π0.5\pi_{0.5}, with task progression increasing from 10.6 to 19.1. The ablation results favor Arm→Move\mathrm{Arm}\rightarrow\mathrm{Move} over independent, bidirectional, and Move→Arm\mathrm{Move}\rightarrow\mathrm{Arm} coordination in overall score, and three experts outperform two or four in the tested configurations. The appendix further analyzes routing dynamics, forced single-expert routing, and linear probing.

Critical strengths
  1. A useful design question and a focused comparison of coordination directions. The paper examines which way information should flow between the mobile-base and arm branches—a question directly relevant to aerial manipulation and worth investigating on other mobile manipulators. Table 2 (d)–(g) provides a focused comparison of independent, bidirectional, Move→Arm\mathrm{Move}\rightarrow\mathrm{Arm}, and Arm→Move\mathrm{Arm}\rightarrow\mathrm{Move} coordination using the same backbone architecture and semantic/grasp feature modules. Arm→Move\mathrm{Arm}\rightarrow\mathrm{Move} achieves an overall score of 45.1, compared with 37.2, 40.4, and 40.5 for the other three configurations. I consider this the most valuable part of the paper. If this advantage persists under matched training budgets and multiple independent training seeds, it would strengthen the case for the design and motivate testing it on other floating-base and mobile manipulation platforms.

  2. Focused architectural changes and joint improvements in safety and task progression. The architectural modifications are concentrated in action generation and are easy to follow. The method retains the shared backbone architecture of π0.5\pi_{0.5}, adds action projections, feature enhancement, and a router, and continues to use the standard flow-matching objective. Appendix B reports the two-stage training setup. Some reported gains are substantial: on Base Manipulation, the total score increases from 45.3 to 62.1, and task progression from 11.0 to 33.0. Compared with single-head π0.5\pi_{0.5}, safety scores improve across all four suites, increasing from 84.7 to 94.7 overall, while overall task progression rises from 10.6 to 19.1. The joint improvement is more informative than a higher safety score alone, since ACT and DP in the same table achieve a safety score of 100 but a task-progression score of 0. These results provide the main empirical support for the paper; the weaknesses below concern how reliably the gains can be attributed to the proposed architecture.

Critical weaknesses
  1. The core comparisons lack sufficient experimental control and statistical evidence. The results for ACT, DP, π0\pi_0-FAST, π0\pi_0, and π0.5\pi_{0.5} in Table 1, as well as row (a) in Table 2, exactly match those in Table 1 of the AIR-VLA benchmark paper. However, this paper does not clarify whether these results were directly taken from that work or obtained through retraining and reevaluation. The original benchmark used default hyperparameters and 10 evaluation trials per task, whereas Appendix B of this paper reports a two-stage training procedure of 30k + 10k steps and 20 evaluation trials per task. It remains unclear whether the baselines and ablation variants were compared under matched training budgets, tuning procedures, and evaluation protocols. Consequently, the improvement from 42.0 to 48.0 cannot yet be confidently attributed to the proposed architecture. The paper also reports neither the number of training seeds nor standard deviations or confidence intervals, making it difficult to assess the stability of these differences. With N=2N=2 or N=4N=4 experts, the overall scores are 39.9 and 43.0, respectively, both below variant (g) without MoE (45.1). This indicates sensitivity to the expert configuration and calls for further evidence of robustness. Furthermore, the paper broadly claims improved UAV positioning, yet its overall positioning score is lower than that of π0.5\pi_{0.5} (70.0 versus 71.3), with lower scores in three of the four task suites. Since the paper’s contribution is primarily empirical, these limitations in experimental control and statistical evidence are the main reasons for my rating.

  2. The proposed explanations for the performance gains are not sufficiently supported. According to Equations (3) and (7) and Appendix B, the manipulation intent passed to the UAV branch, sg⁡(Pmanip(fsuffix))\operatorname{sg}(P_{\mathrm{manip}}(f_{\mathrm{suffix}})), is simply a linear readout of fsuffixf_{\mathrm{suffix}}, which the UAV branch already receives. Under the stated linear-expert implementation, this term can be algebraically absorbed into the expert weights for fixed parameters; the router also does not receive this intent. Thus, the cascade neither adds conditional information nor expands the UAV branch’s forward function class. It may still improve training through its parameterization and gradient routing, but the paper lacks comparisons with and without stop-gradient and capacity-matched reparameterization controls. The current evidence therefore does not distinguish such optimization effects from the claimed mechanism of “observing manipulation intent.” Evidence for task-stage specialization among experts is also limited: the routing weights in Figure 3 vary within a relatively narrow range, while restricting the model to Expert 1 at test time reduces the score only from 48.0 to 47.6 in Table 4. These results do not sufficiently establish an important role for dynamic routing in stage coordination. Moreover, the confusion matrix in Figure 4 gives a linear-probe accuracy of 78.8% for Semantic Understanding, below the 82.8% accuracy obtained by always predicting the majority class. This result does not support the claimed robustness of grasp-interaction perception. Overall, the analyses do not sufficiently establish that the performance gains arise from the mechanisms emphasized in the paper.

Questions for authors
  1. Numerical consistency. Could you reconcile the discrepancies between the reported total scores and the stated weights? The component scores give approximately 46.4 for Ours on Long-Horizon, rather than 44.1, and approximately 18.8 for variant (c) on Base Manipulation, rather than 10.1. How is the overall manipulation score of 53.7 aggregated, given that the unweighted mean across the four suites is approximately 55.8? Which entries are correct, and would corrections change the overall scores or method rankings?

  2. Baseline provenance and training control. Were the baseline rows in Table 1 and row (a) in Table 2 independently rerun, or taken from Sun et al. (AIR-VLA, arXiv:2601.21602)? What results does single-head π0.5\pi_{0.5} achieve with the same data, batch size, total optimization steps, and 20 evaluation trials per task as AIR-VLA+? Are the learning-rate schedule and freezing strategy for corresponding modules matched, including an equivalent second-stage training budget?

  3. Statistical reliability. How many independent training seeds were used for each variant, as distinct from evaluation rollouts? Could you report the mean and standard deviation over at least three independent training seeds for single-head π0.5\pi_{0.5} and variants (d)–(h)? To substantiate the expert-count conclusion, could you also provide these statistics for N=2N=2 and N=4N=4? Does their underperformance relative to the non-MoE variant (g) persist across seeds, and how does this affect the conclusion that three experts are optimal?

  4. Architectural differences and mechanism evidence. What are the exact architectures of the UAV heads in variants (b)–(g), the experts EiE_i, and the shared output projection PmoveP_{\mathrm{move}}? Do the Action-Level Soft/Hard MoE baselines use the same features, dual-branch structure, cascade, expert capacity, router, and training configuration? If PmoveP_{\mathrm{move}} is linear or affine, then for identical expert features hih_i and normalized routing weights wiw_i, the forward computation satisfies Pmove(∑iwihi)=∑iwiPmove(hi)P_{\mathrm{move}}(\sum_i w_i h_i)=\sum_i w_i P_{\mathrm{move}}(h_i). Which non-equivalent implementation or training differences explain the gap between 48.0 and 31.1? Could you provide a comparison of variant (g) with and without stop-gradient, and a control replacing the manipulation-intent input with sg⁡(Rfsuffix)\operatorname{sg}(R f_{\mathrm{suffix}}), where RR is a fixed random linear projection with output dimension and scale matched to the original input? How is the bidirectional variant (e) computed, including its prediction/update order, stop-gradient locations, and additional computation?

  5. Second-stage training and attribution. Is the additional 10k-step stage applied only to the full model, or also to the baselines and ablation variants? What initialization checkpoints and frozen/trainable modules are used for each variant? In particular, does (g) receive the same second-stage training as (h), and does the gain persist under matched training and freezing strategies? Since Appendix B states that this stage also updates the manipulation projection, what controlled evidence distinguishes improved UAV coordination from further manipulation-head optimization as the explanation for the manipulation-score increase from 50.4 to 53.7?

Reviewer_ywz7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AIR-VLA+, which is a flow-matching VLA model built on pi0.5, designed for aerial manipulation tasks. To mitigate interference arising from the heterogeneity between UAV actions and robotic arm actions, AIR-VLA+ replaces the shared action head with two cascaded decoders which are a linear arm head and a UAV decoder. The UAV decoder is conditioned on global semantics, grasp-state features, and the detached arm prediction. It then incorporates a feature-level soft MoE with three experts. On simulated benchmarks, the proposed AIR-VLA+ outperforms pi0.5 (48.0 vs. 42.0).

Critical strengths
  • This paper addresses a well-motivated problem. The model design is lightweight and could be easily plug into other VLA backbones.
  • The proposed approach demonstrate consistent gains over base model on both safety and manipulation.
Critical weaknesses
  • The scope of the evaluation is quite narrow as all experiments were conducted on a single simulated benchmark. It would be better to include real-world or sim-to-real experiments.
  • To properly attribute the improvement of AIR-VLA+, the base model pi0.5 should be finetuned with same training data, budget and evaluation protocol.
Questions for authors

Please refer to the weaknesses.

Reviewer_9WsaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AIR-VLA+ addresses heterogeneous coordination between UAV movement and arm manipulation using a shared VLA backbone and separate action branches. A stop-gradient Arm→Move connection allows UAV movement to condition on manipulation intent. Global semantic features, implicit grasp-state features, and a feature-level MoE further support stage-adaptive movement generation. Evaluation is conducted on the AIR-VLA benchmark in Isaac Sim.

Critical strengths

1.The problem is practically relevant, and asymmetric coordination is well motivated.** UAV movement must account for manipulation state when deciding whether to approach, hover, or transition, making the Arm→Move design consistent with the intended workflow.

2.The architecture balances action specialization and coordination.** Separate output branches retain shared observations and manipulation-intent communication, providing a clear design that can be integrated into existing flow-matching VLAs.

Critical weaknesses

1.The additional benefit of dynamic MoE blending and the claimed stage specialization remain insufficiently established.** The full model scores 48.0, while fixing the routing to Expert 1 at evaluation still achieves 47.6; the task-progression score similarly decreases only from 19.1 to 18.8. This suggests that dynamic blending may provide limited additional benefit over a fixed expert. Although stage-dependent routing visualizations are informative, they do not establish stable expert responsibilities corresponding to approaching, grasping, and transporting.

2.The architecture may incorporate a task-specific workflow bias.** Although experts are not explicitly assigned stages, the design and its interpretation remain closely tied to a staged approach–grasp–transport workflow. Its advantages may therefore be particularly suited to the evaluated task structure. Whether this specialization remains beneficial for more complex tasks with less distinct stages or sustained fine-grained interaction is unclear, limiting conclusions about architectural generality.

3.Real-world performance is difficult to assess without physical experiments.** All results are obtained in simulation, leaving effectiveness and robustness under real flight dynamics, perception noise, execution delays, and physical contact unverified. Representative hardware experiments, or a clearer qualification of the claims, would strengthen the paper.

4.The main paper lacks a Discussion and Limitations section.** A dedicated discussion of applicability, failure modes, architectural assumptions, and sim-to-real limitations would help readers understand the method’s scope.

Questions for authors

1.How do the authors explain the near-equivalent performance of fixed Expert 1 and dynamic blending? Can they further establish the independent benefit of dynamic routing and provide more direct evidence of stage specialization?

2.Can the current MoE architecture accommodate tasks with less structured workflows? If broader applicability is intended, additional experiments on tasks departing from the three-stage approach–grasp–transport workflow would help substantiate this claim. Otherwise, please explicitly address in Discussion and Limitations the architecture’s potential dependence on this particular workflow and its unverified applicability beyond it.

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

Related papers

Discussion (0)

Sign in to comment.