ActiveFly-Bench: Aligning Embodied Question Answering with Vision-Language-Action for Aerial Embodied Perception

2026-07-11

AI Summary

This work addresses the challenge that existing drone agents struggle to cohesively integrate task understanding, behavioral planning, and low-level control in active perception, compounded by the absence of standardized evaluation benchmarks. To bridge this gap, the authors introduce ActiveFly-Bench—the first benchmark specifically designed for embodied aerial perception—which decomposes active perception into three hierarchical components: embodied question answering, observation-driven behavior planning, and fine-grained language-guided control. They further develop a closed-loop agent, ActiveFly, to validate the benchmark. ActiveFly-Bench establishes the first evaluation framework linking cyberspace reasoning with physical-world interaction, featuring a hybrid real-simulated dataset and integrating vision-language models (VLMs) with vision-language-action models (VLAs) for end-to-end control. Experiments reveal significant limitations in current approaches regarding behavior planning, viewpoint adjustment, and task robustness, underscoring the benchmark’s effectiveness and necessity as a new testbed for aerial embodied intelligence.

Abstract

We introduce ActiveFly-Bench, the first benchmark to bridge cyberspace reasoning and physical-world interaction for UAV embodied perception. The benchmark decomposes active perception into three hierarchical tasks: Aerial Embodied Question Answering (Air-EQA), Observation Behavior Planning (OBP), and Fine-grained Language-guided UAV Control (FLUC), explicitly connecting high-level task understanding, behavior planning, and low-level control. The datasets are collected from both real-world and simulated outdoor environments for training and evaluation. We further develop ActiveFly, a closed-loop UAV agent that integrates visual-language reasoning with fine-grained control, and deploy it on a physical UAV platform. Experiments with representative VLMs and VLA models show that current UAV agents still struggle with behavior planning, viewpoint adjustment, and robust task completion in active perception. These results establish ActiveFly-Bench as a new testbed for embodied aerial intelligence.

Problem

• Embodied Perception
• Aerial Intelligence
• Behavior Planning
• Vision-Language-Action
• UAV Control

Innovation

• embodied perception
• vision-language-action
• hierarchical task decomposition
• UAV control
• active perception benchmark