- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper builds on 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 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 , with task progression increasing from 10.6 to 19.1. The ablation results favor over independent, bidirectional, and 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
-
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, , and coordination using the same backbone architecture and semantic/grasp feature modules. 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.
-
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 , 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 , 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
-
The core comparisons lack sufficient experimental control and statistical evidence. The results for ACT, DP, -FAST, , and 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 or 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 (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.
-
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, , is simply a linear readout of , 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
-
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?
-
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 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?
-
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 and variants (d)–(h)? To substantiate the expert-count conclusion, could you also provide these statistics for and ? 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?
-
Architectural differences and mechanism evidence. What are the exact architectures of the UAV heads in variants (b)–(g), the experts , and the shared output projection ? Do the Action-Level Soft/Hard MoE baselines use the same features, dual-branch structure, cascade, expert capacity, router, and training configuration? If is linear or affine, then for identical expert features and normalized routing weights , the forward computation satisfies . 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 , where 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?
-
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?