Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation

Jianli Sun, Bin Tian, Qiyao Zhang, Chengxiang Li, Zihan Song, Zhiyong Cui, Yisheng Lv, Yonglin Tian

TL;DR

AIR-VLA addresses the gap of applying Vision-Language-Action (VLA) models to Aerial Manipulation Systems by introducing the first AMS-focused VLA benchmark. It couples a physics-based NVIDIA Isaac Sim environment with a multimodal dataset of 3000 teleoperated aerial manipulation demonstrations and four task suites that stress 3D navigation, manipulation, semantic understanding, and long-horizon planning. The paper provides a two-layer evaluation framework for both VLA and VLM models, extensive baseline analyses, and insights into the transferability of ground-based VLA paradigms to high-DoF aerial platforms, highlighting both potential and key bottlenecks such as floating-base dynamics and spatial grounding. The AIR-VLA benchmark offers a standardized testbed and dataset to drive future research toward robust, safe, and generalizable aerial embodied intelligence with practical implications for logistics, disaster response, and inspection tasks.

Abstract

While Vision-Language-Action (VLA) models have achieved remarkable success in ground-based embodied intelligence, their application to Aerial Manipulation Systems (AMS) remains a largely unexplored frontier. The inherent characteristics of AMS, including floating-base dynamics, strong coupling between the UAV and the manipulator, and the multi-step, long-horizon nature of operational tasks, pose severe challenges to existing VLA paradigms designed for static or 2D mobile bases. To bridge this gap, we propose AIR-VLA, the first VLA benchmark specifically tailored for aerial manipulation. We construct a physics-based simulation environment and release a high-quality multimodal dataset comprising 3000 manually teleoperated demonstrations, covering base manipulation, object & spatial understanding, semantic reasoning, and long-horizon planning. Leveraging this platform, we systematically evaluate mainstream VLA models and state-of-the-art VLM models. Our experiments not only validate the feasibility of transferring VLA paradigms to aerial systems but also, through multi-dimensional metrics tailored to aerial tasks, reveal the capabilities and boundaries of current models regarding UAV mobility, manipulator control, and high-level planning. AIR-VLA establishes a standardized testbed and data foundation for future research in general-purpose aerial robotics. The resource of AIR-VLA will be available at https://anonymous.4open.science/r/AIR-VLA-dataset-B5CC/.

AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation

TL;DR

AIR-VLA addresses the gap of applying Vision-Language-Action (VLA) models to Aerial Manipulation Systems by introducing the first AMS-focused VLA benchmark. It couples a physics-based NVIDIA Isaac Sim environment with a multimodal dataset of 3000 teleoperated aerial manipulation demonstrations and four task suites that stress 3D navigation, manipulation, semantic understanding, and long-horizon planning. The paper provides a two-layer evaluation framework for both VLA and VLM models, extensive baseline analyses, and insights into the transferability of ground-based VLA paradigms to high-DoF aerial platforms, highlighting both potential and key bottlenecks such as floating-base dynamics and spatial grounding. The AIR-VLA benchmark offers a standardized testbed and dataset to drive future research toward robust, safe, and generalizable aerial embodied intelligence with practical implications for logistics, disaster response, and inspection tasks.

Abstract

While Vision-Language-Action (VLA) models have achieved remarkable success in ground-based embodied intelligence, their application to Aerial Manipulation Systems (AMS) remains a largely unexplored frontier. The inherent characteristics of AMS, including floating-base dynamics, strong coupling between the UAV and the manipulator, and the multi-step, long-horizon nature of operational tasks, pose severe challenges to existing VLA paradigms designed for static or 2D mobile bases. To bridge this gap, we propose AIR-VLA, the first VLA benchmark specifically tailored for aerial manipulation. We construct a physics-based simulation environment and release a high-quality multimodal dataset comprising 3000 manually teleoperated demonstrations, covering base manipulation, object & spatial understanding, semantic reasoning, and long-horizon planning. Leveraging this platform, we systematically evaluate mainstream VLA models and state-of-the-art VLM models. Our experiments not only validate the feasibility of transferring VLA paradigms to aerial systems but also, through multi-dimensional metrics tailored to aerial tasks, reveal the capabilities and boundaries of current models regarding UAV mobility, manipulator control, and high-level planning. AIR-VLA establishes a standardized testbed and data foundation for future research in general-purpose aerial robotics. The resource of AIR-VLA will be available at https://anonymous.4open.science/r/AIR-VLA-dataset-B5CC/.
Paper Structure (54 sections, 12 equations, 5 figures, 4 tables)

This paper contains 54 sections, 12 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Aerial Manipulation Systems
  4. Vision-Language-Action Models
  5. VLA for Aerial Robotics
  6. Benchmarks for Robot Manipulation
  7. AIR-VLA
  8. Task Design Taxonomy
  9. Evaluation Framework
  10. Simulation Environment
  11. Dataset Construction
  12. Data Collection Strategy.
  13. Multimodal Perception System.
  14. Action and State Space.
  15. Scene and Instruction Diversity.
  16. ...and 39 more sections

Figures (5)

  • Figure 1: Overview of AIR-VLA model framework and dataset of aerial manipulation tasks.
  • Figure 2: Overview of the AIR-VLA benchmark. AIR-VLA serves as a full-stack Vision-Language-Action platform tailored for aerial manipulation systems. It integrates a simulation-based teleoperation data acquisition pipeline, an online simulation environment, and diverse multimodal datasets. Furthermore, it provides a comprehensive benchmark for evaluating mainstream VLA and VLM models across varied aerial manipulation tasks.
  • Figure 3: Evaluation pipeline for VLM high-level planning capabilities in aerial manipulation tasks.
  • Figure 4: Task examples in AIR-VLA dataset.
  • Figure 5: Observation Space. This figure illustrates instances of observations from various camera perspectives.