As a typical application of the low-altitude economy, UAV collaborative monitoring contributes to urban management and data collection. The dense distribution of urban buildings leads to limited perception and communication constraints, making it challenging for multi-UAV systems to achieve effective unified situational awareness. Therefore, this paper constructs a dynamic event-triggered communication strategy based on a large language model (LLM) and achieves collaborative control of multiple UAVs through deep reinforcement learning. Firstly, a LLM is used to analyze the environment code to actively understand the state of the UAV and the characteristics of the environment, effectively improving adaptability to complex environments. Then, LLM uses Python code to generate communication trigger conditions through semantic reasoning and makes dynamic adjustments to optimize communication timing and reduce network congestion and resource consumption. In addition, this paper proposes a coder-evaluator framework to solve the problems of high cost, low efficiency and intense subjectivity of the reinforcement learning human feedback (RLHF) method in LLM. In this way, the executable code generated by LLM is optimized to ensure the robustness and efficiency of the communication mechanism. Experiments in the high-fidelity AirSim simulation environment demonstrate significant improvements compared to baseline methods: a 26.58% decrease in communication cost, a 4.92% increase in task success rate, and a 21.43% acceleration in convergence speed.

LiDAR placement and field of view selection play a role in detecting the relative position and pose of vehicles in relocation maps based on high-precision map automatic navigation. When the LiDAR field of view is obscured or the LiDAR position is misplaced, this can easily lead to loss of repositioning or low repositioning accuracy. In this paper, a method of LiDAR layout and field of view selection based on high-precision map normal distribution transformation (NDT) relocation is proposed to solve the problem of large NDT relocation error and position loss when the occlusion field of view is too large. To simulate the real placement environment and the LiDAR obstructed by obstacles, the ROI algorithm is used to cut LiDAR point clouds and to obtain LiDAR point cloud data of different sizes. The cut point cloud data is first downsampled and then relocated. The downsampling points for NDT relocation are recorded as valid matching points. The direction and angle settings of the LiDAR point cloud data are optimized using RMSE values and valid matching points. The results show that in the urban scene with complex road conditions, there are more front and rear matching points than left and right matching points within the unit angle. The more matching points of the NDT relocation algorithm there are, the higher the relocation accuracy. Increasing the front and rear LiDAR field of view prevents the loss of repositioning. The relocation accuracy can be improved by increasing the left and right LiDAR field of view.

Normal Distribution Transform (NDT) algorithm plays the role of detecting the relative pose of the vehicle in high-precision map. In this paper, a method of setting downsampling factor and grid factor for NDT relocation algorithm in dynamic environment is proposed, which can solve the problems of excessive NDT relocation error and location loss caused by dynamic objects accounting for 1 % to 35 % of the volume of scanning point cloud in vehicle environment. To simulate a real dynamic point cloud environment, the single-frame LiDAR point cloud space is voxelized into a mesh. Each grid is assigned a random number evenly distributed between 0 and 1. The threshold value for whether to add a Gaussian noise point is also set. Seven representative dynamic objects on the highway are selected. The threshold value of probability distribution function of Gaussian noise object needs to be set. Then the volume content of the dynamic object in the single frame point cloud space is calculated according to the set threshold value by using the definite integral. By changing the content and volume of dynamic obstacles in a dynamic environment, the effects of the downsampling factor and the grid factor on the accuracy of the repositioning trajectory are obtained. The resampling and mesh coefficients are optimized based on the analysis of the repositioning trajectory accuracy. The results show that When the current sampling factor is fixed, the grid factor of the NDT algorithm is proportional to the RMSE factor. When the NDT grid factor is fixed and the down sampling factor is equal to the side length of the obstacle, the NDT relocation accuracy is the highest and reaches the local optimum.

This paper presents a deep reinforcement learning (DRL)-based Multi-Agent Control for Formation and Obstacle Avoidance (MACFOA) algorithm to solve collaborative formation and obstacle avoidance decision-making for unmanned aerial vehicle (UAV) systems in dense obstacle environments. The algorithm primarily addresses the coupled strategy update challenges that emerge from simultaneous obstacle avoidance and formation control. A distributed control framework is employed to enable efficient formation and obstacle avoidance for UAV swarms in densely obstructed environments. To address the inefficiency and instability of strategy update problems due to sparse data samples in DRL formation control algorithms, an enhanced multi-step continuous experience replay mechanism is introduced. This mechanism stores and leverages experience data from multiple consecutive time steps, linking contextual information while fully accounting for temporal dependencies, ensuring continuous dynamic policy optimization throughout the training process. Comparative simulations were carried out to evaluate the performances of the proposed approach in terms of efficiency and flexibility. The results have shown that employing the MACFOA-MULT4 algorithm, which utilizes a four-step experience replay strategy, leads to optimal performance, significantly enhancing both training efficiency and stability. Compared to MACFOA, it reduces root mean square error (RMSE) by 38.56% and improves by 25.17% over the multi-agent recurrent deterministic policy gradient (MADRPG). In dynamic simulations on the AirSim platform, the algorithm demonstrated strong adaptability and stability, especially in high-obstacle-density environments. Its superior performance in control stability and task efficiency highlights the effectiveness and advantages of the proposed control strategy.

In the Internet of Things (IoT)-based collaborative surveillance systems, communication links are frequently disrupted due to uncrewed aerial vehicle (UAV) node failures and rapid changes in network topology. Existing recovery methods based on deep reinforcement learning (DRL) often rely on static reward functions, which struggle to adapt to dynamic environments and thus impair convergence efficiency. Moreover, waypoints generated by standard algorithms often overlook practical communication range constraints. To address these challenges, this article proposes a collaborative communication link reconstruction (CCLR) framework. A reward adaptation mechanism is designed, which employs a large language model (LLM) to convert real-time environmental context into executable Python reward functions. This enables the agent to dynamically shift its optimization objective from broad coverage to urgent link recovery based on current network conditions. Furthermore, a constrained optimization module is integrated into the training loop, enforcing distance constraints among UAVs strictly through gradient penalty. Experimental results demonstrate that the proposed framework effectively balances adaptive decision-making and physical connectivity constraints, achieving significantly faster convergence and higher network coverage compared to baseline methods.

The cooperative control of UAVs in a complex environment is the key and fundamental problem in intelligent transportation and IoT communication systems. The cooperative control policy of the UAVs is challenging to determine due to the non-stationarity of the environment caused by the interactions between the UAVs and the environment under IoT communication. The use of deep learning algorithms based on a digital twin (DT) system to optimize the control policy has shown significant benefits in terms of efficiency and flexibility. However, the DT is prone to being inconsistent with real-world applications due to sensor errors and data delays, which lead to incorrect policy, even if the learning process is convergent. To address this issue, this article proposes a federated deep reinforcement learning solution to the multi-UAV intelligent control problem, with a newly developed constrained optimization framework. To effectively eliminate the negative impact of uncertainties such as sensor errors and data delays on system performance, the safety constraint function is integrated into the objective function for optimization. The model undergoes training through the constrained distributed federated reinforcement learning framework, avoiding potentially dangerous behavior and thereby improving the safety and robustness of the resulting control strategies. Experimental results have shown that the proposed solution ensures the safety of the strategy, effectively avoids potential dangerous behaviors, and improves the success rate of multi-UAV cooperative control tasks as well as the average task completion time under IoT communication. Experimental results demonstrate that the proposed solution achieves substantial performance improvements while ensuring operational safety. Compared with baseline methods in IoT environments, the algorithm exhibits a minimum of 26.7% faster convergence, a minimum of 1.8% higher success rate across swarm sizes, and a minimum of 12.3% reduction in task completion time.