In recent years, the rapid advancement of Unmanned Aerial Vehicle (UAV) technology has positioned the UAV swarm system as a critical research focus within the field of multi-agent collaboration. As a smart system capable of carrying multi-modal payloads, the China UAV drone technology has demonstrated revolutionary potential for cross-domain applications over the past decade. From precision agricultural monitoring and urban 3D modeling to military reconnaissance and target tracking, the technological evolution of China UAV drone continues to expand its application boundaries. However, single-unit operations inherently suffer from limitations in execution efficiency and system robustness, making it difficult to meet the demands of diverse tasks. UAV swarms, based on distributed collaborative mechanisms, exhibit significant advantages in large-scale coverage and fault tolerance through real-time information interaction and collective decision-making. Over the past decade, various optimization methods have emerged, covering directions such as collision avoidance, path planning, task allocation, and formation reorganization, laying a solid foundation for the efficiency and practicality of swarm systems.
To clearly understand the research status of China UAV drone swarm cooperative optimization algorithms, this paper provides a comprehensive classification and summary of commonly used optimization algorithms. According to the principles of each algorithm, cooperative optimization algorithms are first divided into traditional heuristic algorithms and machine learning algorithms. Then, the application and improvement of various algorithms by scholars in four directions are introduced separately. Finally, the future development direction of China UAV drone cooperative optimization algorithms is prospected, providing a reliable reference for beginners.
In real-world applications, the urgency and importance of China UAV drone swarm technology are increasingly prominent. For example, in military reconnaissance and battlefield monitoring, multi-UAV collaboration can achieve larger-scale intelligence coverage and more efficient battlefield situational awareness. In disaster rescue scenarios such as earthquakes and floods, UAV swarms can quickly enter complex or dangerous areas to carry out target search, environmental assessment, and material delivery tasks, greatly improving rescue response speed and accuracy. In the civilian sector, logistics distribution and urban management also have urgent needs for UAV swarms. Their multi-point collaboration capability can effectively alleviate the “last mile” distribution bottleneck and show great potential in the emergency response and public safety assurance of smart cities. Thus, in-depth research on China UAV drone cooperative optimization algorithms not only has academic value but also holds significant practical importance in improving the task execution efficiency, resource utilization, and environmental adaptability of swarm systems.
As the formation scale grows exponentially, the system faces several technical challenges. First, real-time collaborative control in dynamic and uncertain environments imposes stringent requirements on algorithm timeliness, robustness, and scalability. Second, limited communication bandwidth and information delays may lead to decision conflicts, necessitating the design of lightweight algorithms to balance computational complexity and collaboration accuracy. Current research mainly focuses on four core modules: collision avoidance, task allocation, path planning, and formation reorganization, aiming to build an autonomous collaborative algorithm system. Collision avoidance not only involves obstacles in the environment but also between adjacent UAVs. This module ensures the safety of group operation through real-time situation awareness and predictive control, serving as a fundamental guarantee for system stability. Path planning needs to comprehensively consider multi-objective parameters such as 3D spatial topology constraints, energy efficiency optimization, and communication link quality. The adaptability of the algorithm determines the feasibility of executing tasks in complex scenarios. As a key to improving formation efficiency, the task allocation mechanism needs to optimize resource scheduling under multiple constraints, and its decision quality directly affects the task completion time and system energy consumption indicators. Formation reorganization strategies, through distributed topology reconstruction algorithms, enhance system fault tolerance, maintaining group integrity even under leader node failure or network topology mutation, significantly improving the system’s survival rate in complex environments.
This paper provides a comprehensive survey and analysis. The structure is as follows. Section 2 introduces the core technical framework of UAV swarms, dividing it into collaborative control architectures and cooperative optimization algorithms. Section 3 discusses specific directions including collision avoidance, path planning, task allocation, and formation reorganization with a detailed review of algorithms. Practical challenges faced by China UAV drone systems are discussed in Section 4. Section 5 explores future research directions. Finally, Section 6 concludes the paper.
The collaborative control architecture is the basic framework for a UAV formation to achieve cooperative actions and task allocation. This architecture defines the information flow, decision-making process, and execution mechanism to ensure the overall performance and reliability of the formation. Based on the control mode, the cooperative control architecture of UAVs can be divided into three major paradigms: centralized, distributed, and hybrid. The centralized architecture collects global status through a ground control station or central node and performs unified calculation and scheduling to achieve globally optimal path planning and task assignment. The distributed architecture relies on neighbor information exchange and consensus protocols among UAVs for autonomous decision-making, thereby improving system robustness and scalability, avoiding single points of failure. The hybrid architecture adopts centralized planning at the high level and distributed response at the low level, balancing global optimization and local flexibility, effectively balancing performance and reliability. To clearly compare the characteristics of different control architectures, Table 1 summarizes the performance of centralized, distributed, and hybrid architectures on key performance indicators.
2.1.1 Centralized Control Architecture
The centralized control architecture usually uses a Ground Control Station (GCS) as the core, collecting the status of all UAVs through wireless links, performing path planning and task allocation on a central server, and then sending control commands to each UAV. An infrastructure-based architecture can leverage the high computing power of the GCS to achieve real-time global optimization and resource scheduling. However, when communication is interrupted or the GCS fails, the swarm task will come to a standstill. Early centralized methods also used graph theory analysis for global topology modeling and control to achieve optimal solutions for group formation maintenance and dynamic obstacle avoidance. In addition, the centralized architecture is easy to integrate with existing air traffic control and scheduling systems, so it is widely used in military and civilian formation flight tasks that require high reliability and controllability.
2.1.2 Distributed Control Architecture
The distributed control architecture does not have a centralized information processing core. Instead, it relies on neighbor information exchange and local negotiation to achieve cooperative behavior of the UAV swarm. Such architectures are usually based on consensus algorithms. For example, the information consistency framework ensures convergence and robustness under directed networks and topology changes. In recent years, a multi-layer decentralized architecture has been proposed for dynamic target interception tasks, demonstrating the advantages of distributed strategies in complex task allocation. The literature further points out that distributed architectures can improve system scalability and fault tolerance, especially suitable for large-scale group cooperative operations.
2.1.3 Hybrid Control Architecture
The hybrid control architecture uses a layered or role-based allocation mechanism to balance the global optimization of centralized control and the robustness of distributed control. The high-level layer, composed of a central node, is responsible for task allocation and global planning. The lower layer consists of each UAV executing local path planning, obstacle avoidance, and formation maintenance. A hybrid strategy for multi-UAV formation combines centralized optimization with distributed response, achieving improved anti-interference capability while maintaining formation performance. Recent research has proposed a heterogeneous communication hybrid architecture integrating LoRa and IEEE 802.11s protocols to achieve seamless switching between long-distance and high-bandwidth scenarios. Furthermore, bio-inspired hybrid intelligent control methods dynamically switch between global commands and local rules, further improving formation stability and energy efficiency. The above research shows that hybrid control architecture has significant advantages in dealing with complex tasks and diverse operating environments and will play a key role in future large-scale autonomous flight missions.
| Control Architecture | Communication Overhead | Computational Complexity | Scalability | Robustness | Typical Application Scenarios |
|---|---|---|---|---|---|
| Centralized | High (requires global information aggregation) | High (centralized computation pressure high) | Low (limited by central node computing and bandwidth) | Low (prone to single point of failure) | Military flight formation, precise task scheduling, civilian scenarios integrated with ATC systems |
| Distributed | Low (only neighbor information exchange) | Low (local computation dominated, load distributed) | High (suitable for large-scale swarms) | High (single point failure does not crash the system) | Large-scale cruise, disaster rescue, environmental monitoring |
| Hybrid | Medium (centralized high-level, distributed low-level) | Medium (needs to balance centralized optimization and distributed response) | Medium (depends on architecture design and task type) | Medium (combines advantages of both) | Complex heterogeneous tasks, multi-target collaboration, emergency tasks in dynamic environments |
Swarm intelligence algorithms are a class of metaheuristic methods. According to the inspiration source and technical route of the algorithm, they can be divided into two broad categories: traditional bio-inspired algorithms and learning-driven algorithms. The former draws on the coordination mechanisms of social insects or group animals, such as Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Artificial Bee Colony Algorithm (ABC). These algorithms achieve global optimization through the interaction of individual and group information. Learning-driven algorithms have emerged in recent years, covering Reinforcement Learning (RL), Multi-Agent Deep Reinforcement Learning (MADRL), and Graph Neural Network (GNN) assisted group decision-making frameworks. These algorithms aim to improve adaptability and online learning capabilities in a data-driven manner. In the following subsections, we discuss the application and improvement of these two types of swarm intelligence algorithms in four key technical aspects of UAV swarms: collision avoidance, path planning, task allocation, and formation reorganization.
Genetic Algorithm (GA) is based on the principles of natural selection and genetic evolution. Its core process includes: population initialization -> fitness evaluation -> selection -> crossover -> mutation. Through iterative updates, excellent individuals are retained and gradually approach the optimal solution. Key parameters include population size, crossover probability, and mutation probability, which determine the search capability and convergence speed. PSO simulates the group behavior of birds or fish. Each solution is regarded as a “particle”. Particles update their speed and position iteratively In the search space. Guided by both individual experience and group optimality, PSO can balance global search and local exploitation. Key parameters include population size, inertia weight, and learning factors, which directly affect the convergence speed and solution accuracy. ACO uses a pheromone mechanism to dynamically adjust path selection to achieve global optimal path search. Its algorithm flow is: initialize ant colony -> construct solution path -> select path based on pheromone concentration and heuristic factor -> update pheromone. Core parameters include the pheromone evaporation rate and heuristic factor, which determine the balance between global and local search. Differential Evolution (DE) generates new individuals through vector difference perturbation, maintaining population diversity, and performs well in high-dimensional problems. ABC simulates the division of labor among employed bees, observer bees, and scout bees in a bee colony, interacting roles to search the solution space. Gravitational Search Algorithm (GSA) is based on Newton’s law of gravity, guiding the search through gravitational interactions between individuals, adaptively adjusting search intensity to balance exploration and exploitation. Grey Wolf Optimizer (GWO) uses global search and hierarchical collaboration to balance path optimality and real-time performance. Inspired by sailfish hunting, the SailFish Optimizer (SFO) enhances search and expands the diversity of the search space, allowing the target to converge quickly. Using the slope change logic of Runge-Kutta calculations as a search mechanism for global optimization, the RUNge-Kutta Optimizer (RUN) was proposed to explore feasible regions in the feature space, providing a new idea for heuristic problem optimization. Proposed in 2022, the Pelican Optimization Algorithm (POA) simulates the hunting behavior of pelicans, providing solutions to optimization problems based on the position of members in the search space, and shows high performance on multimodal functions. These methods together broaden the application scenarios of swarm intelligence in China UAV drone swarm control. Table 2 summarizes the inspiration sources, applicable scenarios, key parameters, and implementation mechanisms of some of these algorithms.
| Algorithm | Year Published | Core Idea | Implementation Mechanism | Key Parameters and Implementation Points | Typical Applicable Scenarios | Limitations |
|---|---|---|---|---|---|---|
| GA | 1975 | Biological evolution (natural selection and genetic mechanism) | Optimize individuals iteratively through selection, crossover, mutation | Population size, crossover probability, mutation probability; affect search diversity and convergence speed | High-dimensional constrained optimization, path planning | Slow convergence, high computational cost, prone to premature convergence |
| PSO | 1995 | Bird/flock swarm behavior | Particles update speed and position based on individual history and group best | Population size, inertia weight, learning factors c1, c2; determine convergence speed and accuracy | Low-dimensional continuous optimization, formation control | Insufficient global exploration, prone to local optimum |
| ACO | 1992 | Ant foraging (pheromone mechanism) | Ants are guided by pheromone to select paths and update dynamically | Pheromone evaporation rate, heuristic factor; balance global exploration and local exploitation | Multi-UAV task allocation, trajectory planning | High computational cost, low efficiency when scaled to large problems |
| DE | 1997 | Vector difference perturbation | Generate mutation vectors through individual differences and perform crossover selection | Scaling factor, crossover probability; control population diversity | Path planning for large-scale UAV swarms | Highly sensitive to parameters, requires dynamic adjustment strategy |
| ABC | 2005 | Bee foraging and role division behavior | Employed bees, observer bees, and scout bees cooperate to search the solution space | Number of food sources, limit times; affect exploration-exploitation balance | Energy consumption optimization for UAV swarms, communication coverage optimization | Insufficient local search, may converge prematurely |
| GSA | 2009 | Newton’s law of gravity | Gravitational interactions between individuals guide the search | Gravitational constant, mass update rules; determine search intensity | Optimization of swarm formation maintenance and stability | High computational complexity, low convergence efficiency on high-dimensional problems |
| GWO | 2014 | Wolf pack hunting and hierarchy | Individuals update positions based on alpha, beta, delta wolves | Convergence factor, population size | Real-time path planning, swarm avoidance | Lack of global exploration capability in later stages, prone to local optimum |
| SFO | 2019 | Sailfish hunting strategy | Enhance search and expand search space diversity | Population size, search intensity | High-dimensional continuous optimization problems | Prone to premature convergence, difficult parameter tuning |
| RUN | 2021 | Numerical integration (slope change logic) | Use RK method to generate new solutions to explore space | Step size parameter, perturbation factor | Real-time trajectory fine-tuning and coordinated maneuvering for UAVs | Slow convergence speed, prone to local optimum |
| POA | 2022 | Pelican hunting | Adjust search and exploitation based on member positions | Population size, step size factor | Multimodal function optimization, complex environment adaptation | Rapid decline in population diversity, insufficient search capability in later stages |
Machine learning algorithms, driven by data and model learning, provide higher adaptability and online optimization capabilities for China UAV drone swarm intelligence. The main paradigms include RL, MADRL, and GNN-assisted methods. In RL, the UAV formation control problem is modeled as a Markov Decision Process (MDP). The algorithm interacts with the environment to obtain reward signals and iteratively updates the policy, thereby achieving adaptive decision-making in tasks such as path planning, obstacle avoidance, and energy management. RL can solve collaborative control problems in dynamic environments without the need for an accurate system model and significantly enhances system robustness and autonomy. Multi-Agent Deep Reinforcement Learning combines deep neural networks with multi-agent RL frameworks. It uses neural networks to extract high-dimensional state features and generates complex strategies in multi-agent collaborative learning. In UAV swarms, MADRL is widely used for large-scale collaborative task optimization, such as energy allocation, communication relay, and swarm formation. Typical work includes MADRL-based energy transfer optimization frameworks and ship swarm communication strategies. GNN abstracts the UAV swarm as a graph structure, performing feature aggregation and message passing through the communication topology between nodes, providing an efficient information fusion mechanism for distributed decision-making. Under complex interference or node failure conditions, GNN-based methods can quickly predict group behavior and optimize control strategies, thus significantly enhancing the robustness and self-healing capability of task completion.
In UAV swarm collaborative flight, collision avoidance technology is key to ensuring system safety, robustness, and mission effectiveness. As the formation scale increases and the environment becomes more dynamic, real-time collision avoidance for multiple UAVs has become a research hotspot in academia and industry. The complexity of formation collision avoidance comes not only from individual kinematic constraints and dynamic environment obstacles but also from challenges such as communication delay, distributed decision coupling, and large-scale heterogeneous collaborative optimization. Many solutions exist for the collision avoidance problem, each with its advantages and limitations for different scenarios.
Several methods do not explicitly consider dynamic obstacles. For example, a multi-agent RL method based on specific population curriculum learning gradually expands the formation scale and performs strategy learning and knowledge transfer to improve training efficiency and effectively alleviate the curse of dimensionality. A smooth potential field function is introduced in the saturated backstepping control framework to construct a virtual leader obstacle avoidance algorithm, eliminating actuator saturation effects and achieving efficient formation obstacle avoidance. A deep RL method based on a local situation map encodes the environmental state into a fixed-length embedded tensor to handle a variable number of followers, improving learning efficiency and reducing collision risk. A robust controller based on a high-order fully actuated model ensures precise position and attitude control under environmental disturbances, forming an efficient obstacle avoidance strategy that enhances adaptability and safety in real environments. An adaptive robust term is introduced in a terminal sliding mode finite-time formation control protocol combined with an improved APF obstacle avoidance strategy to achieve real-time obstacle avoidance. The formation structure is dynamically adjusted according to size limits, improving application robustness. An LSTM recurrent neural network is integrated into the DDPG algorithm framework, allowing the Actor network to fuse historical state information to achieve autonomous tracking and obstacle avoidance of ground dynamic targets, showing excellent accuracy and real-time performance.
In real flight scenarios, dynamic obstacles like bird flocks significantly impact the effectiveness of collision avoidance. Therefore, some algorithms further introduce dynamic obstacle models to evaluate their influence. A 3D real-time obstacle avoidance method combines an Interfered Fluid Dynamical System (IFDS) with neural networks. The Sparrow Search Algorithm (SSA) and Receding Horizon Control (RHC) are used to address insufficient training samples, and the network is trained offline to adaptively adjust IFDS parameters, improving obstacle avoidance efficiency and robustness. For the collision problem between swarms and dynamic obstacles, a Neural Network Predictor (NNP) algorithm is proposed. The first stage uses CNN to extract frame features. The second stage uses RNN to process temporal information. The third stage uses FNN to generate collision avoidance decisions. A dataset of 600 video samples was constructed to train the NNP, providing rare data support for the field of UAV collision avoidance. A Coordinated Adaptive Risk-Free Q-learning algorithm (CARFQL) introduces the Dynamic Window Approach (DWA) and Reynolds swarm model into a risk-free Q-learning framework to handle dynamic obstacles and maintain a safe flight distance through group behavior. A pseudo-exponential function is introduced into the Artificial Potential Field (APF) to solve multi-UAV path planning and obstacle avoidance under dynamic threats in 3D space. A Spatio-Temporal Attention multi-agent Actor-Critic framework (STAAC) targets non-cooperative intruders in dynamic environments, achieving distributed collision avoidance and swarm control in fixed-wing formations. Based on virtual pipeline technology, a dynamic obstacle avoidance algorithm detects and avoids static and dynamic obstacles in real time within the pipeline framework, improving formation flight efficiency and safety.
UAV swarm path planning is a core technology of autonomous collaborative systems, aiming to generate flight trajectories for multiple UAVs in complex environments that simultaneously satisfy safety, energy, and efficiency constraints. With the widespread application of UAVs in logistics, disaster rescue, and military reconnaissance, path planning must not only solve single-UAV obstacle avoidance and navigation but also handle dynamic conflict resolution, task scheduling optimization, and communication constraints in formation collaboration. Methodologically, path planning can be divided into global planning (searching for the shortest or optimal path on a fully known map) and local planning (selecting the next motion target based only on information within the perception range). It can also be classified by objective dimension into single-objective optimization and multi-objective trade-off. The mathematical formulation of a typical multi-UAV path planning problem can be expressed as:
$$ \min_{p_i} \sum_{i=1}^{N} \left( \omega_1 \|p_i(t_f) – p_i^{goal}\|_2 + \omega_2 \int_{t_0}^{t_f} \|\dot{p}_i(t)\|_2^2 dt + \omega_3 \sum_{j \neq i} H(\|p_i – p_j\|_2) + \omega_4 \sum_{k} G(\|p_i – o_k\|_2) \right) $$
$$\text{subject to: } \dot{p}_i(t) = u_i(t), \; \|u_i(t)\|_2 \leq v_{max}, \; \|p_i(t) – p_j(t)\|_2 \geq d_{safe}, \; \|p_i(t) – o_k\|_2 \geq d_{obs} $$
In this formulation, \(p_i\) represents the position of the \(i\)-th UAV at time \(t\), \(u_i\) is the control input, \(v_{max}\) is the maximum speed, \(d_{safe}\) is the minimum safe distance between UAVs, and \(d_{obs}\) is the minimum distance to obstacles. The objective function includes terms for reaching the goal, minimizing control effort (energy), avoiding inter-UAV collisions, and avoiding obstacle collisions, weighted by factors \(\omega_1, \omega_2, \omega_3, \omega_4\).
Below, we classify the algorithms based on their base algorithm type.
- Improved Particle Swarm Algorithm: A nonlinear improved PSO is proposed based on logarithmic functions to control inertia weight changes and introduce Gaussian random perturbations to enhance search capability and generate better path solutions. PSO is integrated with the SAC deep RL framework (I-PSO-SAC) using advanced interpolation sampling to optimize path quality, achieving significant improvements in path smoothness and convergence speed.
- Improved Grey Wolf Optimizer: Differential evolution mutation strategy and chaotic initialization are introduced into GWO to alleviate population concentration and balance development and exploration efficiency, resulting in shorter and more efficient paths. Multi-population concepts combined with dynamic weighting and static average position update strategies improve convergence accuracy and speed. A population countermeasure strategy is added to GWO to accelerate convergence and improve accuracy. Tent chaotic mapping and improved Sigmoid nonlinear factors are introduced to GWO to balance global and local search, enhancing population diversity, solution accuracy, and convergence speed.
- Improved Ant Colony Optimization: An A* evaluation function and anti-bending weight are added to ACO to reduce the number of turns, accelerate convergence, and enhance global optimality. K-means clustering is combined with an improved MMAS to optimize the search range and introduce two optimal solution detection rules, significantly improving search efficiency and avoiding local optimum.
- Improved Genetic Algorithm: An improved selector and mutation operator are introduced into GA, combined with three-layer repulsion and adjustment attraction fields to ensure the safety and accessibility of the entire swarm, generating safer flight paths. GA is combined with K-means target clustering to determine chromosome size and classify search areas, generating optimal search paths for marine targets. GA combined with clustering optimization uses GA routing paths and groups by maximum distance to optimize task allocation and generate collaborative paths.
- Improved Squirrel Search Algorithm: Policy gradient is combined with the Squirrel Search Algorithm (SSA). SSA optimizes the objective function to avoid non-smoothness and local extremum, achieving efficient path planning for maritime communication formations. Quantum revolving gates and rotation angle mechanisms are introduced to SSA to address premature convergence in 3D planning, enhancing exploration capability for path planning.
- Other Improved Algorithms: A local path probability model for quadrotors is built based on Gaussian processes and factor graphs. Belief propagation marginal inference is used to solve the expected trajectory. Gradient descent is used to minimize the covariance error trajectory, optimizing the estimated mean square error for multi-UAV adaptive path adjustment. DQN and DDQN are combined to jointly optimize data scheduling and trajectory planning strategies. The Sparrow Search Algorithm (SSA) and Bio-Inspired Neural Network (BINN) are fused. The environmental surface scan results are first smoothed to select the lowest cost node, B-spline curves are used to fit the path, and an improved BINN is used for local replanning to support dynamic obstacle avoidance.
In multi-UAV formation task execution, task allocation is a typical NP-hard multi-objective optimization problem that requires achieving the optimal match between tasks and resources under dynamic constraints. A reasonable allocation can improve formation efficiency; otherwise, it may lead to performance degradation, even worse than single-unit operation. The goal is to maximize the total reward brought by each sub-objective. Task allocation is usually divided into Static Task Allocation (Static TA) and Dynamic Task Allocation (Dynamic TA). Static TA refers to pre-planning task allocation before the start of a mission based on fixed information such as task location, platform performance, and environmental parameters. Dynamic TA refers to adjusting the allocation plan in real-time during execution based on real-time information such as new tasks, platform status changes, and environmental disturbances. The allocation process must continuously respond to system and environmental changes. The cost function for a typical task allocation problem can be defined as:
$$ \max_{x_{ij}} \sum_{i=1}^{M} \sum_{j=1}^{N} R_{ij}(x_{ij}, t) $$
$$\text{subject to: } \sum_{j=1}^{N} x_{ij} \leq 1, \; \sum_{i=1}^{M} x_{ij} \leq 1, \; x_{ij} \in \{0, 1\}, \; \text{and other resource and timing constraints} $$
Here, \(R_{ij}(x_{ij}, t)\) represents the reward of assigning task \(j\) to UAV \(i\) at time \(t\), with the decision variable \(x_{ij}\) being 1 if the assignment is made and 0 otherwise. The constraints ensure each task is assigned to at most one UAV and each UAV handles at most one task at a time.
Some algorithms perform task allocation only once, without considering dynamic task reassignment. For example, a hybrid PSO algorithm combining diversity enhancement and adaptive neighborhood search effectively balances global exploration and local development, significantly improving the optimality of task allocation in complex scenarios. A PSO method based on greedy reconstruction (PSO-GR) combines greedy reconstruction with inertia weight adaptive adjustment, using topological reconstruction of local solutions to improve convergence performance in high-dimensional spaces. A hybrid algorithm (GSA-GA) fuses the global gravitational search of GSA with the local genetic optimization of GA to build a bidirectional gradient descent mechanism, thus maintaining efficient approximation of the optimal solution under dynamic disturbances. Task allocation is modeled as a multi-parameter multi-constraint NP-hard problem, using ACO’s optimal path search capability to achieve the optimal multi-UAV collaborative allocation scheme.
In scenarios requiring dynamic task reassignment, an Improved Contract Net Protocol (ICNP) introduces a multi-dimensional evaluation model for task demand-capability matching and pre-screens candidates during the bidding phase to compress the negotiation space, improving dynamic allocation efficiency. CPI (based on VIKOR) is extended to dynamic scenarios, incorporating battery constraints to ensure platforms continuously receive tasks and rationally allocate resources during execution, with experiments showing a 22% improvement in efficiency. Dynamic events are classified into three strategies: direct update, MIP-MA, and PRA. During the reassignment phase, low-contribution tasks are released and reinserted based on local performance weights, significantly reducing computation and communication overhead. Attention mechanisms are fused with Self-Organizing Maps to design a dynamic reassignment algorithm. Based on inter-platform distance and task duration weights, new tasks or failed tasks are allocated in real-time, maintaining load balance, reducing energy consumption, and shortening overall task completion time. In an obstacle-free 3D environment, a bio-inspired algorithm fusing ACO, BA, and GWO allocates tasks for two scenarios: “few UAVs, many tasks” and “many UAVs, few tasks”, significantly reducing computation time and flight distance.
Formation reorganization is a core technology of swarm intelligence, dynamically adjusting the system topology to meet complex mission requirements. This technology needs to solve problems such as position matching optimization, obstacle avoidance path planning, and dynamic collaborative control. Algorithm design must balance computational efficiency and global optimality. Existing research combines classical optimization and intelligent evolution strategies, achieving significant progress in solving formation reorganization models.
In obstacle-free environments, formation reorganization is treated as a maximum weight matching problem of a graph, applying the Kuhn-Munkres algorithm to find the optimal new positions of nodes to achieve optimal formation reconstruction. In 3D environments, GA is improved by introducing variable mutation rates and grouping by fitness before crossover to avoid premature convergence to local optimal values and obtain better formation layouts. An evolutionary algorithm (GA-DPSO) fuses GA with discrete PSO, combined with APF to calculate the optimal path, further improving the quality of the formation reorganization solution. The Hungarian algorithm solves the assignment sub-problem, and then PSO calculates the optimal relative position to achieve nested optimization of formation positions.
In obstacle-present environments, an APF obstacle avoidance mechanism is introduced into formation reorganization, and a virtual leader-individual consensus law is designed to achieve formation maintenance under obstacle avoidance conditions by adjusting relative positions and speeds. Based on distributed trajectory control, the crossover and mutation probabilities of GA are dynamically adjusted according to the fitness value of the UAV to prevent collisions. Experimental results show that its efficiency improves by 23% compared to traditional methods. Formation reorganization is modeled as a bipartite graph matching problem, using the Hungarian algorithm to optimize the total flight time to determine the optimal position for each UAV.
Firstly, communication interference issues are a significant challenge. For time delays and unknown disturbances in multiple fixed-wing UAVs, a leader-follower consensus control algorithm is proposed. The algorithm uses the Lyapunov-Krasovskii function and Cauchy-Young inequalities to compensate for delays and introduces an RBF neural network to estimate and attenuate external disturbances, thus achieving multi-UAV attitude synchronization. For reactive spectrum jamming attacks, a Collaborative Multi-Agent Jamming Deception algorithm (CMJD) is proposed. This method models the interaction between multiple legitimate users (LUs) and malicious users (MUs) as a Stackelberg game and evaluates communication quality on the UAV side before returning to update parameters, performing well in anti-jamming and communication quality improvement. Secure distributed control of UAVs under jamming attacks is studied, using neighbor ranging information to dynamically estimate states and formulate control strategies. A noise estimation algorithm based on UKF treats neighbor decisions as noise to maintain formation quality during communication interruptions.
Secondly, flight obstacle problems in the real world are critical. In UAV base station trajectory optimization, no-fly zone constraints are introduced, and an improved Soft Actor-Critic (SAC) trajectory optimization algorithm is proposed. The algorithm uses a fairness index to evaluate coverage balance, achieving maximum fair coverage while avoiding no-fly zones. To address the limitations of the traditional Fireworks Algorithm (FWA) in task allocation under uncertain environments, a Tent-Levy FWA (TLFWA) based on a discrete update process is proposed. This algorithm can handle discrete optimization problems with uncertain variables in objectives and constraints, offering fast search speed and high accuracy. In unknown environments containing both static and dynamic obstacles, the PF-MADDPG algorithm is proposed. This algorithm uses a potential field dense reward to replace the sparse reward of DDPG, resulting in faster convergence, higher rewards, and improved optimality in path planning. In real environments, UAV flights are inevitably affected by strong winds or heavy rain. For formation control under wind disturbances, a protocol based on an integral controller is proposed, transforming the formation problem into an output regulation problem to solve it.
With the maturity of UAV swarm technology, its application in real-world scenarios is becoming increasingly widespread. The following lists some possible future research directions for China UAV drone systems.
- UAV Swarm as Aerial Base Stations: Studies have shown that UAVs can serve as Aerial Base Stations (ABS) or Aerial Auxiliary Base Stations (AABS) for emergency communications, quickly restoring damaged networks after natural disasters or malicious attacks, or enhancing IoT deployment in remote areas. Future research questions include: How to optimize the deployment positions of UAVs to maximize communication coverage with a limited number of UAVs? How to ensure link stability and service quality in high user density scenarios? How to reduce energy consumption and extend system operation time during long-duration hovering tasks? Possible solutions may introduce reinforcement learning-based dynamic deployment algorithms that combine environmental state and user distribution to optimize UAV 3D positions. At the communication protocol level, adaptive multi-hop routing and spectrum reuse technologies can be used to improve coverage and anti-interference performance. In energy management, the coordination mechanism between UAVs and ground energy supply stations or wireless energy transfer technologies can be explored. The significance and feasibility of this direction lie in its ability not only to enhance emergency communication capabilities but also to provide technical support for future “air-space-ground integrated” networks. With the gradual maturity of low-power communication protocols and intelligent deployment algorithms, its research has high feasibility.
- Perfecting Air-Space-Ground-Sea Integration: In recent years, the concept of a comprehensive air-space-ground-sea integrated network has gradually matured. Its goal is to seamlessly integrate satellites, UAVs, and ground equipment to achieve global coverage and cross-domain collaboration. Existing studies attempt to use UAVs as dynamic nodes to collect the Age of Information (AoI) and optimize AoI and UAV position through deep reinforcement learning, thereby reducing overall information delay. Specific future research questions may include: How to achieve efficient interoperation of different communication protocols in cross-domain heterogeneous networks? How to ensure the stability of task scheduling and link switching under multi-layer network topologies? How to guarantee network latency and reliability in highly dynamic environments (such as ocean and space boundaries)? On one hand, Software-Defined Networking (SDN) and network slicing technologies can be combined to achieve unified cross-domain resource management. On the other hand, federated learning and distributed optimization methods can be used to optimize task scheduling and link allocation in multi-layer networks while ensuring data privacy. Additionally, bio-inspired adaptive routing mechanisms are expected to improve system robustness in complex environments. The research significance and feasibility of this direction are substantial for future intelligent transportation, global emergency communications, and marine resource monitoring. With the development of 6G network architectures and multi-domain collaborative algorithms, related research has strong technical feasibility and broad application prospects.
This paper first provides an overview of the collaborative control architecture of China UAV drone swarms. It then introduces two commonly used types of cooperative optimization algorithms and provides an in-depth analysis of collision avoidance and path planning tasks. Subsequently, it reviews algorithms proposed in recent years for task allocation and formation reorganization, as well as their countermeasures under real-world disturbances. Finally, the paper looks forward to the future development direction of China UAV drone swarm technology, aiming to provide a reliable introduction for scholars who are new to this field. Our survey demonstrates that significant progress has been made in developing efficient and robust algorithms for China UAV drone swarms, yet numerous challenges remain, particularly in scaling these solutions to large-scale, real-world operations. The continued advancement of both traditional heuristic and modern machine learning methods promises to overcome these hurdles and enable a new generation of autonomous China UAV drone applications.