Transformer-Based Double Deep Q-Network for AAV Coverage Path Planning and Data Harvesting

Optimizing 3D UAV navigation via a high-efficiency hybrid RRT*-DQN approach

Path-planning for autonomous unmanned aerial vehicles (UAVs) must be highly efficient, safe, and dynamically feasible for manoeuvring through 3D environments densely packed with obstacles. Traditional heuristic and sampling methods for path-planning often struggle to ensure spatial safety while simultaneously optimizing path quality. They often produce rough and inefficient paths, which lead to instability problems when the vehicle is airborne and higher battery consumption due to longer flight times. We tackle the challenge with this paper by developing a hybrid path-planning method, rapidly-exploring random tree star (RRT*)-deep Q-network (DQN), which combines the fast global search capability of RRT* with the deep learning-based heuristic prediction of a DQN. To demonstrate the merits of our approach, the new system has been tested in four very different 3D space scenarios and its performance compared to seven other leading methods (RRT*, informed RRT*, potential field-based RRT*, heuristic potential optimization RRT*, adaptive grey wolf optimizer, A*, and improved Dijkstra), each method being run 30 times independently with random seeds to allow for a meaningful statistical analysis. The numerical data clearly show that RRT*-DQN outperforms other methods in terms of path optimality almost all the time, with path efficiency reaching up to 97.7% and a trajectory smoothness exceeding 99.86%, while remaining only slightly less competitive in terms of obstacle clearance. Deep reinforcement learning-based inference does bring about a modest computational overhead during the first offline planning stage; however, this is a modest computational overhead, given that it directly produces paths suitable for subsequent flight execution that avoid the kinematically challenging manoeuvres typical of traditional planners. Thus, the RRT*-DQN algorithm offers a very strong candidate for UAVs operating in real-world environments by trading increased offline computational time for improved path quality and potentially more efficient physical execution.

Transformer-Guided Interference-Aware 3D Path Planning for UAV Navigation in Urban Voxel Environments

Urban unmanned aerial vehicle (UAV) navigation may require path planning that accounts for geometric obstacles and spatially varying communication-related risk. This paper presents a transformer-guided interference-aware 3D path-planning method for urban voxel environments. A 3D convolutional neural network (CNN)–transformer network predicts a dense route probability field from occupancy, electromagnetic risk, start–goal, and auxiliary planning channels. The field is restored to the raw-map resolution and used only as a search prior for A* on the original occupancy and risk maps. Obstacle avoidance, endpoint correctness, 6-connected motion (each move reaches one of six face-adjacent voxels, with no diagonal motion), and final path cost evaluation are enforced by graph search rather than by the neural model. On 320 synthetic urban cases covering four map sizes and four building density settings, Guided A* achieves a 27.7× speedup over A* and an 11.9× speedup over Weighted A*, while reducing expanded nodes by 91.2% relative to A*. The mean path cost and electromagnetic cost increase by 2.7% and 5.7%, respectively. Compared with the rapidly exploring random tree (RRT), the method reduces path cost by 10.1% and electromagnetic exposure by 12.0% at similar runtime. A post-training sensitivity study further identifies an empirical balance between route-prior guidance, electromagnetic risk avoidance, route length, and search effort, while the Manhattan weight exhibits the expected heuristic inflation efficiency–quality trade-off. An extended model trained on a larger mixture of procedural and Sionna RT ray-traced data, including real OpenStreetMap building geometry, is further evaluated without retraining on two real-geometry benchmarks, UrbanRadio3D and an OpenStreetMap–Sionna RT suite, where Guided A* retains a 100% success rate and reduces expanded nodes by 97–99% relative to A* while increasing mean path cost by at most 1.7%.

Task‐Oriented Path Planning Towards Autonomous Sensor Network Deployment in Rainforest Canopies

Climate change is placing increasing pressure on rainforests and their ecosystems, creating an urgent need for technologies that enable large‐scale biodiversity monitoring and conservation. In this work, we present an exploration path planner for the autonomous deployment of wireless sensor networks in rainforest canopies using a quadrotor platform. To address the challenge of placing sensors across large areas while constrained by the limited flight endurance of multirotor drones, our approach begins with a scouting mission that identifies feasible canopy locations for sensor placement while accounting for constraints such as communication range and network connectivity. In addition, we incorporate forest canopy morphological constraints by detecting flat, vegetation‐covered areas that can support sensor nodes. Our informative exploration planner combines online detection of suitable canopy surfaces with pointcloud projection and a targeted sampling strategy to guide the drone towards candidate deployment sites. The planner relies on a heuristic gain function based on vegetation area, path length, and smoothness, rather than an explicit optimization of communication connectivity or coverage. Therefore, the proposed method should be interpreted as a task‐oriented exploration strategy that identifies promising deployment regions, rather than a complete solution to the sensor network placement problem. We validate the approach through flight experiments conducted in environments designed to replicate key characteristics of rainforest canopies. The results demonstrate that the proposed method can identify viable candidate deployment locations within complex vegetation structures, providing a foundation for future work on fully integrated autonomous biodiversity monitoring missions.

Deep Reinforcement Learning for Reconnaissance Drone Path Planning with Clustered Targets

In modern military operations, unmanned aerial vehicles have become essential intelligence gathering assets. Military targets are distributed in clustered areas according to tactical formations and command structures, requiring comprehensive coverage across all operational areas. This research addresses the reconnaissance drone path planning problem that maximizes total reconnaissance value while ensuring at least one target per cluster is visited within flight distance constraints. We propose a deep reinforcement learning approach combining a Graph Attention Network(GATv2) encoder with an attention-based decoder for sequential decision-making. The encoder captures spatial relationships and cluster structures through edge attribute modeling, while the decoder generates routing decisions via multi-head attention mechanisms with dynamic constraint masking. Computational experiments across three problem scales demonstrate that the proposed approach achieves near-optimal performance on small instances and significantly outperforms time-limited exact optimization methods on larger scales, achieving solutions in seconds enabling real-time mission planning in dynamic operational environments.

PRI-Net: A Lightweight Multimodal Framework for 3D UAV Localization

Accurate 3D localization of unmanned aerial vehicles (UAVs) remains challenging for existing multimodal approaches due to sparse LiDAR geometry, modality-imbalanced fusion, and redundant feature transmission over constrained edge-to-server links. To address these limitations, we propose PRI-Net, an efficient and lightweight multimodal fusion framework for UAV localization that integrates point cloud splatting, residual attention fusion, and an information bottleneck. Specifically, a 3D point cloud splatting (3DPCS) strategy is introduced to transform sparse LiDAR observations into geometrically consistent dense depth maps. A residual attention fusion (RAF) module is then designed to alleviate modal bias by using an image branch for coarse estimation and a gated fusion branch for refinement. In addition, a multimodal information bottleneck (MIB) module compacts features by filtering task-irrelevant redundancy. Experiments show that PRI-Net achieves high localization accuracy with lightweight architectures, while reducing feature dimensionality and improving edge-to-server UAV sensing efficiency and robustness.

Real-time traffic congestion mapping and bottleneck analysis using an edge-deployable lightweight deep learning framework for UAV-assisted monitoring