Reinforcement learning has achieved significant progress in UAV autonomous navigation. However, existing methods typically rely on either purely discrete or purely continuous action spaces to define the UAV's maneuver mode. Discrete action space is simple to implement and converge quickly but lack sufficient control granularity. In contrast, continuous action space provides higher control resolution but often leads to inefficient training and susceptibility to local optima. Existing RL methods cannot adaptively switch maneuver modes within a unified framework because discrete and continuous action spaces differ fundamentally in structure and control objectives. To address this issue, we propose a hierarchical reinforcement learning framework with hybrid action space (HAS-HRL). Specifically, the high-level policy adaptively selects the maneuver mode according to the environment context, while the low-level policy consists of a set of primitive navigation skills associated with the hybrid maneuver modes. These skills generate executable control commands, enabling the UAV to perform smooth maneuvers in dense obstacle regions while cruising efficiently in open spaces. Furthermore, an event-triggered control rule is introduced to provide structured prior guidance during the early training stage, thereby improving exploration efficiency and convergence stability. Experiments in various simulation environments demonstrate that the proposed HAS-HRL framework consistently outperforms single-layer RL and HRL baselines in terms of success rate, obstacle-avoidance performance, and training stability. The results show that the hybrid maneuver modes effectively balance flight safety and navigation efficiency, offering a robust and efficient solution for UAV autonomous navigation in complex scenarios.

Autonomous navigation in unknown environments remains a critical challenge for the small unmanned aerial vehicle. Recently, reinforcement-learning methods have been widely employed to improve flight efficiency and obstacle-avoidance performance. However, most existing methods rely on implicit end-to-end decision mappings and homogeneous action representations, which makes it difficult to explicitly characterize maneuver intent and coordinate heterogeneous maneuver requirements, such as fast cruising in open regions and smooth obstacle avoidance in cluttered regions. This limitation reduces the adaptability and reliability of learned navigation policies in constrained airspace. To address this problem, this paper proposes an intent-guided hierarchical reinforcement learning framework with hybrid maneuver modes, HMM-IGHRL. The framework consists of a high-level intent policy and two low-level maneuver execution policies. The high-level policy generates a composite intent consisting of a maneuver mode and a local sub-goal, establishing an explicit link between strategic navigation intent and local motion execution. Conditioned on the selected maneuver mode, the framework invokes the corresponding low-level policy, enabling the UAV to switch adaptively between aggressive maneuvering with discrete motion primitives and smooth maneuvering with continuous control commands. Furthermore, an intent-driven and risk-predictive event-triggered mechanism evaluates collision risk, path-intent feasibility, and maneuver-mode environmental compatibility to refresh the high-level intent only when necessary, thereby reducing redundant decisions while maintaining timely responsiveness to local environmental changes. Experiments in multiple constrained navigation environments, together with ablation studies, maneuver-mode selection statistics, and parameter sensitivity analyses, show that the proposed method achieves a higher task success rate, improved collision-avoidance performance, and higher execution efficiency than representative flat and hierarchical reinforcement-learning baselines. High-fidelity AirSim simulations further demonstrate the execution feasibility of the proposed framework at the navigation decision-making layer in complex aerial scenarios.

Unmanned Aerial Vehicle (UAV) technology is playing an increasingly vital role in the low-altitude economy, while intelligent task allocation has become a key factor in improving operational efficiency in logistics and emergency response. However, practical applications face three interrelated challenges, i.e., the random emergence of tasks, the heterogeneity of task skill requirements, and the secure audit trail of the AI-generated task allocation process. To address these challenges, this paper proposes a proactive multi-agent allocator (ProMAA), an AI-assisted blockchain framework for dynamic UAV task allocation. Specifically, we first propose a spatio-temporal skill Mamba (STSMamba) model to forecast future task locations, types, and arrival times. Then, Auxiliary Rewards Multi-Agent Reinforcement Learning allocator (ARMARL) leverages these predictions and auxiliary rewards to produce structured task, UAV assignments that respect skill compatibility. Finally, we design a model trace merkle tree (MTM-tree), a blockchain-optimized data structure that exploits the UAV-centric and time-ordered nature of AI outputs to enable efficient storage, batched updates, and authenticated queries. Experimental results on datasets demonstrate that ProMAA achieves up to a 99.63% task completion rate and a 99.1% response rate, significantly outperforming the state-of-the-art baseline. Furthermore, the proposed MTM-tree maintains sub-millisecond range query times and sustains a peak throughput of 215 QPS, effectively resolving the overhead of traditional blockchain storage.

Unmanned Aerial Vehicles (UAVs) are increasingly deployed as mobile nodes in intelligent transportation systems (ITS), facilitating traffic monitoring, emergency response, and urban air mobility. These applications demand precise modeling of UAV behavior and energy efficiency under diverse real-world traffic and environmental conditions. However, existing datasets often overlook crucial aspects such as UAV flight dynamics and multi-UAV collaboration in complex airspace. To bridge this gap, we introduce AMOVFLY, a comprehensive UAV flight status dataset captured from multiple AMOVLAB UAV platforms. AMOVFLY offers substantial improvements over existing datasets: agile heterogeneous devices, comprehensive and extended flight records, diverse flight conditions (including fixed and dynamic altitude/speed profiles), and synchronized multi-UAV flight data under uniform environmental conditions. These features provide a rich foundation for developing predictive models such as Temporal Convolutional Networks (TCN) to support energy consumption forecasting, multi-agent coordination, and adaptive flight control—key challenges in transportation-oriented UAV deployments, such as aerial logistics and traffic surveillance. Our evaluation further quantifies the fragility of static models against real-world dynamics. This finding underscores the critical need for online updates and robust trajectory planning to ensure reliable energy prediction and mission safety in dynamic ITS environments. AMOVFLY thus opens new avenues for improving UAV mission planning, energy prediction, and time-series forecasting in complex, real-world transportation environments. Dataset page can be found at https://github.com/YujiaoHu/AMOVFLY-Dataset

The deployment of deep learning algorithms has significantly enhanced swarm intelligence. However, pre-trained models often perform poorly in dynamic and unpredictable environments. While online training with distributed updates offers a promising solution, the heterogeneity of computing resources in swarm systems poses substantial challenges for efficient model updating. To address this, we propose a Distributed Training Scheduler for Heterogeneous Swarms (DTSHS), a reinforcement learning-based scheduling framework that dynamically balances workload across devices. DTSHS aims to optimize the trade-off between training time and energy consumption by learning adaptive task offloading strategies. Experimental results demonstrate that DTSHS outperforms traditional methods in both time and energy efficiency.

Modular methods, which decompose complex joint policies into function-specific sub-policies, have been widely adopted to enhance asymptotic performance in single-task cooperative multi-agent reinforcement learning (MARL). However, modular policies trained on source tasks often struggle to generalize to unseen scenarios due to variations across tasks, such as mismatched action spaces and divergent state dynamics. To address this challenge, we propose Mutual Information-Guided Subtask Selection(MIGSS), a novel framework that enhances zero-shot generalization in MARL through two key innovations: a Discriminative Group Trajectory Encoder and Global Attention-Driven Coordination. Specifically, the Discriminative Group Trajectory Encoder remaps agent trajectories by maximizing mutual information between agent trajectories and dynamically assigned groups. This optimizes cross-task consistent group trajectory with broader embedding distributions. This encourages agents in distinct states to select specialized subtasks, effectively promoting functional modularity. Meanwhile, the Global Attention-Driven Coordination employs a global attention mechanism to integrate state information, coordinating group trajectories for expressive credit assignment. Extensive experiments in StarCraft II cooperative scenarios demonstrate that MIGSS significantly outperforms superior zero-shot generalization baselines in both single-task and multi-task settings.Visualization analyses confirm that the learned group trajectories successfully disperse agent trajectories into a consistent and broader embedding space, thereby enhancing subtask modularization.

The integration of deep neural network (DNN) intelligence into embedded mobile devices is expanding rapidly, supporting a wide range of applications. DNN compression techniques, which adapt models to resource-constrained mobile environments, often force a trade-off between efficiency and accuracy. Distributed DNN inference, leveraging multiple mobile devices, emerges as a promising alternative to enhance inference efficiency without compromising accuracy. However, effectively decoupling DNN models into fine-grained components for optimal parallel acceleration presents significant challenges. Current partitioning methods, including layer-level and operator or channel-level partitioning, provide only partial solutions and struggle with the heterogeneous nature of DNN compilation frameworks, complicating direct model offloading. In response, we introduce AdaKnife, an adaptive framework for accelerated inference across heterogeneous mobile devices. AdaKnife enables on-demand mixed-granularity DNN partitioning via computational graph analysis, facilitates efficient cross-framework model transitions with operator optimization for offloading, and improves the feasibility of parallel partitioning using a greedy operator parallelism algorithm. Our empirical studies show that AdaKnife achieves a 66.5% reduction in latency compared to baselines.

Unmanned Aerial Vehicles (UAVs) are emerging as a pivotal component in the field of intelligent transportation systems. Leveraging virtual-physical interactions, digital twin technology significantly enhances the adaptability of UAVs in complex traffic environments. However, current approaches still pose three major challenges: contextual adaptability, timely responsiveness, and effective multi-UAV coordination. In this paper, we introduce EnFlexiTwin, a digital twin enhancement assistance platform seamlessly integrated with AdaSor, a lightweight adaptive data selector. EnFlexiTwin automates the construction of incremental learning datasets, enabling real-time enhancements that allow digital twins to adapt to new time series data while preserving historical knowledge. We test EnFlexiTwin on a real-world dataset from low-altitude small-parcel delivery. The results show improved performance and adaptability of digital twins. Furthermore, time-varying simulations on real-world dataset and experiments on a practical air-ground cooperative UAV swarm application highlight that EnFlexiTwin achieves superior enhancements under varying real-time requirements and swarm scale compared to baseline approaches.

In recent years, reinforcement learning has received significant attention and has been widely applied to UAV autonomous navigation tasks. However, most existing studies assume that UAV operates in static environments, overlooking randomly occurring dynamic obstacles. Such obstacles are often difficult for conventional sensors to detect in a timely manner, which poses a serious threat to flight safety. To address autonomous navigation in dynamic environments, this paper introduces the Event-camera, a novel dynamic vision sensor, to capture environmental information with high dynamic range and microsecond-level temporal resolution. To efficiently process the sparse and asynchronous event stream generated by the Event-camera, we develop spiking reinforcement learning framework based on a spiking neural network, enabling low-latency and high-efficiency control and decision-making. Furthermore, inspired by the advances in biological neural dynamics, we propose a biologically plausible plasticity spiking threshold mechanism, which enables spiking neurons to dynamically adjust their firing thresholds in response to the mean membrane potential and depolarization rate. This mechanism enhances the robustness and adaptability of neural information encoding. Extensive experiments in multiple complex environments within the Airsim simulator demonstrate that the proposed method consistently outperforms baseline methods in dynamic environments across various objective evaluation metrics, achieving higher navigation success rates and flight speeds. Moreover, it maintains competitive performance in previously unknown environments, indicating a certain degree of generalization capability.

Multiagent routing problems, arising from practical applications, such as logistics, transportation, and emergency response, face challenges due to the exponential growth of the search space with increasing problem scales. This article proposes RouteMaker to address the often-overlooked multiagent routing problems involving dedicated multiple depots. RouteMaker leverages role-interaction-based graph neural network (RIGNN) to realize effective locations assignments and integrates an advanced planner to plan travel path for each agent. RouteMaker is trained on small-scale problems and can produce comparable or superior approximate optimal solutions compared with the best heuristic baselines. Notably, the learned RouteMaker generalizes seamlessly to large-scale problems and real-world problems without the need for fine-tuning, delivering significantly higher quality solutions in relatively less time. For scenarios involving 40 agents and 1000 locations, RouteMaker achieves over $600\times $ speed improvement and more than 88% cost reduction, compared with the representative classical heuristic solver (ORTools).

A neural processing unit (NPU) is a microprocessor which is specially designed for various types of neural network applications. Because of its high acceleration efficiency and lower power consumption, the airborne embedded system has widely deployed NPU to replace GPU as the new accelerator. Unfortunately, the inherent scheduler of NPU does not consider real-time scheduling. Therefore, it cannot meet real-time requirements of airborne embedded systems. At present, there is less research on the multi-task real-time scheduling of the NPU device. In this article, we first design an NPU resource management framework based on Kubernetes. Then, we propose WAMSPRES, a workload-aware NPU performance model based soft preemptive real-time scheduling method. The proposed workload-aware NPU performance model can accurately predict the remaining execution time of the task when it runs with other tasks concurrently. The soft preemptive real-time scheduling algorithm can provide approximate preemption capability by dynamically adjusting the NPU computing resources of tasks. Finally, we implement a prototype NPU scheduler of the airborne embedded system for the fixed-wing UAV. The proposed models and algorithms are validated on both the simulated and realistic task sets. Experimental results illustrate that WAMSPRES can achieve low prediction error and high scheduling success rate.

UAV-to-Satellite geolocalization aims to estimate the location of an aerial-view query image taken by an Unmanned Aerial Vehicle (UAV) by matching it to satellite images annotated with known locations. However, it is difficult for existing methods to distinguish neighboring satellite images that exhibit a high degree of visual similarity. To address this issue, we introduce a plug-and-play adaptive confusion-aware contrastive loss (ACCL) to explicitly enhance the model’s discriminative ability, which gives more tolerance to high confusion query samples by means of elaborating a confusion metric function. As a plug-and-play loss module, ACCL can be easily incorporated into various UAV-to-Satellite geolocalization methods without additional modifications. To demonstrate the effectiveness of our proposed method, we conduct extensive experiments on one publicly available geolocalization dataset (i.e. NewYorkFly) and to further prove the effectiveness of our method in different scenarios, we collect two new geolocalization datasets (LasVegasFly and HollywoodFly), which contain drone-captured aerial images and dense sampled satellite images in various geomorphic regions. Experimental results indicate that our method can achieve an obvious performance improvement over the state-of-the-art methods on all three datasets. Our code and collected datasets are available at https://github.com/NWPU-CPS/ACCL.

In this paper,a deep deterministic policy gradient algorithm based on Partially Observ-able Weighted Mean Field Reinforcement Learning(PO-WMFRL)framework is designed to solve the problem of path planning in large-scale Unmanned Aerial Vehicle(UAV)swarm operations.We establish a motion control and detection communication model of UAVs.A simulation envi-ronment is carried out with No-Fly Zone(NFZ),the task assembly point is established,and the long-term reward and immediate reward functions are designed for large-scale UAV swarm path planning problem.Considering the combat characteristics of large-scale UAV swarm,we improve the traditional Deep Deterministic Policy Gradient(DDPG)algorithm and propose a Partially Observable Weighted Mean Field Deep Deterministic Policy Gradient(PO-WMFDDPG)algo-rithm.The effectiveness of the PO-WMFDDPG algorithm is verified through simulation,and through the comparative analysis with the DDPG and MFDDPG algorithms,it is verified that the PO-WMFDDPG algorithm has a higher task success rate and convergence speed.

Recently, developing multi-UAVs to cooperatively pursue a fast-moving target has become a research hotspot in the current world. Although deep reinforcement learning (DRL) has made a lot of achievements in the UAV pursuit game, there are still some problems such as high-dimensional parameter space, the ease of falling into local optimization, the long training time, and the low task success rate. To solve the above-mentioned issues, we propose an improved twin delayed deep deterministic policy gradient algorithm combining the genetic algorithm and maximum mean discrepancy method (GM-TD3) for multi-UAV cooperative pursuit of high-speed targets. Firstly, this paper combines GA-based evolutionary strategies with TD3 to generate action networks. Then, in order to avoid local optimization in the algorithm training process, the maximum mean difference (MMD) method is used to increase the diversity of the policy population in the updating process of the population parameters. Finally, by setting the sensitivity weights of the genetic memory buffer of UAV individuals, the mutation operator is improved to enhance the stability of the algorithm. In addition, this paper designs a hybrid reward function to accelerate the convergence speed of training. Through simulation experiments, we have verified that the training efficiency of the improved algorithm has been greatly improved, which can achieve faster convergence; the successful rate of the task has reached 95%, and further validated UAVs can better cooperate to complete the pursuit game task.

Multi-edge cooperative computing that combines constrained resources of multiple edges into a powerful resource pool has the potential to deliver great benefits, such as a tremendous computing power, improved response time, more diversified services. However, the mass heterogeneous resources composition and lack of scheduling strategies make the modeling and cooperating of multi-edge computing system particularly complicated. This paper first proposes a system-level state evaluation model to shield the complex hardware configurations and redefine the different service capabilities at heterogeneous edges. Secondly, an integer linear programming model is designed to cater for optimally dispatching the distributed arriving requests. Finally, a learning-based lightweight real-time scheduler, CoRaiS, is proposed. CoRaiS embeds the real-time states of multi-edge system and requests information, and combines the embeddings with a policy network to schedule the requests, so that the response time of all requests can be minimized. Evaluation results verify that CoRaiS can make a high-quality scheduling decision in real time, and can be generalized to other multi-edge computing system, regardless of system scales. Characteristic validation also demonstrates that CoRaiS successfully learns to balance loads, perceive real-time state and recognize heterogeneity while scheduling.

Communication plays a vital role in multi-agent systems, fostering collaboration and coordination. However, in real-world scenarios where communication is bandwidth-limited, existing multi-agent reinforcement learning (MARL) algorithms often provide agents with a binary choice: either transmitting a fixed number of bytes or no information at all. This limitation hinders the ability to effectively utilize the available bandwidth. To overcome this challenge, we present the Dynamic Size Message Scheduling (DSMS) method, which introduces a finer-grained approach to scheduling by considering the actual size of the information to be exchanged. Our contribution lies in adaptively adjusting message sizes using Fourier transform-based compression techniques, enabling agents to tailor their messages to match the allocated bandwidth while striking a balance between information loss and transmission efficiency. Receiving agents can reliably decompress the messages using the inverse Fourier transform. Experimental results demonstrate that DSMS significantly improves performance in multi-agent cooperative tasks by optimizing the utilization of bandwidth and effectively balancing information value.

The aggregation task is one of the typical tasks of Unmanned A erial Vehicle (UAV) operations and has always been a hot research topic. With the development of technology and the continuous proves of war, multi-UAV have significant advantages in combat compared to single UAV. To address the intelligent control problem of multi-UAV aggregation task, this paper proposes a Twin Delayed Deep Deterministic policy gradient (TD3) algorithm based on an evolutionary algorithm and introduces an evolutionary reinforcement learning framework. A learning cross operator is proposed to make offspring inherit parent features through training. To avoid high interaction costs, a critic network-assisted evaluation mechanism is proposed, and a response control decision is designed to combine real and virtual fitness to improve accuracy. Therefore, a Twin Delayed Deep Deterministic policy gradient algorithm combining Genetic Algorithm enhanced by learning cross factors and auxiliary evaluation (GA-TD3) is proposed. Simulation results show that GA-TD3 has better training effects and generalization.

In autonomous robotic systems, the parallel processing of multiple tasks often competes for limited resources, affecting system performance and the robot's responsiveness to environmental changes. Traditional computational task scheduling methods often overlook the dynamic nature of task priorities in autonomous robotic systems, where task importance can shift based on interactions with the external environment. Therefore, there's a crucial need for a mechanism capable of adaptively adjusting task scheduling in response to environmental changes, ensuring timely access to resources for critical tasks. To address this challenge, this study presents Priorest, a neural network model that incorporates multimodal data processing and multitask learning. Priorest integrates sensor data with logs monitoring computational device performance to predict events influencing task priority, enabling task adjustments while preserving essential resource allocations. When deployed in autonomous robotic systems, Priorest's event-prediction-based adjustment strategy reduced critical task completion times by 18.7%, which demonstrates the effectiveness of Priorest in enhancing parallel task scheduling.

Multi-agent deep reinforcement learning (MDRL) has attracted attention for solving complex tasks. Two main challenges of MDRL are non-stationarity and partial observability from the perspective of agents, impacting the performance of agents’ learning cooperative policies. In this study, Controlled and Targeted Communication with the Centralized Critic (COTAC) is proposed, thereby constructing the paradigm of centralized learning and decentralized execution with partial communication. It is capable of decoupling how the MAS obtains environmental information during training and execution. Specifically, COTAC can make the environment faced by agents to be stationarity in the training phase and learn partial communication to overcome the limitation of partial observability in the execution phase. Based on this, decentralized actors learn controlled and targeted communication and policies optimized by centralized critics during training. As a result, agents comprehensively learn when to communicate during the sending and how to target information aggregation during the receiving. Apart from that, COTAC is evaluated on two multi-agent scenarios with continuous space. Experimental results demonstrated that partial agents with important information choose to send messages and targeted aggregate received information by identifying the relevant important information, which can still have better cooperation performance while reducing the communication traffic of the system.

Unmanned swarm systems are now employed in a wide range of applications. Agriculture, commerce, and the military will all benefit from the advancement of unmanned swarm technology. Scenarios should have a standardized description and definition as system input to support numerous applications and accelerate scenario validation, allowing for more precise and specialized processing to standardize the validation process. A scenario description language for unmanned swarm (SL4U) is introduced in order to implement the unmanned swarm scenario. The scenario is divided into environment and task. The environment is defined, including the task space, target, and home locations, and the initial settings of the unmanned swarm. In terms of tasks, four categories of tasks, four temporal relationship constraints, the event handler, as well as other elements are identified. The dual engines of parsing and micro-embedded collaboration we developed will handle the SL4U. We employ a search and rescue scenario in both simulation and physics to test SL4U. Through the processing of the dual engines, the unmanned swarm’s search and rescue operations in a predefined environment are accomplished successfully. In more complex scenarios, just add more descriptions to SL4U and implement them in dual engines rather than starting from scratch. The use of SL4U in verification minimizes the cost and time consumption of unmanned swarm, particularly in physical verification.