APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal Navigation

TL;DR

APEX tackles the challenging problem of Aerial Object Navigation by decoupling perception, decision-making, and target grounding into three specialized modules. It introduces a Dynamic Spatio-Semantic Mapping framework that builds a 3D grid memory with Attraction, Exploration, and Obstacle maps, coupled with an RL-based Action Decision Module using PPO and a Target Grounding Module employing an open-vocabulary detector. The system operates asynchronously to mitigate VLM latency, achieving state-of-the-art results on UAV-ON with improvements in SR and SPL, while maintaining high safety via the Obstacle Map and continuous Target Grounding. Ablation studies validate the necessity of each component, and offline map generation with precomputed attraction maps enables efficient RL training. The work demonstrates that a hierarchical, memory-rich, and parallel architecture yields practical, reliable, and efficient aerial navigation in open-world scenarios, with potential for further VLM pretraining and alternative RL strategies.

Abstract

Aerial Object Goal Navigation, a challenging frontier in Embodied AI, requires an Unmanned Aerial Vehicle (UAV) agent to autonomously explore, reason, and identify a specific target using only visual perception and language description. However, existing methods struggle with the memorization of complex spatial representations in aerial environments, reliable and interpretable action decision-making, and inefficient exploration and information gathering. To address these challenges, we introduce \textbf{APEX} (Aerial Parallel Explorer), a novel hierarchical agent designed for efficient exploration and target acquisition in complex aerial settings. APEX is built upon a modular, three-part architecture: 1) Dynamic Spatio-Semantic Mapping Memory, which leverages the zero-shot capability of a Vision-Language Model (VLM) to dynamically construct high-resolution 3D Attraction, Exploration, and Obstacle maps, serving as an interpretable memory mechanism. 2) Action Decision Module, trained with reinforcement learning, which translates this rich spatial understanding into a fine-grained and robust control policy. 3) Target Grounding Module, which employs an open-vocabulary detector to achieve definitive and generalizable target identification. All these components are integrated into a hierarchical, asynchronous, and parallel framework, effectively bypassing the VLM's inference latency and boosting the agent's proactivity in exploration. Extensive experiments show that APEX outperforms the previous state of the art by +4.2\% SR and +2.8\% SPL on challenging UAV-ON benchmarks, demonstrating its superior efficiency and the effectiveness of its hierarchical asynchronous design. Our source code is provided in \href{https://github.com/4amGodvzx/apex}{GitHub}

Figures (8)

• Figure 1: APEX addresses three critical limitations in prior works. 1) Memory: APEX employs a rich 3D dynamic map for robust spatial reasoning. 2) Decision-Making: APEX uses decoupled modules, separating semantic understanding from action generation for more reliable and interpretable control. 3) Efficiency: APEX employs a parallel and asynchronous framework that mitigates inference latency and boosts exploration proactivity.
• Figure 2: The overall framework of APEX. Module 1 extracts and integrates visual information into 3D dynamic spatio-semantic maps, serving as the agent's memory. Module 2 utilizes the stored information from Module 1 for action decision-making via an RL-based model. Module 3 continuously performs object detection on visual observations to obtain precise target locations. The three modules are integrated into a hierarchical asynchronous parallel system.
• Figure 3: Detailed architecture of the Dynamic Map Module and the Asynchronous Parallel Framework.
• Figure 4: A case study of the APEX agent navigating to the target "Umbrella". The figure shows the dynamic integration of the Attraction, Exploration, and Obstacle Maps in the mission, leading to successful navigation. More visualized results are demonstrated in Appendix B.
• Figure 5: The architecture of policy network.