Paper deep dive
LAPF: LLM-Agent-Based Path Finder Using the UAVScenes Dataset
Yousef Emami, Mohammadhossein Homaei, Hao Zhou, Miguel Gutiérrez Gaitán, Atefeh Hajijamali Arani, Rui Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 8/18/2026, 5:50:37 AM
Summary
The paper introduces LAPF (LLM-Agent-Based Path Finder), a framework for autonomous UAV navigation in town-scale outdoor environments. LAPF integrates perception, memory, planning, and action modules into a closed-loop cognitive architecture, leveraging Large Language Models (LLMs) with Chain-of-Thought (CoT) reasoning. Evaluated on the UAVScenes dataset, LAPF demonstrates superior path efficiency and safety compared to standard CoT prompting, achieving near-zero clamp events and significant path length reductions by coupling hazard detection with bounded corrective actions.
Entities (6)
Relation Signals (5)
LAPF → evaluatedon → UAVScenes
confidence 100% · LAPF is evaluated using the UAVScenes dataset [27]...
LAPF → employs → Chain-of-Thought
confidence 95% · The proposed agent leverages prior navigation experiences, performs Chain-of-Thought (CoT) reasoning...
UAV → equippedwith → Livox Avia
confidence 90% · The system consists of a single UAV equipped with... a Livox‑Avia LiDAR
LAPF → outperforms → Chain-of-Thought
confidence 90% · LAPF achieves mean path lengths... corresponding to path length reductions of 17.2% and 15.6% relative to CoT prompting
LAPF → uses → Qwen2-VL-7B
confidence 90% · Using the same Qwen2-VL-7B [26] backbone... LAPF reduces mean path length...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Uncrewed aerial vehicles (UAVs) are increasingly deployed for autonomous navigation in complex outdoor environments, where dynamic conditions and mission requirements require intelligent adaptive decision-making. Existing optimization-based, Machine Learning (ML), and Reinforcement Learning (RL) approaches often rely on predefined models or task-specific training, limiting their generalization and adaptability in uncertain scenarios. Recent Large Language Model (LLM)-assisted approaches offer promising reasoning capabilities but remain constrained by limited agentic functionality, including insufficient memory, planning, and tool interaction this http URL paper proposes an LLM-Agent-Based Path Finder (LAPF) framework for autonomous UAV navigation in town-scale outdoor environments. LAPF extends LLM-assisted navigation by integrating perception, memory, planning, and action modules into a closed-loop cognitive architecture. The proposed agent leverages prior navigation experiences, performs Chain-of-Thought (CoT) reasoning, couples each detected hazard to a bounded corrective action, and dynamically refines waypoint decisions based on environmental this http URL three independent trials per method demonstrate that LAPF achieves mean path lengths of 512.83 m and 506.37 m, compared to the straight-line optimum of 497.33 m, corresponding to path length reductions of 17.2% and 15.6% relative to CoT prompting and absolute path efficiencies of 97.1% and 98.1% in open-field and obstacle-injected scenarios, respectively. Furthermore, LAPF is the only evaluated approach that couples every detected hazard to a bounded, metric-neutral corrective action while maintaining near-goal stability, with zero clamp events in both scenarios, whereas CoT prompting increases from 9.7 to 14.0 events.
Tags
Links
- Source: https://arxiv.org/abs/2608.15175v1
- Canonical: https://arxiv.org/abs/2608.15175v1
Trouble viewing inline? Open PDF directly →
Full Text
88,455 characters extracted from source content.
Expand or collapse full text
LAPF: LLM-Agent-Based Path Finder Using the UAVScenes Dataset Yousef Emami Mohammadhossein Homaei Hao Zhou Miguel Gutiérrez Gaitán and Atefeh Hajijamali Arani and Rui Zhang Thanks: Copyright (c) 2026 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to pubs-permissions@ieee.org. Abstract Unmanned aerial vehicles (UAVs) are increasingly deployed for autonomous navigation in complex outdoor environments, where dynamic conditions and mission requirements require intelligent adaptive decision-making. Existing optimization-based, Machine Learning (ML), and Reinforcement Learning (RL) approaches often rely on predefined models or task-specific training, limiting their generalization and adaptability in uncertain scenarios. Recent Large Language Model (LLM)-assisted approaches offer promising reasoning capabilities but remain constrained by limited agentic functionality, including insufficient memory, planning, and tool interaction mechanisms. This paper proposes an LLM-Agent-Based Path Finder (LAPF) framework for autonomous UAV navigation in town-scale outdoor environments. LAPF extends LLM-assisted navigation by integrating perception, memory, planning, and action modules into a closed-loop cognitive architecture. The proposed agent leverages prior navigation experiences, performs Chain-of-Thought (CoT) reasoning, couples each detected hazard to a bounded corrective action, and dynamically refines waypoint decisions based on environmental feedback. The three independent trials per method demonstrate that LAPF achieves mean path lengths of 512.83512.83 m and 506.37506.37 m compared to the straight-line optimum 497.33497.33 m corresponding to path length reductions of 17.2%17.2\% and 15.6%15.6\% relative to CoT prompting and absolute path efficiencies of 97.1%97.1\% and 98.1%98.1\% in open-field and obstacle-injected scenarios, respectively. Furthermore, LAPF is the only evaluated approach that couples every detected hazard to a bounded, metric-neutral corrective action while maintaining near-goal stability with zero clamp events in both scenarios, whereas CoT prompting increases from 9.79.7 to 14.014.0 events. Index Terms: UAV Navigation, Large Language Model Agents, Vision-Language Models, Episodic Memory, Chain-of-Thought Reasoning, Tool-Augmented Reasoning, Multimodal Perception I Introduction Uncrewed Aerial Vehicles (UAVs) provide three-dimensional mobility, dynamic Line-of-Sight (LoS) establishment, and high data throughput, making them suitable for operation in complex outdoor environments such as public safety, environmental monitoring, agriculture, and parcel delivery [20, 1]. Their ability to dynamically adjust trajectories according to mission requirements enables flexible coverage in dense urban areas where terrestrial infrastructure may be obstructed or unavailable. Furthermore, UAVs can serve as aerial base stations, data collectors, relay platforms, and sensing nodes, extending their role from simple transportation platforms to intelligent networked systems[31]. Figure 1: Architectural comparison between standalone LLM prompting and LLM agents. Conventional prompting (top) executes a single prompt-to-answer pass without autonomous correction capabilities. In contrast, the LLM agent (bottom) establishes a closed-loop system via an internal Plan→ cycle and an Update plan feedback mechanism. As illustrated on the right, these paradigms map directly onto a control-autonomy scale, highlighting the trade-off between direct human control and system autonomy. UAV navigation and communication systems have traditionally relied on optimization-based methods and, more recently, Machine Learning (ML) and Reinforcement Learning (RL) approaches to address challenges such as trajectory planning, resource allocation, and autonomous decision-making [2]. Although these methods have achieved significant progress, their effectiveness is often limited by predefined system models, task-specific training requirements, and insufficient generalization to dynamic and uncertain environments. In particular, emergency response, cooperative UAV networking, and large-scale autonomous operations require decision-making mechanisms that can rapidly adapt to changing conditions without extensive retraining or manual redesign. Large Language Models (LLMs) have recently demonstrated remarkable capabilities in reasoning, contextual understanding, and task automation across domains such as finance, healthcare, and networking [5, 22]. Techniques like In-Context Learning (ICL) and Chain-of-Thought (CoT) reasoning [28] enable LLMs to solve complex tasks by leveraging contextual information and structured reasoning processes. These emerging capabilities have motivated their exploration in UAV and autonomous system applications [12, 8]. However, directly applying standalone LLMs to UAV navigation remains challenging due to three fundamental limitations: 1. Multimodal perception gap: Conventional text-based LLMs cannot directly interpret heterogeneous UAV sensory inputs, including RGB images, LiDAR point clouds, depth information, and 6-Degrees of Freedom (DoF) localization data, which are critical for reliable navigation. Addressing this gap requires a backbone capable of ingesting visual evidence natively, together with an interface that renders the non-visual channels in a form the backbone can reason over [6]. 2. Limited agentic capabilities: Autonomous UAV operation requires memory, adaptive planning, long-term spatial reasoning, and continual interaction with the environment, which are beyond the capabilities of conventional single-pass LLM prompting[21]. 3. Lack of tool integration: Practical navigation relies on external modules, including simulators, obstacle detection systems, localization tools, and trajectory planners. Standalone LLM prompting lacks mechanisms for reliable tool invocation and closed-loop execution[13]. Moreover, existing UAV navigation approaches based on optimization, ML, and RL often require extensive environment modeling or task-specific training, offer limited interpretability, and struggle to integrate heterogeneous perception information into a unified decision-making framework [18, 14]. These limitations motivate the development of more flexible cognitive architectures capable of combining perception, reasoning, memory, and action. LLM agents offer a promising solution by extending conventional LLMs with additional cognitive modules, including memory, planning, tool use, and action execution. As illustrated in Fig. 1, unlike standalone prompting approaches that perform a single prompt-to-response operation, LLM agents establish a closed-loop interaction process consisting of perception, reasoning, planning, execution, and feedback-based refinement. Such architectures enable autonomous systems to adapt their decisions based on environmental changes, retrieve prior experiences, and interact with external computational resources. Therefore, LLM agents provide a potential pathway toward intelligent UAV navigation with enhanced situational awareness, cross-layer coordination, and long-horizon decision-making capabilities. Nevertheless, several challenges remain, including reliable reasoning, real-time operation, domain-specific adaptation, and integration with existing UAV control frameworks[32, 13, 17]. Motivated by these opportunities, this work proposes the LLM-Agent-Based Path Finder (LAPF), a cognitive navigation framework for UAVs designed to enable adaptive waypoint generation and autonomous replanning. LAPF is evaluated using the UAVScenes dataset [27], whose town-scale outdoor sequences provide a realistic evaluation environment for multimodal UAV perception and navigation. UAVScenes is a multimodal UAV perception benchmark combining RGB images, LiDAR, 6-DoF poses, semantic labels, and 3D maps. It supports diverse tasks such as segmentation, localization, place recognition, depth estimation, and novel view synthesis. The dataset enables evaluation of advanced learning-based and LLM-assisted UAV navigation systems. The LAPF agent integrates perception, memory, planning, and action modules to perform CoT reasoning, retrieve previous navigation experiences, invoke external tools such as obstacle checkers and trajectory validators, and generate candidate waypoints. During execution, newly observed environmental changes can trigger autonomous replanning, enabling continuous refinement of navigation decisions through a closed-loop perception–reasoning–action process. The main contributions of this paper are summarized as follows: • LAPF, an LLM-agent-based autonomous UAV navigation framework, extends conventional LLM prompting by integrating multimodal perception, episodic memory, and tool-enforced action within a closed-loop architecture. It grounds reasoning in UAVScenes observations by combining native RGB input to an open-weight Vision-Language Model (VLM) with symbolic LiDAR and 6-DoF pose representations in a single inference call. Running entirely on a consumer laptop, LAPF demonstrates that its performance gains stem from the agent architecture rather than a proprietary backbone. The framework leverages memory and external tools for adaptive waypoint generation, safety validation, and trajectory refinement. Its two-tier Action module guarantees safety-critical hazard checks at every decision step while allowing discretionary tool use, ensuring hazards always trigger corrective actions, a reliability capability absent in prompting-only methods. Consequently, LAPF enables adaptive perception–reasoning–action cycles without requiring additional model fine-tuning. • The effectiveness of LAPF is validated on the UAVScenes AMtown03 sequence under open-field and obstacle-injected scenarios. Using the same Qwen2-VL-7B [26] backbone and identical multimodal observation sequences for all methods, LAPF reduces mean path length by 17.2%17.2\% and 15.6%15.6\% relative to CoT prompting across the two scenarios, reaches path efficiencies of 97.05%97.05\% and 98.09%98.09\% respectively, and maintains zero near-goal clamp events in both, indicating that its advantage is preserved rather than eroded as environmental complexity increases. The remainder of this paper is organized as follows. Section I reviews related work. Section I describes the system model, and Section IV formulates the path-planning objective. Section V details the proposed LAPF framework. Section VI presents the numerical results and discussion. Finally, Section VII concludes the paper. Table I: Comparative Analysis of LAPF Against Related UAV and LLM-Agent Frameworks Framework Domain Multimodal Input Explicit Memory Tool Use Fine-tuning Required UAV- specific Real Sensor Fusion [16] Urban UAV safety path planning ✓ × × × ✓ × [23] Urban UAV obstacle avoidance × × × × ✓ × [19] Urban UAV reconstruction ✓ × × × ✓ ✓ [15] Risk-aware UAV path planning ✓ × × × ✓ × [4] UAV optimization-based path planning (LGPSO) × × × × ✓ × [30] GNSS-aware urban UAV navigation ✓ × × × ✓ ✓ [18] Wireless slicing × ✓ ✓ × × × [29] Wireless multi-task optimization × × × ✓ × × [3] 6G network engineering × ✓ ✓ ✓ × × [25] Wireless agent benchmark × × ✓ – × × [12] LLM-UAV general framework – – – – ✓ – [11] UAV scheduling × × × × ✓ × [9] UAV trajectory and velocity optimization × × × × ✓ × [10] UAV AoI and flight resource allocation × × × × ✓ × [24] UAV autonomous navigation (AutoFly) ✓ × × ✓ ✓ ✓ [33] UAV outdoor VLN (FreeFly-thinking) ✓ × × ✓ ✓ × [7] Self-evolving VLN (SE-VLN) ✓ ✓ × ✓ × × LAPF (this work) UAV navigation ✓ ✓ ✓ × ✓ ✓† †Real recorded UAVScenes streams. The RGB frame is consumed natively by the VLM backbone; the LiDAR return and 6-DoF pose enter as a symbolic state at the perception interface. “–” denotes not reported or not applicable. I Related Work This section reviews classical UAV path-planning approaches, recent LLM-agent frameworks for wireless networks, and emerging LLM-assisted UAV navigation methods. Finally, we identify the limitations of existing approaches and highlight the research gap that motivates the proposed LAPF framework. I-A Classical Path Planning Iwashina et al.[16], Sui et al.[23], and Kuang et al.[19] study UAV path planning in urban environments, focusing on safety, efficiency, and autonomy. Iwashina et al. propose a V2X-based dynamic risk map with RRT for collision-aware navigation in changing urban conditions. Sui et al. apply RRT to obstacle-dense environments and demonstrate faster computation, greater adaptability, and smoother paths compared with A* and ACO. Kuang et al. develop a framework for real-time urban mapping that combines path generation, height estimation, SLAM-based image acquisition, and coverage prediction to improve reconstruction quality and reduce operational time. Hu et al.[15] suggest a comprehensive UAV risk assessment framework for safe urban operations by quantifying collision risks involving people, vehicles, and manned aircraft. Their approach integrates risk costs into UAV path planning by generating risk maps and optimizing flight trajectories. Cheng et al.[4] present an improved particle swarm optimization-based UAV path planning algorithm (LGPSO) to address limitations of traditional PSO, such as convergence to local optima and instability in complex urban environments. By incorporating safety constraints, stochastic inertia weights, asymmetric learning factors, Lévy flight strategies, and a greedy evaluation mechanism, the proposed method enhances path optimization and search efficiency. Zhang et al.[30] address the challenge of Global Navigation Satellite System (GNSS) positioning errors in urban UAV operations caused by multipath effects and Non-Line-of-Sight (NLOS) conditions. They propose a risk-aware path planning approach that predicts GNSS errors using ray-tracing simulations and 3D building models, then integrates the predicted positioning error into an enhanced A* path planning algorithm. The results show that accounting for localization uncertainty during route planning enables UAVs to select safer paths with fewer positioning errors compared with conventional distance-based planning methods. I-B LLM Agent for Wireless Networks Tong et al. [18] introduce WirelessAgent, an LLM-agent framework for intelligent wireless network management built on four modules: perception, memory, planning, and action. Using LangGraph-based agent workflows, it enables autonomous decision-making and achieves higher bandwidth utilization than prompt-based methods while approaching the performance of rule-based optimization. Yu et al. [29] present WirelessLLM-Agent, a multi-task wireless optimization framework that uses semantic state serialization and Mixture-of-Experts (MoE)-LoRA for efficient knowledge sharing. A two-stage training strategy that combines supervised fine-tuning with Group Relative Policy Optimization (GRPO)-based RL improves long-term decision-making. Chen et al.[3] develop 6GAgentGym and 6G-Forge, a closed-loop tool-based environment for training and evaluating wireless agents through NS-3 simulations, enabling small open-source models to achieve competitive performance on long-horizon 6G tasks. To address the limitations of existing benchmarks, Tong et al. [25] introduce a tolerance-aware benchmark covering reasoning, resource allocation, and mobility-aware decisions, with tool-based evaluation and catastrophic-error detection to identify practical agent failures. I-C LLMs for UAV Networks Emami et al. have investigated the integration of LLMs with UAV networks through a series of contributions that establish a foundation for intelligent UAV decision-making. In their survey [12], they present a unified framework for LLM-assisted UAV systems, summarizing existing architectures, methodologies, and applications, and highlighting the potential of LLMs as high-level reasoning modules for autonomous aerial networks. Building on this perspective, they propose ICL-based solutions for UAV missions, including Search-and-Rescue (SAR) operations, where LLM-assisted scheduling enables adaptive data collection by jointly considering dynamic channel conditions, battery constraints, and queue states [11]. Extending the role of LLMs from task scheduling to real-time control, they further investigate public safety applications by employing LLM-assisted ICL for UAV trajectory planning and adaptive velocity control [9]. To address freshness-aware sensing scenarios, they introduce FRSICL, an online resource allocation framework that integrates LLM-assisted reasoning with optimization techniques to jointly determine data collection schedules and UAV velocities, thereby minimizing the average Age of Information for ground sensors [10]. Beyond pure LLM-assisted reasoning, recent works have explored the integration of multimodal foundation models to enable UAVs with richer perception and embodied intelligence. Sun et al.[24] introduce AutoFly, an end-to-end Vision-Language-Action (VLA) model that enables UAVs to navigate unknown environments using coarse guidance rather than predefined trajectories. By incorporating pseudo-depth spatial reasoning, progressive multimodal training, and a real-world autonomous navigation dataset, AutoFly improves UAV decision-making and achieves a 3.9% higher success rate than existing VLA baselines. Similarly, Zhou et al.[33] suggest an end-to-end Vision-Language Navigation framework that enables UAVs to perform outdoor navigation through explicit reasoning over visual observations and language instructions. Their framework combines a dedicated UAV navigation dataset with CoT reasoning and a two-stage training strategy consisting of supervised fine-tuning and reinforcement learning, improving robustness and generalization in unseen environments. Going beyond static model deployment, Dong et al.[7] introduce a self-evolving Vision-Language Navigation framework that allows agents to continuously improve during testing through experience-driven learning. By integrating hierarchical memory, retrieval-augmented reasoning, and reflection mechanisms, their approach enhances long-term adaptation and achieves significant performance improvements in unseen environments. In contrast, among the reviewed approaches, LAPF combines explicit memory, tool-enforced action, and no fine-tuning requirement. It is also grounded in multimodal UAV data (RGB, LiDAR, and 6-DoF pose) from a real-world UAV dataset (UAVScenes), rather than synthetic states. The reviewed wireless-domain agents show the value of memory and tool use. However, they lack multimodal grounding and aerial safety-critical design. UAV-specific LLM works avoid fine-tuning through prompting alone, but lack both memory of prior outcomes and a tool-use path to act on an unsafe classification. LAPF’s Action module addresses this gap. Classical planners remain efficient for narrowly defined geometric objectives, such as risk-aware navigation, obstacle avoidance, and coverage mapping. They do not natively interpret natural-language mission intent, retain memory across missions, or jointly reason over heterogeneous constraints without planner redesign. LAPF is designed to complement rather than replace classical planners. Overall, the reviewed literature reveals a consistent gap: existing LLM-agent frameworks are either agentic but domain-general (equipped with memory and/or tool use, but validated on wireless network management rather than physical navigation, and without multimodal sensor grounding), or UAV-specific but architecturally shallow (effective at single-shot or in-context optimization of scheduling and trajectory parameters, but reliant on prompting alone, with no persistent memory of past navigation outcomes and no enforced mechanism to translate a detected hazard into corrective action). To the best of our knowledge, few UAV navigation systems combine multimodal grounding, episodic memory, and tool-enforced safety control in a unified LLM-agent architecture. LAPF is designed to address this gap by combining these capabilities within a closed-loop UAV navigation architecture. Table I provides a comparative analysis. Figure 2: A typical UAV path planning pipeline with LLM agents. The agent receives multimodal input from the UAVScenes dataset and, through an iterative Reason→ → reasoning cycle, leverages external tools and memory in a bidirectional exchange to generate the UAV’s navigation decisions, one waypoint per decision step. I System Model This section presents the system model of the proposed LAPF framework for UAV navigation. The system consists of a single UAV equipped with a multi‑modal sensor suite, an onboard or edge‑based LLM agent, with the UAVScenes dataset serving as the source of environmental observations for evaluation, as depicted in Fig. 2. The proposed LAPF enables adaptive path planning through closed‑loop interaction between the UAV and the LLM agent. The UAV is equipped with a synchronized camera–LiDAR configuration, matching that used in UAVScenes and comprising a Livox‑Avia LiDAR and a high‑resolution RGB camera (2448×20482448× 2048 pixels at 10 Hz) as used in UAVScenes. This combination provides dense 2D texture information and accurate 3D spatial measurements. The system also includes GNSS/inertial measurement unit (IMU) measurements with Real-time Kinematic (RTK) support, which contributes to 6‑DoF pose estimation (3‑DoF position and 3‑DoF orientation). All sensor streams are hardware‑synchronised to produce frame‑wise aligned image–LiDAR pairs. The UAV operates in environments represented in the UAVScenes dataset, which include towns, valleys, airports, and islands. However, extension to dense urban environments remains future work because UAVScenes currently does not include high‑rise buildings, narrow streets, and pedestrian crowds. The UAV is equipped with: • An RGB camera capturing high‑resolution images (2448×20482448× 2048 pixels) at 10 Hz, synchronized with the LiDAR. • A Livox Avia LiDAR sensor providing frame‑wise 3D point clouds at 10 Hz. • A hardware‑synchronized camera‑LiDAR suite ensuring pixel‑wise alignment via pre‑calibrated extrinsics. • An onboard localization system that supplies 6-DoF poses, reconstructed using structure from motion (SfM) in DJI Terra and refined from the original 4‑DoF GNSS/RTK data. These sensors generate multi‑modal observations at each time step t, denoted as: tfull=ℐt,t,t,t,ℒtO_t^full=\I_t,P_t,D_t,S_t,L_t\ (1) where • ℐtI_t is the RGB image. • tP_t is the raw LiDAR point cloud (Livox Avia, after noise filtering). • tD_t is the dense depth map derived from projecting tP_t onto the image plane using the known calibration. • tS_t is the semantic segmentation mask, comprising 16 static classes (e.g., road, building, and vegetation), 2 dynamic object classes (sedan and truck only; pedestrians and other small dynamic objects are not annotated), and 1 background class – for a total of 19 classes. Annotations are provided for both the image and the LiDAR point cloud on a per‑frame basis. • ℒtL_t is the 6‑DoF pose (x,y,z,ψ,θ,ϕ)(x,y,z,ψ,θ,φ) (position and Euler angles) reconstructed via SfM and aligned with the 3D map. The median positional accuracy on benchmark splits is approximately 6.1 m for state‑of‑the‑art methods, indicating that the benchmarked methods do not achieve submeter localization accuracy. Eq. (1) specifies the full sensor suite available in UAVScenes; the subset instantiated in this work is denoted by tO_t and defined in (2), as summarized in Table I. The three instantiated channels reach the backbone through two distinct pathways. The RGB frame ℐtI_t is registered, time-matched and resampled to 1024×768, and is then passed natively to the VLM backbone as an image input, so that visual scene evidence is available to the reasoning stage without an intermediate captioning step. The LiDAR point cloud and the pose, by contrast, are not visual quantities and are rendered symbolically: tP_t enters through the summary statistic σt _t of (3) rather than as raw points, and the pose enters as ℒtL_t. Fusion of the visual and symbolic pathways therefore occurs inside the backbone at inference time, whereas the LiDAR return is reduced to a geometric abstraction at the interface. This separation is deliberate: it keeps the safety-critical clearance test of (5) on a deterministic geometric channel, independent of the backbone’s visual inference, while still exposing the scene appearance to the planner. The dense depth map tD_t and semantic mask tS_t, although annotated in UAVScenes, are not supplied in this work. Table I: Instantiation of the system model in the present implementation Symbol System model Instantiation ℐtI_t RGB image Passed natively to the VLM backbone at 1024×768; registered and time-matched tP_t Raw point cloud Entered via the summary σt _t, (3) ℒtL_t 6-DoF pose Pose (Rt,pt)(R_t,p_t) with the time-matched RTK fix tD_t Dense depth map Not fused (future work) tS_t Semantic mask Not fused (future work) MtM_t Episodic memory k=3k=3 most recent episodes, recency-based hth_t Hazard indicator [ct≤dsafe]1[c_t≤ d_safe], (5) ℛR Corrective operator Deterministic contraction toward pgoalp_goal, (6) Figure 3: Overview of the proposed LAPF framework. The observation reaches the backbone through two distinct pathways: the RGB frame is passed natively as an image block and is never textualized, whereas the LiDAR return and the pose are reduced to a symbolic state; fusion of the two pathways occurs inside the backbone at inference time, within a single multimodal call. The two-tier Action module evaluates the clearance test with guaranteed invocation at every decision step, so that a registered violation is always coupled to a bounded corrective replan, which is metric-neutral by construction; the discretionary tier remains under agent authority. IV Problem Formulation This section formulates the UAV navigation problem addressed by LAPF. We first define the observation model and sequential decision process, followed by the memory-conditioned decision policy and the enforced hazard-response mechanism. Finally, we present the optimization objective and the performance metrics used to evaluate the proposed framework. This paper aims to minimize the projected mission length L~ L of (10) while ensuring an enforced hazard response during autonomous navigation in the considered environment. The proposed LAPF leverages multimodal environmental information obtained from the UAVScenes dataset, including RGB images, summarized LiDAR returns, and 6-DoF localization states, to enable path planning through an LLM-assisted autonomous agent. The UAV agent dynamically reasons about obstacle distributions, traversability conditions, localization confidence, and mission objectives to generate adaptive navigation trajectories. IV-A Decision Process and Observation Model The mission is modelled as a discrete-time, partially observable sequential decision process over a horizon TmaxT_ . At decision step t the agent holds a commanded state wpt∈ℝ3wp_t ^3, initialized at wp0=p0wp_0=p_0, and receives an observation t=ℐt,σt,ℒt,O_t= \I_t,\; _t,\;L_t \, (2) where ℐtI_t is the RGB frame, ℒt=(Rt,pt)∈SE(3)L_t=(R_t,p_t)∈ SE(3) is the 6-DoF pose, and σt _t is a compact statistic of the LiDAR return, σt=(nt,rtmin,rtmax,btmin,btmax,ct),ct=minixi,\!\! _t= (n_t,\,r_t ,\,r_t ,\,b_t ,\,b_t ,\,\!c_t ), c_t= _i\;x_i, (3) in which xix_i is the boresight coordinate of return ip_i, ntn_t is the return count, [rtmin,rtmax][r_t ,r_t ] are the bounds of the Euclidean range, [btmin,btmax][b_t ,b_t ] are the lateral bounds in the plane orthogonal to the boresight, and ctc_t is the smallest boresight coordinate. Since ct≤∥i∥c_t≤ _i for the corresponding return ip_i, ctc_t lower-bounds the Euclidean range. The test in (5) is therefore conservative by construction and cannot under-declare a hazard. The Livox-Avia unit in UAVScenes is nadir-mounted and hardware-aligned with the RGB camera, so ctc_t is the clearance between the platform and the nearest surface below it. Across the evaluated sequence the returns span a 77∘77 cone, matching the nominal sensor field of view, and the boresight coordinate is tightly concentrated within each frame, as expected for a downward survey over near-planar terrain. The sequence is flown at survey altitude throughout, and the sensed clearance ctc_t remains above 1515 m in every frame, the lowest values arising from tree canopy. Scope of the safety test. The enforcement mechanism of Section IV-C is defined over a scalar hazard signal and is agnostic to the channel supplying it. Whatever geometric quantity is placed in σt _t, its registration is guaranteed to be coupled to the bounded response of (6). This is a property of the action layer rather than of the sensor, and it is what prompting-only pipelines lack. The channel instantiated here is the boresight clearance ctc_t, which measures separation from the nearest surface below the platform. Two properties of this channel are stated explicitly. First, ctc_t certifies clearance along the boresight and not along the direction of commanded motion, so (15) guarantees a prescribed response to a registered clearance violation rather than horizontal collision avoidance. This is consistent with the disclaimer accompanying Property 1. Second, the return resolves range but not bearing within the sensed footprint, so a waypoint-level clearance ∥wp−o∥ -o cannot be evaluated without assuming a bearing that σt _t does not certify. The test is consequently expressed at the level of the observation, which is the conservative choice under this abstraction. This scope follows from the UAVScenes sensor configuration rather than from LAPF. Under the common-observation protocol it applies identically to all three evaluated methods and therefore does not affect the comparison. Lateral scene structure remains available to the planner through the nadir RGB frame, which the VLM backbone consumes directly. That channel is excluded from enforcement because a safety-critical guarantee cannot depend on a stochastic visual inference. Extending enforcement to a bearing-resolved channel is identified as future work. IV-B Memory-Conditioned Decision Policy Let Mt=(j,wpj,yj,ρj)j=t−kt−1M_t= \(O_j,wp_j,y_j, _j) \_j=t-k^t-1 denote the episodic memory of the k most recent decision episodes, where yjy_j is the execution outcome and ρj _j the generated reflection, and let ϕt _t be the scalar performance-feedback signal derived from the trailing window of path-length statistics. The agent implements the policy π:(sys,,t,Mt,ϕt)⟼(wp^t,Toolst),π:\; (P_sys,\,G,\,O_t,\,M_t,\, _t )\; \; ( wp_t,\;Tools_t ), (4) where G is the mission specification, wp^t wp_t the proposed waypoint and ToolstTools_t the requested tool set. IV-C Enforced Hazard Response and Stability Regulation The proposed waypoint wp^t wp_t is not executed directly. Given the minimum admissible clearance dsafed_safe, the hazard indicator ht=[ct≤dsafe]h_t=1 [\,c_t≤ d_safe\, ] (5) is evaluated at every decision step with guaranteed invocation, so that its activation does not depend on whether the corresponding tool is named in a particular sampling of ToolstTools_t;here, [⋅]1[·] denotes the indicator function, the agent retains full authority over wp^t wp_t and over the discretionary tier of ToolstTools_t. When ht=1h_t=1, the model-authored proposal is discarded and replaced by the deterministic corrective operator ℛ(wpt−1)=wpt−1+κ(pgoal−wpt−1),κ∈(0,1),R(wp_t-1)=wp_t-1+κ (p_goal-wp_t-1 ), κ∈(0,1), (6) which is a function of the commanded state and the mission goal alone. This operator serves strictly as a low-level fail-safe guardrail rather than an intelligent planner. Its role is not to optimize the trajectory, but to provide a prescribed and bounded reaction in critical states. The cognitive value of the framework lies in the agent’s ability to use memory and continuous reasoning to anticipate hazards, thereby learning to navigate efficiently to minimize the activation of this deterministic shield. When triggered, the executed displacement becomes ∥ℛ−wpt−1∥=κ∥pgoal−wpt−1∥ -wp_t-1 =κ p_goal-wp_t-1 , which is bounded and reproducible across trials, and independent of temperature sampling, in contrast to an arbitrary model-authored coordinate. The magnitude κ∥pgoal−wpt−1∥κ p_goal-wp_t-1 is a commanded set-point rather than an instantaneous displacement: the operator specifies where the vehicle is directed once a hazard is registered, and the rate at which that set-point is tracked is delegated to the low-level flight controller, which is outside the scope of the decision layer modelled here. Enforcing a decision-layer displacement bound jointly with C, so that set-point and kinematic feasibility are regulated by a single mechanism, is identified as future work. We state explicitly that ℛR enforces a bounded and deterministic response to a detected hazard, and not a geometric clearance guarantee, which (3) does not support. Although ℛR contracts toward pgoalp_goal, it is not an implicit optimizer of the reported objective: Property 1 below shows that it leaves L~t L_t exactly invariant. A second operator regulates step magnitude in the terminal phase, where independently resampled coordinates would otherwise be charged as travelled distance. For a goal tolerance ε and a maximum near-goal displacement δmax _ , (w)=wpt−1+δmaxw−wpt−1∥w−wpt−1∥,∥wpt−1−pgoal∥<ε,∥w−wpt−1∥>δmax,w,otherwise.C(w)= caseswp_t-1+ _ w-wp_t-1 w-wp_t-1 ,& subarrayc _t-1-p_goal < ,\\ w-wp_t-1 > _ subarray,\\[6.0pt] w,&otherwise. cases (7) together with a hold-position operator, invoked when the agent selects WaypointAdjust, ℋ(w)=wpt−1,∥wpt−1−pgoal∥<ε,w,otherwise.H(w)= caseswp_t-1,& _t-1-p_goal < ,\\ w,&otherwise. cases (8) The executed waypoint is obtained by the ordered composition wpt _t =ℋ 1[WaypointAdjust∈Toolst](htℛ(wpt−1) =H^\,1[ WaypointAdjust _t] (h_t\,R(wp_t-1) (9) OPEN+(1−ht)(wp^t)). +(1-h_t)\,C ( wp_t ) ). Equation (9) defines a shielded policy πsafe _safe: the language model retains authority over the waypoint on every step classified as hazard-free, while C constrains only model-authored proposals and is deliberately not applied to ℛR, giving hazard response strict precedence over step-magnitude regulation. IV-D Objective and Reported Estimators Let Lt=∑i=1t∥wpi−wpi−1∥L_t= _i=1^t _i-wp_i-1 denote the travelled length and L⋆=∥pgoal−wp0∥L = p_goal-wp_0 the straight-line optimum. The projected mission length, path efficiency, detour ratio and clamp count are defined as L~t=Lt+∥pgoal−wpt∥,ηt=L⋆L~t,ϱt=ηt−1, L_t=L_t+ p_goal-wp_t , _t= L L_t, _t= _t^-1, (10) Nc=∑t=1Tmax[is active at step t],Nh=∑t=1Tmaxht.N_c= _t=1^T_ 1 [\,C\ is active at step t\, ], N_h= _t=1^T_ h_t. (11) The estimator L~t L_t augments the travelled length with the residual distance to the goal, so that a trajectory is not credited for terminating early away from pgoalp_goal. By the triangle inequality L~t≥L⋆ L_t≥ L , hence ηt≤1 _t≤ 1, with equality attained exactly when the commanded sequence is collinear with, and monotone along, the start–goal segment. Reported efficiencies are means of the per-trial ratios ηt _t rather than the ratio of the mean path lengths; by Jensen’s inequality the two differ slightly. Property 1 (Metric neutrality of ℛR). Let Δt=L~t−L~t−1 _t= L_t- L_t-1. The triangle inequality gives Δt≥0 _t≥ 0, with equality if and only if wptwp_t lies on the segment [wpt−1,pgoal][wp_t-1,p_goal], which (6) places it on for every κ∈(0,1)κ∈(0,1). Since L~0=L⋆ L_0=L , this yields the exact decomposition L~Tmax=L⋆+∑t:ht=0Δt,Δt≥0. L_T_ =L +\!\! _t\,:\,h_t=0\!\! _t, _t≥ 0. (12) Hence the enforcement pathway contributes identically zero to ηt _t and is scored through NhN_h rather than through ηt _t. This is a designed separation of concerns: the shield is required to be metric-neutral so that safety enforcement and trajectory quality are never conflated. The decomposition (12) is an exact identity on the realized trajectory and so invokes no counterfactual. Neutrality is moreover not obtained at the expense of goal progress, since ∥pgoal−ℛ(wpt−1)∥=(1−κ)∥pgoal−wpt−1∥ p_goal-R(wp_t-1) =(1-κ) p_goal-wp_t-1 : the residual distance entering L~t L_t is strictly reduced, so enforcement neither charges the metric on the step it acts nor imposes any additional penalty on the evaluation of subsequent steps. We claim metric neutrality and monotone reduction of the residual term only, and not a geometric reachability guarantee, which (3) does not support. The role of ℛR is to render the hazard response bounded, deterministic and reproducible across trials, a guarantee unavailable to the prompting-only baselines evaluated in this work, which register the same hazard classification but possess no pathway to act on it. The metric therefore admits its supremum for a geometrically direct command sequence and does not by itself encode a per-step kinematic bound. The navigation problem solved by LAPF is minπ[L~Tmax] _π\;\;E [ L_T_ ] (13) subject to the goal-reaching condition ∥wpTmax−pgoal∥≤ε _T_ -p_goal ≤ (14) and to the enforcement condition ht=⟹∥wpt−wpt−1∥=κ∥pgoal−wpt−1∥,∀t≤Tmax,h_t\!=\!1\!\! \;\!\!\! _t-wp_t-1 =κ p_goal-wp_t-1 ,\!\!\!\!\!\!\!\! ∀\,t≤ T_ , (15) which requires the hazard response to be a prescribed, model-independent displacement rather than an arbitrary model-authored coordinate, and which (9) satisfies by construction. Condition (14) is evaluated at the terminal step of the fixed horizon rather than used as a stopping rule, so that all methods are scored over an identical step budget, the residual term of L~t L_t penalizing any trajectory that has not met it. Table I: Summary of the Proposed LAPF Framework Modules Module Input Output Main Functions Perception RGB image, LiDAR point cloud, and 6-DoF pose RGB image and symbolic state Registers and synchronizes RGB images for native VLM inference, converts LiDAR and localization data into a compact symbolic representation, and summarizes traversability, obstacle proximity, environmental semantics, and navigation constraints to provide situational awareness for downstream reasoning. Memory Current observation, executed waypoint, and execution outcome Retrieved navigation experience Stores one navigation episode per decision step together with an LLM-generated reflection, retrieves relevant historical experiences, and maintains persistent memory to support temporal continuity, long-term spatial awareness, and adaptive reasoning. Planning Current observation and retrieved memory Candidate waypoint Performs high-level multimodal reasoning using CoT to generate one candidate waypoint per decision step, together with any required corrective tool calls, according to the current scene and historical context. Action Candidate waypoint and agent tool calls Executable waypoint Coordinates external tool invocation, executes discretionary waypoint-adjustment operations when required, and translates high-level planning decisions into executable navigation commands. V Proposed LAPF This section presents the proposed LAPF framework for autonomous UAV navigation. We first describe its overall architecture and operational workflow, followed by the agent’s perception, memory, planning, and action modules. Finally, the structured prompting strategy and implementation details are introduced to illustrate how LAPF performs adaptive, closed-loop navigation. LAPF provides a framework for autonomous UAV navigation using an LLM agent. The system comprises a UAV tasked with navigation in the considered environment, while an edge server hosts the agent and supports memory updates. LAPF casts navigation as a closed loop rather than a single observation-to-waypoint mapping, with five stages per step t tied to the formulation of Section IV: • Perception: produces the observation t=ℐt,σt,ℒtO_t=\I_t, _t,L_t\ of (2), keeping the RGB frame as an image and reducing the LiDAR return and pose to a symbolic state. • Memory: retrieves the k=3k=3 most recent episodes MtM_t, each carrying an executed waypoint, its outcome, and a reflection, supplying temporal context a single pass lacks. • Planning: evaluates the policy π:(sys,,t,Mt,ϕt)↦(wp^t,Toolst)π:(P_sys,G,O_t,M_t, _t) ( wp_t,Tools_t) of (4), conditioning the proposed waypoint jointly on observation, memory, and feedback. • Action: does not execute wp^t wp_t directly; the guaranteed-invocation hazard test of (5) runs first, a registered hazard is replaced by the bounded operator ℛR of (6), and the executed waypoint follows the shielded composition of (9). This is the only stage that performs external tool invocation. • Feedback: scores the step, updates ϕt _t, and writes the episode back into MtM_t, closing the loop. The non-agentic baselines collapse this loop to the special case Mt=∅,ϕt=∅M_t= ,\ _t= , receiving the identical observation and the same operators, so the measured advantage isolates the memory and feedback channels rather than reasoning depth. Fig. 3 illustrates the end-to-end operation of the proposed LAPF framework, integrating multimodal perception, memory-augmented reasoning, and tool-enforced action within a closed-loop navigation cycle. The process begins with the acquisition of multimodal environmental observations. The full sensor suite tfullO_t^full of (1) is available in UAVScenes; the pipeline evaluated in this work instantiates the subset tO_t of (2), as detailed in Table I. The depth and semantic channels are reserved for future integration and are therefore not shown in the figure. The scenario evaluated in Section VI uses injected clearance violations, as described there. The pair (ℐt,symbolic prompt)(I_t,\,symbolic prompt) constitutes the interface between the perception pipeline and the agent backbone, referred to as the Cognitive Navigation Core. The framework is organized into four interdependent modules, namely Perception, Memory, Planning and Action, summarized in Table I; Fig. 3 arranges them by execution stage. The Perception module registers and time-matches the RGB frame for native ingestion by the VLM backbone, and separately reduces the LiDAR point cloud, obstacle range, and localization state into the compact symbolic representation that accompanies the image in the same inference call. This module provides step-synchronous situational awareness to the agent by summarizing traversability conditions, collision risks, environmental semantics, and navigation constraints, so that downstream reasoning is based on up-to-date environmental observations. The Memory module stores one episode per decision step, comprising the observation, the executed waypoint, the execution outcome, and an LLM-generated reflection on the discrepancy between the two, and maintains a persistent state representation that enables the LLM agent to recall prior navigation experiences. This memory-enhanced mechanism supports temporal continuity, adaptive reasoning, and long-term spatial awareness, allowing the agent to adapt navigation decisions over time through accumulated environmental knowledge. These two modules jointly inform the Planning module, which performs high-level CoT reasoning and autonomous decision-making to support UAV path planning and trajectory generation using techniques such as ICL and CoT reasoning. Based on the current scene understanding and historical context, the module generates a single candidate waypoint per decision step, aiming to minimize path length while supporting successful mission completion. The action module translates the planned navigation strategies into executable UAV control commands, such as waypoint updates, through textual outputs, API calls, or low-level navigation instructions. Table I provides a summary of the Proposed LAPF Framework Modules. Tool invocation is mediated through a dedicated external tool suite, comprising an Obstacle Checker, Trajectory Validator, Localization service, Simulator, Vector Database, and Knowledge Base, enabling the agent to evaluate candidate trajectories against safety and feasibility constraints prior to execution. The evaluation reported in Section VI instantiates the two tiers of this suite that bear on the reported metrics, namely the guaranteed-invocation Obstacle Checker and the discretionary waypoint adjuster; the remaining interfaces are defined by the architecture and left for future instantiation. Upon successful validation, the resulting waypoint is transmitted to the UAV controller for execution; the waypoint sequence wpt\wp_t\ accumulates in logT_log over the horizon. This constitutes a single decision cycle within the broader closed-loop autonomous navigation process depicted at the bottom of Fig. 3, wherein the UAV iteratively perceives its environment, encodes the non-visual channels into a structured prompt that accompanies the RGB frame, reasons over the resulting multimodal input using both memory and external tools, generates and refines a candidate trajectory, validates that trajectory for safety and feasibility, executes the resulting validated path, observes the subsequent environmental state t+1O_t+1, and updates its memory to reflect the outcome of the completed action. This iterative perception-reasoning-action cycle enables LAPF to adapt its navigation strategy in response to evolving environmental conditions without requiring gradient-based retraining of the underlying language model, in contrast to conventional reinforcement-learning-based UAV navigation approaches during execution. The cycle described here is the general operation of the architecture; under the replay-based protocol described in Section VI the recorded observation stream is fixed, so the executed waypoint does not alter subsequent observations; all methods therefore receive an identical observation sequence. Algorithm 1 details the operational flow of the proposed LAPF framework. The procedure takes as input the UAVScenes dataset SD_\!S, a mission specification G, the UAV U, an edge server ℰE hosting the LLM agent, a system prompt sysP_sys, and a maximum step budget TmaxT_ . It initializes an empty memory store ℳM and an empty trajectory log logT_log, and subsequently enters a fixed-horizon control loop that executes while step<Tmaxstep<T_ (lines 1–4). Perception (lines 5–8): at each decision step, the algorithm retrieves the current frame’s multimodal data directly from the dataset, comprising the RGB image ℐtI_t, the LiDAR point cloud tP_t, and the 6-DoF pose ℒtL_t. The point cloud is reduced to the summary statistic σt _t of (3), and the resulting quantities are assembled into the observation tO_t of (2). The non-visual channels (σt,ℒt)( _t,L_t) are then rendered into the symbolic state sts_t, whereas ℐtI_t is retained as an image and is not textualized. Memory (lines 9–11): the memory store ℳM is queried for the k most recent prior episodes MtM_t (with k=3k=3), so that retrieval is recency-based rather than similarity-based, enabling the agent to condition its subsequent reasoning on previously encountered trajectories, failures, and successful outcomes. The scalar performance-feedback signal ϕt _t of (4) is computed over the trailing window of path-length statistics. Planning (lines 12–14): the mission goal G, the symbolic state sts_t, the retrieved memory MtM_t, and the feedback signal ϕt _t are combined into a context CtxCtx. This context, together with the system prompt sysP_sys and the image ℐtI_t, is passed to the agent backbone ℰ.Agent(⋅)E.Agent(·), which returns a reasoning trace ThoughtThought, a proposed waypoint WaypointWaypoint, and a set of tool calls ToolCallsToolCalls in a single multimodal inference call. Action (lines 15–24): Hazard evaluation is performed with guaranteed invocation at every decision step, so that its activation does not depend on whether the corresponding tool is named in the tool set returned at line 14. It evaluates the sensed clearance ctc_t against the threshold dsafed_safe, returning the hazard indicator h of (5). If h=1h=1, the deterministic corrective operator (6) overwrites WaypointWaypoint with a value computed from the commanded state and the mission goal; otherwise the stability operator (7) is applied to the model-authored proposal. If WaypointAdjust appears in ToolCallsToolCalls, the hold-position operator ℋH of (8) is subsequently applied, completing the ordered composition of (9). Execution (lines 25–26): the resulting waypoint, potentially modified by the Action step, is transmitted to the UAV via U.execute(Waypoint)U.execute(Waypoint), and outcome feedback is subsequently retrieved via U.get_feedback()U.get\_feedback(). Memory Update (lines 27–30): the edge server evaluates the discrepancy between the original reasoning trace ThoughtThought and the observed OutcomeOutcome, producing a ReflectionReflection. The complete episode (t,Waypoint,Outcome,Reflection)(O_t,Waypoint,Outcome,Reflection) is then stored in memory ℳM for retrieval in subsequent iterations, and the UAV’s current pose is appended to the trajectory log logT_log. The step counter is incremented (line 31) and the loop repeats until TmaxT_ steps have elapsed, at which point the complete trajectory log logT_log is returned. Algorithm 1 Agentic LAPF Framework 1: UAVScene dataset SD_\!S, mission G, UAV U, edge server ℰE, system prompt sysP_sys, maximum steps TmaxT_ 2: Navigation trajectory logT_log 3: Initialize memory ℳ←∅M← 4: Initialize trajectory log log←∅T_log← 5: step←0step← 0 6: while step<Tmaxstep<T_ do 7: Perception 8: (ℐt,t,ℒt)←S[step](I_t,P_t,L_t) _\!S[step] 9: σt←summarize(t) _t (P_t); t←ℐt,σt,ℒtO_t←\I_t, _t,L_t\ 10: st←encode_to_text(σt,ℒt)s_t \_to\_text( _t,L_t) ⊳ ℐtI_t kept as image 11: Memory 12: Mt←ℳ.retrieve_recent(k=3)M_t .retrieve\_recent(k=3) 13: ϕt←feedback(log) _t (T_log) ⊳ trailing window 14: Planning 15: Ctx←,st,Mt,ϕtCtx←\G,s_t,M_t, _t\ 16: (Thought,Waypoint,ToolCalls)←ℰ.Agent(sys,ℐt,Ctx)(Thought,Waypoint,ToolCalls) .Agent(P_sys,I_t,Ctx) 17: Action 18: h←[ct≤dsafe]h 1[\,c_t≤ d_safe\,] ⊳ guaranteed invocation 19: if h=1h=1 then 20: Waypoint←ℛ(wpt−1)Waypoint (wp_t-1) ⊳ Eq. (6) 21: else 22: Waypoint←(Waypoint)Waypoint (Waypoint) ⊳ Eq. (7) 23: end if 24: if WaypointAdjust∈ToolCalls\,∈ ToolCalls then 25: Waypoint←ℋ(Waypoint)Waypoint (Waypoint) ⊳ Eq. (8) 26: end if 27: U.execute(Waypoint)U.execute(Waypoint) 28: Outcome←U.get_feedback()Outcome← U.get\_feedback() 29: Memory Update 30: Reflection←ℰ.Reflect(Thought,Outcome)Reflection .Reflect(Thought,Outcome) 31: ℳ.store(t,Waypoint,Outcome,Reflection)M.store(O_t,Waypoint,Outcome,Reflection) 32: log.append(U.pose())T_log.append(U.pose()) 33: step←step+1step← step+1 34: end whilereturn logT_log LAPF Prompt Template System Prompt You are an expert LAPF UAV navigation agent. Use CoT reasoning for safe and efficient path planning while minimizing path length. Mission Mission=mission Goal=(gx,gy,gz) Current=(x,y,z) Image input ⟨ ⟩ ℐtI_t (1024×768) --- passed natively to the VLM backbone as a separate content block, not as text Symbolic State Pose=(x,y,z), RTK=(lat,lon,alt), LiDAR=n_pointspts, Clearance=cm, range=[rmin,rmax], lateral=[bmin,bmax] Retrieved Memory Waypoint=wp; Outcome=outcome; Reflection=reflection; k most recent episodes (k=3k=3) Performance Error=X m; Trend=degrading; Efficiency=Y%; Large deviation → choose a more direct waypoint. Reasoning 1) Assess current state; 2) Retrieve memory; 3) Plan safe & efficient waypoint; 4) Select required tools. JSON Output "situation_analysis":"...", "thought":"...", "waypoint":[x,y,z], "tool_calls":["Tool1","Tool2"] Figure 4: Multimodal prompt supplied to the VLM backbone at each planning cycle. The RGB frame is passed natively as an image block; all remaining fields are text, instantiated at runtime from σt _t, ℒtL_t, the mission goal, the retrieved episodic memory MtM_t, and the feedback signal ϕt _t. The returned JSON is parsed into Thought, Waypoint, and ToolCalls by the Planning module. To ground the agent’s reasoning in a machine-parseable format at every decision step, LAPF employs a fixed structured template rather than free-form querying. Fig. 4 illustrates this template, comprising seven input fields and a JSON output format. A system prompt fixes the agent’s role and CoT objective. A mission field gives the goal coordinates and current position. The RGB frame is supplied as a separate image block. A symbolic state field carries the pose, the RTK geolocation, and the summarized LiDAR return, namely the point count, the clearance ctc_t, and the range and lateral bounds. A retrieved-memory field lists the k most recent episodes. A performance field reports the current path-length error, trend, and efficiency score. A reasoning field instructs the agent to assess the situation, consult memory, plan a candidate waypoint, and select any required tools. To allow the Planning module to parse Thought, Waypoint and ToolCalls deterministically at every step, the template constrains the output to a single raw JSON object with no surrounding text. The reasoning trace is thereby retained as the thought field of the structured response rather than as unconstrained free text, so that CoT content remains available for the reflection step without compromising machine-parseability. Table IV: LAPF implementation hyperparameters Parameter Value LLM inference Backbone Qwen2-VL-7B-Instruct (open-weight VLM) Serving backend mlx-vlm, locally hosted Image input resolution 1024×768 (resampled from 2448×20482448× 2048) Temperature 0.7 Max tokens (planning) 8192 Max tokens (reflection) 150 Call timeout (planning) 600 s (LAPF), 120 s (baselines); scaled to context length Mission configuration Decision horizon TmaxT_ 110 steps Dataset frame skip 20 Frames in evaluated subset 41 Independent trials per configuration 3 Memory & feedback Retrieved past experiences k 3 Feedback trend window 5 steps Critical-deviation threshold >10.0>10.0 m Warning-deviation threshold >5.0>5.0 m High-detour-ratio threshold >1.5×>1.5× Low-efficiency threshold <60%<60\% High-efficiency threshold >85%>85\% Safety & action Clearance threshold dsafed_safe 5.05.0 m Near-goal feedback trigger <2.0<2.0 m Near-goal step bound δmax _ 2.02.0 m Replan contraction κ 0.50.5 Reliability Max retries per LLM call 20 Retry backoff min(2attempt,15) (2^attempt,15) s VI Numerical Results and Discussions This section evaluates the performance of the proposed LAPF framework through numerical experiments on the UAVScenes dataset. We first describe the implementation settings and baseline methods, and then present comparative results under different navigation scenarios to assess path efficiency, safety, and overall navigation performance. VI-A Implementation of LAPF LAPF is implemented in Python and evaluated on a MacBook Air equipped with an Apple M5 processor and 16 GB unified memory under macOS. The framework uses Qwen2-VL-7B-Instruct as the open-weight vision-language backbone, served locally through ’mlx-vlm’, enabling the complete evaluation to run on a single laptop without relying on proprietary large-scale models. RGB inputs are resized from their original 2448×2048 resolution to 1024×768 before inference to control the visual token budget. The structured prompt template enforces JSON-based outputs, with invalid responses handled through retries. A geometric fallback mechanism is monitored through the llm_fallback_steps log; across the two scenarios and three trials, LAPF executes 2×3×110=6602× 3× 110=660 decision steps in total, over which this fallback was never activated. The evaluation uses three independent trials per configuration with a maximum horizon of 110 steps on the AMtown03 subset containing 41 frames. Since the decision horizon Tmax=110T_ =110 exceeds the 4141-frame observation subset, the recorded stream is cycled once exhausted, so that every method is scored over an identical 110110-step budget irrespective of sequence length. The remaining implementation parameters, including inference settings, memory configuration, feedback thresholds, and safety-related parameters, are summarized in Table IV. All methods follow the same observation schedule, ensuring a fair comparison where performance differences arise from decision-making strategies rather than input variations. VI-B Baselines Description Pure LLM represents a non-agentic baseline that directly generates navigation decisions from the input state without external feedback, memory, or iterative action correction. CoT LLM represents a reasoning-enhanced but still non-agentic baseline, in which the model is prompted to perform step-by-step reasoning before generating navigation actions. All three methods are driven by the same Qwen2-VL-7B backbone under identical decoding settings. Crucially, all three receive the identical multimodal observation: the RGB frame ℐtI_t is passed natively as an image input in every configuration, and the symbolic fields derived from σt _t and ℒtL_t are byte-identical across the three prompts, as is the output schema. All three are additionally evaluated under the same near-goal stability operator C of (7) and the same hold-position operator ℋH of (8), which are applied uniformly rather than only to the proposed method. VI-C Open-Field Navigation (AMtown03) Figure 5: Open-field scenario (AMtown03, no injected obstacles). Panels (a) and (b) show path length and efficiency (mean ± SD) across three trials per method, relative to the 497.33497.33 m straight-line optimum shown as a dashed red line. Panel (c) shows mean clamp events (near-goal step truncations). LAPF converges to near-optimal path length and efficiency with zero clamp events, while CoT LLM shows the largest deviation from optimal and the only nonzero clamp-event rate even in the absence of obstacles. Figure 6: Obstacle-injected scenario (AMtown03, five injected clearance violations, four below dsafe=5d_safe=5 m). Panels (a) and (b) show path length and efficiency (mean ± SD) across three trials per method, relative to the 497.33497.33 m straight-line optimum shown as a dashed red line. Panel (c) shows mean unsafe-event and clamp-event counts. LAPF converges to near-optimal path length and efficiency while incurring zero clamp events, whereas CoT LLM shows the largest deviation and the highest clamp-event frequency, consistent with less regulated step proposals under obstacle constraints. Fig. 5 summarizes path length (panel a), efficiency (panel b), and clamp events (panel c) relative to the theoretical straight-line optimum (dashed red line). This scenario contains no injected obstacles, so ht=0h_t=0 at every step and Nh=0N_h=0 in all trials: by (9) every executed waypoint here is the model-authored proposal under the same operator C applied to the baselines. As shown in Fig. 5(a), LAPF achieves the shortest mean path length (512.83 ± 26.78 m), closely tracking the optimal-path reference line, with T1 (497.33497.33\,m) and T2 (497.40497.40\,m) nearly coincident with it. Since Nh=0N_h=0, the coincidence of T1 with L⋆L reflects a collinear command sequence issued by the policy itself, not a geometric fallback. Pure LLM yields a longer mean path length of 556.37 ± 15.02 m, while CoT LLM exhibits the largest deviation from the optimal trajectory, averaging 619.16 ± 29.01 m. LAPF operates through a closed-loop agent architecture that enables the LLM to perceive outcomes, maintain memory of previous actions, and iteratively refine future decisions. The reduced path length therefore reflects not only improved reasoning, but also the ability of the LLM agent to ground its decisions in feedback and continuously correct its behavior. This trend is further supported by Fig. 5(b), where LAPF achieves the highest mean efficiency (97.05 ± 5.09%) and the lowest mean detour ratio (1.032 ± 0.056), approaching the 100% efficiency reference. In comparison, Pure LLM achieves intermediate performance (89.43 ± 2.45% efficiency, with a 1.119 detour ratio), while CoT LLM performs worst (80.37 ± 3.85% efficiency, 1.246 detour ratio). The observed advantage is consistent with interaction between reasoning and externalized control mechanisms, where observations, memory, and corrective actions allow the system to adapt its behavior rather than relying on prompting alone. Fig. 5(c) provides additional evidence of this distinction. Both Pure LLM and LAPF achieve zero clamp events across all trials, indicating stable waypoint generation and appropriate step-size regulation. In contrast, CoT LLM produces an average of 9.7 clamp events despite operating in an obstacle-free environment. Without an execution-feedback loop, the model may generate actions that are internally plausible but poorly calibrated. LAPF avoids this issue by coupling LLM decision-making with explicit action constraints and feedback, allowing the agent to better regulate waypoint execution. The near-optimal values in panels (a) and (b) warrant explicit interpretation. By (10), ηt _t attains unity exactly when the commanded sequence is collinear with, and monotone along, the start–goal segment, so a value of 100%100\% is admissible by construction and is not an artefact of the measurement. It does, however, indicate that the metric rewards geometric directness without separately certifying per-step kinematic realizability, since (9) constrains step magnitude only in the terminal phase through C. Path efficiency is therefore reported here as a measure of trajectory directness relative to the straight-line optimum, and is complemented by the clamp count NcN_c of (11), which is sensitive precisely to unregulated step magnitudes. Introducing a global displacement bound ∥wpt−wpt−1∥≤Δmax _t-wp_t-1 ≤ _ applied uniformly to all methods, so that directness and kinematic feasibility are scored jointly, is identified as future work. The ability to maintain memory, observe execution outcomes, enforce constraints, and revise actions enables LAPF to produce more efficient and stable trajectories. These results suggest that effective LLM-assisted autonomy depends not only on model reasoning but also on the surrounding agent architecture that converts reasoning into adaptive behavior. VI-D Obstacle-Injected Navigation (AMtown03) Fig. 6 extends the open-field evaluation by injecting five obstacles into the environment, and compares Pure LLM, CoT LLM, and LAPF across three trials (T1–T3) relative to the same 497.33 m straight-line optimum. Obstacles are injected by overriding the sensed clearance ctc_t at five pre-selected decision steps, identical across methods and trials. Four of these are set below dsafe=5d_safe=5 m and therefore register as hazards. Injection is applied to the observation stream rather than to the commanded state, which preserves the common-observation control and accounts for the identical unsafe-event count in panel (c). Controlled injection is also what exercises the enforcement pathway at all, since the recorded sequence is flown as a nadir survey and ctc_t remains above 1515 m in every frame, so it never approaches dsafe=5d_safe=5 m at any point of the flight. The results below therefore establish mechanism correctness under a controlled stimulus. Evaluation against naturally occurring hazards requires a lower-clearance sequence or a forward-facing sensor and is identified as future work. Fig. 6 further demonstrates the importance of agentic capabilities when an LLM-assisted navigation system is exposed to increased environmental complexity. The obstacle-injected scenario introduces a setting where a single-pass generation strategy is insufficient: the system must not only produce a trajectory but also incorporate environmental constraints, regulate actions, and adapt its decisions based on feedback. As shown in Fig. 6(a), LAPF achieves the shortest mean path length (506.37 ± 16.68 m), remaining closest to the optimal-path reference line despite the presence of additional obstacles. In comparison, Pure LLM achieves 581.51 ± 35.20 m, while CoT LLM reaches 599.89 ± 37.20 m, showing the largest deviation from the optimum. Although all methods experience some increase in path length compared with the open-field condition due to obstacle avoidance requirements, LAPF exhibits almost no degradation. This robustness is consistent with the agentic structure enabling the LLM to integrate observations, maintain state information, and revise its trajectory when environmental constraints change, rather than relying on a fixed pre-generated plan. The same trend appears in Fig. 6(b), where LAPF maintains an efficiency of 98.09 ± 3.17% and a detour ratio of 1.020 ± 0.034, closely matching its obstacle-free performance. In contrast, Pure LLM and CoT LLM undergo larger efficiency reductions, achieving 85.73 ± 5.05% and 82.96 ± 5.30%, respectively. These results suggest that additional reasoning depth alone is insufficient for effective obstacle handling, which also requires interaction and adaptation. Although CoT LLM performs additional internal reasoning, it lacks the external feedback mechanism needed to assess whether its generated actions remain appropriate under changing conditions. LAPF’s closed-loop agent architecture grounds reasoning in perception, memory, and action correction, enabling it to preserve high path efficiency even as the environment becomes more challenging. The stability of ηt _t under obstacle injection is a designed property rather than an incidental one. By Property 1 the enforcement pathway is metric-neutral by construction, so the efficiency reported here measures the agentic decision policy and not the shield, and enforcement is instead assessed through NhN_h and the clamp count. This separation is what makes the present comparison interpretable. By Property 1 the hazard-classified steps contribute exactly zero to ηt _t, so the 93.593.5 m and 75.175.1 m advantages of LAPF over CoT LLM and Pure LLM are predominantly attributable to the decision policy, which under an identical backbone, identical decoding settings and an identical observation stream is most consistently explained by the agentic mechanism that distinguishes LAPF, namely the joint conditioning of every proposal on episodic memory and on the performance-feedback signal ϕt _t. Fig. 6(c) shows identical unsafe-event counts across all methods (mean 4.0), corresponding to the same set of injected obstacles below the safety threshold. This identity is intentional and serves as an experimental control: it certifies that the three methods were exposed to the same hazard set, so the divergence observed in the response metrics cannot be ascribed to unequal environmental difficulty. What distinguishes the methods is not how often a hazard is registered, because this is common to all methods, but whether registration is coupled to a prescribed response: LAPF converts all four instances into enforced replans through (15), whereas the prompting-only baselines classify the same four instances with no pathway to act on them. A more meaningful distinction appears in the clamp-event results, where LAPF maintains zero clamp events, Pure LLM shows a small increase (2.3 events), and CoT LLM increases substantially to 14.0 events. This distinction is not an artefact of the exemption of hazard steps from C: the exemption spans at most four steps, whereas the separation from CoT LLM is 14.014.0, and in the open-field scenario Nh=0N_h=0, so C is active on all 110110 steps of every method and Nc=0N_c=0 is attained by LAPF with no exemption whatsoever. The high clamp rate of CoT LLM suggests that additional reasoning does not guarantee improved action regulation. Without an execution-feedback loop, the model may produce logically plausible but poorly calibrated waypoint predictions, particularly when jointly considering goal progression and obstacle constraints. Overall, panels (a)–(c) show that LAPF’s advantage is consistent with its agentic capabilities: perceiving the environment, preserving memory, enforcing action constraints, and iteratively refining decisions. Unlike non-agentic LLM baselines, whose performance becomes less stable as task complexity increases, LAPF maintains efficient, controlled navigation under environmental perturbations. These results suggest that reliable LLM-assisted autonomy depends not only on the model’s reasoning ability but also on the surrounding agent framework that translates reasoning into adaptive, feedback-driven action. VI-E Discussion The results are best interpreted through the closed loop of (4), in which each waypoint is conditioned not on the current observation alone but jointly on the episodic memory MtM_t and the feedback signal ϕt _t. The non-agentic baselines are the special case Mt=∅,ϕt=∅M_t= ,\ _t= ; since all three receive the byte-identical observation and the same operators C and ℋH, the measured gap isolates these two channels rather than reasoning depth or perception quality. Role of memory. Recency-based retrieval over the k=3k=3 most recent episodes, each carrying an executed waypoint, its outcome, and a reflection on their discrepancy, lets the agent condition every proposal on how its own recent commands actually resolved. This is what turns a per-step reaction into a corrected trajectory: in the open-field case it yields near-collinear command sequences (T1 coincident with L⋆L at Nh=0N_h=0), and under obstacle injection it preserves efficiency at 98.09%98.09\% rather than degrading, because past hazard episodes bias subsequent waypoints toward directions that anticipate the shield instead of repeatedly triggering it. Without MtM_t, the baselines regenerate each waypoint from scratch, a mechanism most consistent with their larger detour ratios and, for CoT LLM, the increase in clamp events from 9.79.7 to 14.014.0. Role of tools. Tool use contributes a distinct and complementary effect: it separates what the agent decides from what is guaranteed. The guaranteed-invocation Obstacle Checker of (5) tests clearance at every step independently of ToolstTools_t, so a registered hazard is always coupled to the bounded response of (6), while the discretionary waypoint adjuster of (8) regulates step magnitude near the goal. Fig. 6(c) shows that all methods register the same four unsafe events, yet only LAPF converts each into an enforced replan; by Property 1 this pathway is metric-neutral, so the two effects, safer response and shorter paths, are scored through separate estimators (NhN_h versus ηt _t) and never conflated. Taken together, memory and feedback shorten and stabilize the trajectory, while tool-enforced action guarantees the hazard response; neither alone reproduces the full behavior. This is why additional reasoning depth, as in CoT LLM, is insufficient: without a feedback channel to test whether generated actions remain appropriate, and without an enforced path to act on detected hazards, deeper reasoning yields internally plausible but poorly calibrated waypoints, particularly when goal progression and obstacle constraints must be considered jointly. In the obstacle-injected scenario, LAPF preserves both path efficiency and trajectory quality with minimal degradation: its mean path length changes only from 512.83512.83 m to 506.37506.37 m and its efficiency rises from 97.05%97.05\% to 98.09%98.09\%, with zero clamp events in both, whereas CoT LLM rises from 9.79.7 to 14.014.0 clamp events and Pure LLM from 0.00.0 to 2.32.3. The advantage is therefore not limited to unconstrained navigation but is sustained as complexity increases. LAPF is not intended to replace classical geometric planners when accurate maps and deterministic objectives are available; it targets environments that require semantic interpretation, adaptive replanning, and integration of heterogeneous information sources, providing a higher-level cognitive layer that formulates, revises, and validates navigation strategies. Several limitations remain: the evaluation is confined to UAVScenes and excludes dense urban settings with narrow corridors, tall buildings, and pedestrian interactions; the semantic and depth channels are not yet fused into the reasoning pipeline; and the safety layer operates on a conservative scalar clearance rather than a full 3D collision model. Finally, the contribution of the visual channel is not isolated here: the image is a control variable rather than a treatment, since all three methods receive it identically, and an image-ablated configuration is left to future work. VII Conclusion This paper presents LAPF, an LLM-agent-based framework for autonomous UAV navigation that integrates multimodal perception, episodic memory, planning, and tool-enforced action within a closed-loop architecture. Using observations from UAVScenes, LAPF combines native RGB input with symbolic LiDAR and 6-DoF pose representations through a compact open-weight VLM, enabling edge-scale multimodal navigation on a single consumer laptop. The proposed memory and action mechanisms support adaptive waypoint generation and guarantee safety-critical hazard responses during navigation. Experimental results show that LAPF consistently outperforms conventional prompting and CoT prompting in both open-field and obstacle-injected scenarios, achieving higher path efficiency, shorter trajectories, and improved navigation stability while maintaining zero clamp events under obstacle injection. In our experiments, the most decisive advantage…of the agentic design is the coupling of memory-grounded feedback with tool-enforced action: this pairing, rather than reasoning depth alone, allows LAPF to both shorten trajectories and guarantee a bounded hazard response, and most reliably distinguishes it from prompting-only methods as task complexity grows. Although the current evaluation is limited to a single scene, one backbone, and three trials per configuration, the results demonstrate the effectiveness of integrating memory, reasoning, and tool interaction for autonomous UAV navigation. Future work will extend the framework to larger-scale evaluations, interactive closed-loop simulation, and real-world UAV deployment. References [1] A. H. Arani, X. Fernando, O. Alhussein, and Y. Zhu (2026) Deep reinforcement learning for resource sharing and UAV trajectory optimization in multi-operator UAV-assisted wireless networks. IEEE Transactions on Vehicular Technology (), p. 1–16. Cited by: §I. [2] A. H. Arani, P. Hu, and Y. Zhu (2024) UAV-assisted space-air-ground integrated networks: a technical review of recent learning algorithms. IEEE Open Journal of Vehicular Technology 5 (), p. 1004–1023. External Links: Document Cited by: §I. [3] J. Chen, J. Tang, X. Yang, and Z. Lv (2026) 6GAgentGym: tool use, data synthesis, and agentic learning for network management. arXiv. External Links: Document Cited by: Table I, §I-B. [4] Q. Cheng, Z. Zhang, Y. Du, and Y. Li (2024) Research on particle swarm optimization-based uav path planning technology in urban airspace. Drones 8 (12), p. 701. Cited by: Table I, §I-A. [5] Z. Chkirbene, R. Hamila, A. Gouissem, and U. Devrim (2024) Large language models (llm) in industry: a survey of applications, challenges, and trends. In IEEE 21st International Conference on Smart Communities: Improving Quality of Life using AI, Robotics and IoT (HONET), Vol. , Doha, Qatar, p. 229–234. Cited by: §I. [6] C. Cui, Y. Ma, X. Cao, W. Ye, Y. Zhou, K. Liang, J. Chen, J. Lu, Z. Yang, K. Liao, T. Gao, E. Li, K. Tang, Z. Cao, T. Zhou, A. Liu, X. Yan, S. Mei, J. Cao, Z. Wang, and C. Zheng (2024) A survey on multimodal large language models for autonomous driving. In IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), Vol. , Waikoloa, HI, p. 958–979. Cited by: item 1. [7] X. Dong, H. Zhao, J. Gao, H. Li, X. Ma, Y. Zhou, F. Chen, and J. Liu (2025) Se-vln: a self-evolving vision-language navigation framework based on multimodal large language models. arXiv preprint arXiv:2507.13152. Cited by: Table I, §I-C. [8] H. El Alami and D. B. Rawat (2025) Lightweight language models in autonomous systems: in-context learning with bidirectional alignment for fault detection. In IEEE International Conference on Communications, Montreal, QC, Canada, p. 4395–4400. Cited by: §I. [9] Y. Emami, H. Zhou, M. G. Gaitán, K. Li, L. Almeida, and Z. Han (2026) From prompts to protection: large language model-assisted in-context learning for smart public safety uav. IEEE Wireless Communications (), p. 1–10. Cited by: Table I, §I-C. [10] Y. Emami, H. Zhou, M. Gutiérrez Gaitán, K. Li, and L. Almeida (2026) FRSICL: llm-enabled in-context learning flight resource allocation for fresh data collection in uav-assisted wildfire monitoring. IEEE Internet of Things Journal 13 (10), p. 21613–21622. Cited by: Table I, §I-C. [11] Y. Emami, H. Zhou, S. Nabavirazavi, and L. Almeida (2025) LLM-enabled in-context learning for data collection scheduling in uav-assisted sensor networks. IEEE Internet of Things Journal 12 (23), p. 51664–51676. Cited by: Table I, §I-C. [12] Y. Emami, H. Zhou, R. Reddy, A. H. Arani, B. Wang, K. Li, L. Almeida, and Z. Han (2026) Large language model-assisted uav operations and communications: a multifaceted survey and tutorial. arXiv preprint arXiv:2602.19534. Cited by: Table I, §I, §I-C. [13] M. A. Ferrag, A. Lakas, N. Tihanyi, and M. Debbah (2026) LLM and ai agents for autonomous systems: a survey of applications, datasets, and security challenges. IEEE Open Journal of Intelligent Transportation Systems 7, p. 615–657. Cited by: item 3, §I. [14] M. A. Ferrag, N. Tihanyi, and M. Debbah (2026) From llm reasoning to autonomous ai agents: a comprehensive review. IEEE Access 14 (), p. 84237–84285. Cited by: §I. [15] X. Hu, B. Pang, F. Dai, and K. H. Low (2020) Risk assessment model for uav cost-effective path planning in urban environments. IEEE Access 8 (), p. 150162–150173. Cited by: Table I, §I-A. [16] Y. Iwashina, M. Kunibe, S. Kato, and H. Shigeno (2022) UAV path planning in urban environments with dynamic risk-map generation by vehicle and pedestrian perception. In IEEE 96th Vehicular Technology Conference (VTC2022-Fall), Vol. , London, United Kingdom, p. 1–5. Cited by: Table I, §I-A. [17] F. Jiang, C. Pan, K. Wang, P. Michiardi, O. A. Dobre, and M. Debbah (2026) From large ai models to agentic ai: a tutorial on future intelligent communications. IEEE Journal on Selected Areas in Communications. Cited by: §I. [18] T. Jingwen, G. Wei, S. Jiawei, W. Qiong, L. Zijian, L. Zehong, and Z. Jun (2026) WirelessAgent: large language model agents for intelligent wireless networks. China Communications 23 (3), p. 265–285. External Links: ISSN 1673-5447, Document Cited by: Table I, §I, §I-B. [19] Q. Kuang, J. Wu, J. Pan, and B. Zhou (2020) Real-time uav path planning for autonomous urban scene reconstruction. In IEEE International Conference on Robotics and Automation (ICRA), Vol. , p. 1156–1162. External Links: Document Cited by: Table I, §I-A. [20] H. Kurunathan, H. Huang, K. Li, W. Ni, and E. Hossain (2024) Machine learning-aided operations and communications of unmanned aerial vehicles: a contemporary survey. IEEE Communications Surveys & Tutorials 26 (1), p. 496–533. Cited by: §I. [21] D. Lee and H. Kim (2026) A survey on llm agents: architecture, applications, and challenges. In 40th International Conference on Information Networking (ICOIN), Vol. , Hanoi, Vietnam, p. 979–982. Cited by: item 2. [22] M. A. K. Raiaan, M. S. H. Mukta, K. Fatema, N. M. Fahad, S. Sakib, M. M. J. Mim, J. Ahmad, M. E. Ali, and S. Azam (2024) A review on large language models: architectures, applications, taxonomies, open issues and challenges. IEEE access 12, p. 26839–26874. Cited by: §I. [23] C. Sui, Q. Liu, J. Wang, R. Zhang, and Y. Wang (2025) UAV path planning in urban environments based on rrt algorithm. In 3rd International Conference on Signal Processing and Intelligent Computing (SPIC), Vol. , p. 404–407. External Links: Document Cited by: Table I, §I-A. [24] X. Sun, W. Si, W. Ni, Y. Li, D. Wu, F. Xie, R. Guan, H. Xu, H. Ding, Y. Wu, et al. (2026) Autofly: vision-language-action model for uav autonomous navigation in the wild. arXiv preprint arXiv:2602.09657. Cited by: Table I, §I-C. [25] J. Tong, F. Liu, L. Xv, S. Lu, K. Li, Y. Zhang, Y. Song, Z. Xue, and J. Zhang (2026) WirelessBench: a tolerance-aware llm agent benchmark for wireless network intelligence. arXiv. External Links: Document Cited by: Table I, §I-B. [26] P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y. Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin (2024) Qwen2-vl: enhancing vision-language model’s perception of the world at any resolution. External Links: 2409.12191, Link Cited by: 2nd item. [27] S. Wang, S. Li, Y. Zhang, S. Yu, S. Yuan, R. She, Q. Guo, J. Zheng, O. K. Howe, L. Chandra, et al. (2025) UAVScenes: a multi-modal dataset for uavs. arXiv preprint arXiv:2507.22412. Cited by: §I. [28] J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. (2022) Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, p. 24824–24837. Cited by: §I. [29] G. Yu, I. Butler, and A. Collins (2026) WirelessLLM-agent: a unified llm-based agent framework for multi-task wireless communication decision-making. Preprints. External Links: Document Cited by: Table I, §I-B. [30] G. Zhang and L. Hsu (2019) A new path planning algorithm using a gnss localization error map for uavs in an urban area. Journal of Intelligent & Robotic Systems 94 (1), p. 219–235. Cited by: Table I, §I-A. [31] L. Zhang, W. Chen, J. Chang, and Y. Xue (2025) A survey on navigation technologies of unmanned aerial vehicle swarm. In International Conference of Mechanical Engineering on Aerospace (CoMEA), Vol. , p. 01–20. External Links: Document Cited by: §I. [32] H. Zhou, C. Hu, D. Yuan, Y. Yuan, D. Wu, X. Chen, H. Tabassum, and X. Liu (2024) Large language models (llms) for wireless networks: an overview from the prompt engineering perspective. arXiv preprint arXiv:2411.04136. Cited by: §I. [33] J. Zhou, S. Wang, Z. Yang, Z. Yu, and T. Li (2026) FreeFly-thinking: aligning chain-of-thought reasoning with continuous uav navigation. arXiv preprint arXiv:2603.07181. Cited by: Table I, §I-C.