Technical Summary: AIR-VLA+

1. Problem Statement

Aerial manipulation systems, which combine Unmanned Aerial Vehicles (UAVs) with robotic arms, face a fundamental challenge in end-to-end control: representation coupling. The control requirements for the two subsystems are highly heterogeneous:

• UAV Movement: Involves macroscopic, large-scale relative positional changes (3D translation and yaw) with smooth, continuous dynamics.
• Arm Manipulation: Involves microscopic, local 3D high-precision alignment (joint angles and gripper states) with drastic, non-linear variations.

Existing Vision-Language-Action (VLA) models, particularly those based on flow matching, typically optimize these two distinct action spaces within a single shared latent vector field. This joint optimization triggers severe coordination conflicts, manifesting physically as UAV drifting, jittering, and uncoordinated transitions between task stages (e.g., hovering vs. moving). The paper argues that forcing these topologically conflicting action features into a shared manifold degrades performance and stability.

2. Methodology: AIR-VLA+

The authors propose AIR-VLA+, a flow matching action generation architecture designed to achieve "cognitively highly coordinated and physically absolutely decoupled" control. The architecture consists of five core components:

A. Cascaded Dual-Action Decoders

Instead of a single shared decoder, AIR-VLA+ employs two specialized, asymmetric decoders:

1. Arm Manipulation Decoder: Directly predicts high-precision 3D alignment intent from shared latent features.
2. UAV Movement Decoder: Predicts macroscopic movement intent.

• Unidirectional Information Cascade: A critical design feature is the use of a stop-gradient (.detach()) operation on the manipulator's output. This allows the UAV decoder to unidirectionally observe the arm's intent for workflow coordination (e.g., knowing when to hover) while strictly isolating the UAV's movement backpropagation from the arm's manifold. This protects the pure inverse kinematics mapping of the manipulator.

B. Input Feature Enhancement Module for UAV

Recognizing that UAV movement relies heavily on high-level semantics and task state transitions, the UAV decoder receives an enhanced input vector constructed by concatenating:

1. Shared Foundational Visual Features: From the shared VLA backbone.
2. Compressed Global Semantic Features: A global intent vector derived via mean pooling of prefix tokens (images and text), broadcasted across the action horizon to maintain task awareness.
3. Implicit Visual Grasp Features: Extracted via a miniature implicit perception network (2-layer MLP) that mines physical contact representations (gripper-object state) from shared visual features.
4. Isolated Manipulator Intent: The detached output from the arm decoder.

C. Asymmetric Feature-Level Mixture of Experts (MoE)

Within the UAV Movement Decoder, the authors deploy a Feature-Level MoE rather than an action-level MoE.

• Routing Mechanism: A router network computes activation weights based on global semantics and grasp states.
• Dense Soft Blending: Multiple expert networks ($E_1, \dots, E_K$) process the enhanced input independently. Their outputs are combined via dense soft blending on the feature manifold.
• Benefit: This allows different "movement experts" to spontaneously exhibit capacity inclinations for specific task stages (e.g., approaching, grasping, moving) without disrupting the continuous flow field required for ODE integration.

D. Action Space Hard Stitching

After the decoders generate their respective derivative fields ($\hat{v}_{move}$ and $\hat{v}_{manip}$), the system performs physical hard stitching before the ODE solver.

• The first 4 dimensions (UAV motion) are taken strictly from the Movement MoE.
• The remaining dimensions (arm joints and gripper) are taken strictly from the Arm Decoder.
• This "hard stitch" ensures that the pure kinematic mapping of the arm is preserved, avoiding the trajectory discontinuities and solver collapse seen in action-level soft or hard MoE attempts.

3. Key Contributions

1. Architecture Design: The proposal of AIR-VLA+, an asymmetric feature-level MoE flow matching architecture with cascaded dual-action decoders, specifically tailored to the heterogeneous characteristics of aerial manipulation.
2. Feature Enhancement & Decoupling: The design of an input feature enhancement module that integrates implicit visual grasp representations and global semantics, coupled with a unidirectional information cascade. This resolves cross-modal feature submergence and protects arm stability from UAV movement backpropagation.
3. State-of-the-Art Performance: Comprehensive evaluation on the standardized AIR-VLA benchmark demonstrating that the method outperforms all baselines, including single-head $\pi_{0.5}$ and other MoE variants.

4. Experimental Results

Experiments were conducted in Isaac Sim using a quadrotor with a 7-DoF manipulator across four task suites: Base Manipulation, Object & Spatial, Semantic Understanding, and Long-Horizon.

• Overall Performance: AIR-VLA+ achieved an overall average score of 48.0, surpassing all baselines.
• Improvement over Baseline: Compared to the single-head $\pi_{0.5}$ policy, the method improved the task completion score by 80.2%.
• Specific Gains:
  • Base Manipulation: Success rate leap of 200%.
  • Long-Horizon Tasks: Success rate leap of 150%.
  • Safety & Stability: The method effectively mitigated UAV jittering and drifting, achieving higher positioning accuracy (70.0 vs. 78.8 for base, but with significantly higher task completion and safety scores in complex scenarios) and safer interactions.
• Ablation Studies:
  • Removing the stop-gradient operation caused severe manipulator jitter and reduced efficacy.
  • Omitting the visual grasp projector led to "semantic amnesia" and a cognitively blind state.
  • Expert Count: The optimal number of experts was found to be 3 ($N=3$), corresponding to the latent spaces of approaching, grasping, and moving to target. $N=2$ failed to partition transitions smoothly, while $N=4$ introduced redundancy without gains.
• Interpretability: Analysis of expert routing weights confirmed that the three implicit experts spontaneously specialized for specific task stages (e.g., Expert 1 for grasping, Expert 2 for approaching, Expert 3 for moving), validating the emergence of functional division of labor.

5. Significance and Claims

The paper claims that AIR-VLA+ addresses the critical bottleneck of representation coupling in composite robots. By decoupling the physical control of the chassis and the manipulator while maintaining cognitive coordination through a unidirectional cascade, the architecture enables:

• Stable Flight: Mitigation of UAV jittering and drifting during complex interactions.
• Precise Manipulation: Preservation of the arm's inverse kinematics integrity.
• Task Adaptability: Enhanced ability to transition between task stages (hovering, moving, grasping) via the MoE's soft blending on the feature manifold.

The authors position this work as a necessary architectural evolution for applying VLA models to aerial manipulation, moving beyond simple foundation model transfers to deep structural adaptations that respect the heterogeneous dynamics of the system. The method establishes a new state-of-the-art on the AIR-VLA benchmark, offering a novel multi-expert routing paradigm for heterogeneous collaborative control.