PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Daoxuan Zhang, Ping Chen, Xiaobo Xia, Xiu Su, Ruichen Zhen, Jianqiang Xiao, Shuo Yang

arXiv:2602.00551v1cs.ROcs.CV

TL;DR

Aerial ObjectNav demands autonomous search in complex 3D settings, where existing methods face spatial-memory, control, and latency challenges. APEX combines specialized mapping, action-decision, and target-grounding modules in an asynchronous framework; on UAV-ON, it reports state-of-the-art performance with improved SR and SPL.

  • Problem

    Aerial navigation methods face challenges in retaining complex spatial information, producing reliable and interpretable actions, and operating efficiently under VLM inference latency.

    Source: §1 · pp. 2–3

  • Method

    APEX combines dynamic 3D spatio-semantic mapping, an RL-based action policy, and target grounding in a hierarchical asynchronous parallel architecture.

    Source: §1 · p. 3, §3.3 · p. 4, §3.4 · p. 5

  • Results

    4.16% improvement in Success Rate (SR) and 2.77% improvement in Success-weighted Path Length (SPL) on the UAV-ON benchmark, with the highest Oracle Success Rate (OSR).

    Source: §4.2 · p. 8

  • Takeaways & Limitations

    On the UAV-ON Benchmark, APEX achieves state-of-the-art performance and exhibits superior reliability and higher operational efficiency.

    Source: §5 · p. 9

    A failure case involved collision with a tree, suggesting the Obstacle Map may have difficulty detecting sparse obstacles.

    Source: §7.1 · p. 12

Abstract

from arXiv · show

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}

1. Introduction

Aerial navigation extends language-guided navigation into complex 3D spaces, where autonomous search must contend with instruction dependence, spatial-memory demands, unreliable control, and VLM latency. APEX addresses these challenges with specialized mapping, action-decision, and target-grounding modules in an asynchronous parallel framework.

1. Introduction

  • Fine-grained, step-by-step language instructions constrain autonomy and limit the scalability and applicability of existing aerial navigation systems in open-ended exploration.Aerial ObjectNav instead asks a UAV to find an object from a high-level textual cue using onboard visual sensors.Source: §1 · p. 2
  • Existing approaches often fail to integrate historical observations with spatial information, while topological maps can sacrifice geometric and metric fidelity needed for fine-grained reasoning and planning.These limitations are associated with inefficient exploration, repeated errors, and impaired reasoning in complex aerial terrains.Source: §1 · p. 2
  • End-to-end VLM decision-making leaves a gap between semantic reasoning and low-level control, and can require extensive demonstrations while producing unstable, unreliable, and less interpretable policies.Large-VLM inference also creates latency that conflicts with the continuous motion needed for practical UAV operation.Source: §1 · pp. 2–3
  • APEX decouples aerial object navigation into dynamic spatio-semantic mapping, RL-based action decision, and target-grounding modules within an asynchronous parallel framework.Its contributions emphasize long-horizon spatial memory, reliability and interpretability, and exploration efficiency.Source: Figure 2 · p. 2, §1 · p. 3

2. Related Work

Related work traces Object Goal Navigation from modular pipelines toward learned policies and multimodal reasoning, while aerial navigation raises additional demands for memory and long-horizon efficiency. APEX is positioned as a response to instruction dependence and loss of spatial detail in existing aerial methods.

2.1. Object Goal Navigation

  • Object Goal Navigation asks an agent to find a target in an unseen environment; methods evolved from classical modular pipelines to IL/RL policies and multimodal approaches such as NavGPT-2 and VoroNav.Aerial ObjectNav remains an open problem for efficient, reliable long-horizon tasks.Source: §2.1 · p. 3

2.2. Aerial Vision-Language Navigation

  • AVLN methods may depend on detailed language instructions and discard geometric detail in topological maps, while APEX addresses these limits with simple goal descriptions and dynamic 3D maps.Prior work includes aerial simulators, end-to-end agents, topological memories, VLM training, and point-cloud geometry methods.Source: §2.2 · pp. 3–4

3. Methodology

APEX formulates aerial object navigation as memory-augmented decision-making and separates mapping, action selection, and target confirmation. Its 3D maps guide PPO control, while asynchronous modules operate at different frequencies.

3.1. Task Formulation

  • The agent maps observations, its state, and a goal description to actions, with persistent spatial-semantic memory supporting long-horizon exploration and recall.The policy makes decisions using accumulated memory, and the goal is to reach the target efficiently.Source: §3.1 · p. 4

3.2. Overview

  • APEX uses dynamic Attraction, Obstacle, and Exploration maps, PPO-based action selection, and a detector to confirm the target’s location.The modules run asynchronously in parallel to address latency bottlenecks.Source: §3.2 · p. 4

3.3. 3D Dynamic Grid-based Map

  • A persistent 3D grid stores spatio-semantic information, using back-projection and object-conditioned attraction scores to update its map channels.Object masks are projected into the grid; attraction updates use majority ownership and favor closer observations, exploration gains come from ray-cast visible voxels, and occupied voxels populate the obstacle map.Source: §3.3 · pp. 4–5

3.4. RL-based Action Decision Module

  • A PPO policy maps the three constructed 3D maps to low-level control actions through separate CNN extractors and shared actor and critic features.Its composite reward combines terminal success or failure with dense attraction and exploration rewards.Source: §3.4 · pp. 5–6

3.5. Target Grounding Module

  • The Target Grounding Module continuously applies an open-vocabulary detector to RGB observations and accepts a target when its top confidence exceeds a threshold.A reconstruction projector computes the target’s precise 3D world coordinate.Source: §3.5 · p. 6

3.6. Asynchronous Parallel Framework

  • The asynchronous framework shares Attraction and Exploration maps across modules running at different frequencies, while action control updates obstacle information from current sensors.The slower mapping module performs VLM inference and ray casting; high-frequency target detection runs independently of that mapping cycle.Source: §3.6 · p. 6

4. Experiment

APEX is evaluated on UAV-ON against basic, ground-navigation, and aerial-navigation baselines, with results and ablations assessing its performance and core design choices.

4.1. Experiment Settings

  • UAV-ON comprises 14 photorealistic, large-scale outdoor environments and 10,000 navigation episodes, evaluated with four standard navigation metrics.Success Rate (SR), Navigation Error (NE), Oracle Success Rate (OSR), and Success-weighted Path Length (SPL) measure success, final distance, oracle-assisted success, and success-weighted path efficiency.Source: §4.1 · pp. 6–7

4.2. Experimental Results

  • APEX improves Success Rate (SR) by 4.16% and Success-weighted Path Length (SPL) by 2.77%, and achieves the highest Oracle Success Rate (OSR).It also has the lowest step latency and highest average safe distance among the three top-performing baselines; the authors attribute higher Navigation Error (NE) to proactive exploration in unsuccessful episodes.Source: Table 1 · p. 7, Table 2 · p. 7, §4.2 · p. 8

4.1. Experiment Settings

  • The baseline set spans random and frontier exploration, semantic and language-model methods, ground ObjectNav, and aerial navigation.It includes Random Exploration, Frontier-based Exploration, MLLM-Direct Navigation, CLIP-H, L3MVN-Zeroshot, TravelUAV, and AOA-F.Source: §4.1 · p. 7
  • The Action Decision policy maps 3D grid-based maps to discrete actions and is trained with PPO using four parallel environments across 36 tasks.Training first uses exploration and terminal-penalty rewards, then adds attraction and success rewards to fine-tune goal-directed behavior.Source: §4.1 · p. 7

4.3. Ablation Study

  • An attraction-exploration trade-off parameter α of 0.2 achieves the best overall performance among the evaluated values.The authors describe this setting as balancing attraction and exploration rewards.Source: Table 3 · p. 8, §4.3 · p. 8
  • Ablations show that removing the 3D map impaired convergence, a heuristic action policy became trapped in local optima, and simple attraction-map targeting substantially reduced performance.The results support the roles of the map, RL-based Action Decision Module, and specialized Target Grounding Module.Source: Table 4 · p. 8, §4.3 · p. 8

5. Conclusion

The conclusion presents APEX as a hierarchical agent for Aerial Object Goal Navigation, combining three specialized modules through an asynchronous parallel framework. UAV-ON experiments show state-of-the-art performance, reliability, and operational efficiency.

5. Conclusion

  • APEX comprises Dynamic Spatio-Semantic Mapping, RL-based Decision Making, and Target Grounding modules for Aerial Object Goal Navigation.Source: §5 · p. 8
  • The modules are integrated through an Asynchronous Parallel Framework.Source: §5 · p. 9
  • Experiments on UAV-ON demonstrate state-of-the-art performance, superior reliability, and higher operational efficiency.Source: §5 · p. 9
  • Future work includes fine-tuning pre-trained VLM and MLLM models for Aerial ObjectNav and comparing different RL methods in APEX.Source: §5 · p. 9

6. Algorithm Flow

The agent runs its modules in parallel, updating shared memory and executing policy actions until target detection enables navigation.

6. Algorithm Flow

  • A multi-process architecture enables parallel execution of the three modules.The algorithm workflow takes observations, states, and the target description as inputs.Source: §6 · p. 12
  • The map-updating process repeatedly applies CAP and SEG, then updates shared MEM with the resulting mask and attraction scores.MEM includes the attraction, exploration, and obstacle maps, among other shared information.Source: §6 · p. 12
  • The action-decision process updates memory, obtains an action from the policy, and executes it until detection triggers navigation to p.The workflow also includes a target-grounding process, but these passages do not specify its full operation.Source: §6 · p. 12

7. More Visualization Results

The visualizations show successful searches and a failure attributed to a tree collision, alongside trajectories for four additional cases and the main-paper example.

7.1. Cases

  • A tree collision caused the Sandbox search failure, suggesting the Obstacle Map may miss sparse obstacles.The other three illustrated searches target a Car, Playground, and Umbrella.Source: §7.1 · p. 12
  • The four episodes include three successful searches and one Sandbox failure.The failure is attributed to a tree collision.Source: §7.1 · p. 12

7.2. Trajectories

  • Figure 6 shows complete trajectories for Cases 1–4 alongside the main-paper example, Case 0.The passage identifies the trajectory examples but does not report a comparative outcome.Source: §7.2 · p. 12

8. Details of Heuristic Algorithm

The heuristic evaluates simulated actions using attraction and exploration rewards, excluding unsafe outcomes before selecting the highest-reward action.

8. Details of Heuristic Algorithm

  • The heuristic simulates each possible action and selects the one with the highest predicted reward from attraction and exploration criteria.The action with the greatest predicted reward is chosen.Source: §8 · p. 12

9. More Training Details

Training uses precomputed attraction maps to avoid VLM inference in the RL loop, while a multi-map policy network processes spatial features to select actions.

9.1. Offline Map Generation

  • For each of 36 tasks, a flight path generates a complete attraction map before RL training, avoiding VLM calls during training.The agent uses depth to identify visible grid cells and looks up their precomputed values to populate its local attraction map.Source: §9.1 · p. 12

9.2. Policy Network Architecture

  • The policy network uses separate CNNs for attraction, exploration, and obstacle maps, then shares concatenated features between Actor and Critic MLP heads.The z-axis is treated as the channel dimension, and the Actor outputs the action policy.Source: §9.2 · pp. 12–13, Figure 5 · p. 13
  • Attraction-map updates match visible grid cells to masks by projected-pixel overlap and keep the score when the new observation is closer.The update iterates over visible cells and selects the mask with the largest intersection with the pixels projected to each cell.Source: §9.2 · p. 13
  • Exploration-map updates trace rays through traversed cells and increase each cell’s value according to distance and its current value.The update traces rays from the UAV position to 3D points obtained from the observation.Source: §9.2 · p. 13
  • Obstacle-map updates mark grid cells hit by depth rays as obstacles.The provided visualization captions include attraction maps, complete trajectories, and successful and failed navigation examples.Source: §9.2 · p. 13, Figure 7 · p. 14, Figure 6 · p. 14

Related papers

Loading 2602.00551v1…