Drone delivery has emerged as a transformative logistics model, and achieving efficient collaborative task planning among multiple unmanned aerial vehicles (UAVs) has become a pressing technical challenge for the logistics industry, particularly within the rapidly evolving China UAV ecosystem. Task planning technology, as the core of efficient collaborative operations for multiple logistics UAVs, encompasses two critical components: task allocation and trajectory planning. It aims to optimize resource allocation and improve task execution efficiency. In this survey, we first outline the basic concepts and control structures of multi-UAV task planning. We then analyze the solution methods for task allocation and trajectory planning, discuss the centralized and layered decoupling strategies for solving the overall problem, and finally propose future development trends for multi-logistics UAV task planning, with a strong emphasis on the context of China UAV applications.
1. Introduction
The low-altitude economy, driven by both manned and unmanned aerial vehicle operations, has been formally recognized as a key growth sector in China since the 2024 National People’s Congress. China UAV technology, as a typical representative of new productive forces, integrates information technology, digitalization, aviation, and drone expertise. Within this domain, logistics UAVs are playing an increasingly vital role, especially in last-mile delivery, emergency supply transport, and remote area logistics. As the scale of China UAV deployment expands, the efficient coordination of multiple UAVs for task planning becomes a critical issue that the logistics industry must address. In this article, we provide a comprehensive survey of multi-UAV task planning for logistics, focusing on the unique constraints and opportunities presented by the China UAV ecosystem.
2. Problem Formulation
Multi-UAV task planning in logistics is a complex combinatorial optimization problem. It involves assigning a set of delivery tasks to a fleet of UAVs and planning feasible flight paths subject to numerous constraints. We define the problem as follows:
Let \( \mathcal{U} = \{U_1, U_2, \dots, U_{N_u}\} \) be the set of UAVs and \( \mathcal{T} = \{T_1, T_2, \dots, T_{N_t}\} \) be the set of delivery tasks. Each task \( T_j \) has a location \( (x_j, y_j, z_j) \), a time window \( [e_j, l_j] \), a priority \( p_j \), and a weight \( w_j \). Each UAV \( U_i \) has a maximum payload \( C_i \), maximum range \( R_i \), speed \( v_i \), and battery capacity constraints. The objective is to determine an assignment \( A: \mathcal{T} \to \mathcal{U} \) and a set of trajectories \( \mathcal{P} = \{P_1, P_2, \dots, P_{N_u}\} \) such that all tasks are completed while minimizing a weighted combination of total flight distance, total time, risk, and delay penalties.
The general optimization model can be stated as:
$$
\begin{aligned}
\min & \quad f(\mathbf{x}) = \alpha_1 \cdot D(\mathbf{x}) + \alpha_2 \cdot T(\mathbf{x}) + \alpha_3 \cdot R(\mathbf{x}) + \alpha_4 \cdot P(\mathbf{x}) \\
\text{s.t.} & \quad \sum_{i=1}^{N_u} x_{ij} = 1, \quad \forall j \in \{1,\dots,N_t\} \\
& \quad \sum_{j=1}^{N_t} w_j \cdot x_{ij} \leq C_i, \quad \forall i \\
& \quad \text{distance}(P_i) \leq R_i, \quad \forall i \\
& \quad t_{\text{arrival},j} \in [e_j, l_j], \quad \forall j \\
& \quad \text{no collisions between UAVs and with obstacles}
\end{aligned}
$$
where \( x_{ij} \) is a binary variable indicating whether task \( j \) is assigned to UAV \( i \), and \( D, T, R, P \) represent distance, time, risk, and penalty cost functions.
3. Constraints and Evaluation Criteria
In the China UAV logistics context, we identify four main categories of constraints:
| Category | Specific Constraints | Description |
|---|---|---|
| UAV Constraints | Number, range, payload, dynamics | Each China UAV has limited battery, weight capacity, and kinematic limits (speed, climb angle, turn radius). |
| Task Constraints | Coverage, time windows, priority | Every delivery must be completed exactly once; time windows (soft or hard) must be respected; high-priority tasks are favored. |
| Environmental Constraints | Weather, obstacles, no-fly zones | Urban buildings, trees, power lines, and no-fly zones (e.g., government buildings) must be avoided. Wind and rain affect flight stability. |
| Swarm Constraints | Collision avoidance, temporal/spatial coordination | UAVs must maintain safe separation distance; tasks may require simultaneous arrival or sequential execution. |
The evaluation metrics for multi-UAV task planning typically include:
- Total flight distance – minimize energy consumption.
- Total mission time – critical for time-sensitive deliveries.
- Risk – minimize exposure to hazards (e.g., flying over populated areas).
- Delay penalty – cost incurred for late deliveries.
- Number of UAVs used – resource efficiency.
- Task value – maximize total priority points achieved.
These objectives are often conflicting; we employ weighted sum, hierarchical optimization, or Pareto-based multi-objective methods.
4. Control Structures
The control architecture of a multi-UAV system significantly influences the solution approach. We classify three typical structures:
| Structure | Advantages | Disadvantages | Applicability |
|---|---|---|---|
| Centralized | Global optimum possible; strong coupling between task allocation and path planning | Single point of failure; high computational load; poor scalability | Small numbers of UAVs (~10); static environments |
| Distributed | High robustness and scalability; real-time adaptability | May miss global optimum; requires frequent communication; conflict resolution needed | Large swarms (~100+); dynamic environments |
| Hybrid | Balances global optimization and local autonomy; resilient | Complex coordination mechanisms; design trade-offs | Medium-sized fleets; mixed static/dynamic tasks |
In many China UAV logistics applications, hybrid control structures are increasingly favored due to the combination of centralized dispatch from ground control stations and distributed local decisions for obstacle avoidance.
5. Task Allocation Methods
Task allocation determines which UAV performs which task. We categorize approaches into centralized, distributed, and hybrid algorithms.
5.1 Centralized Algorithms
Centralized algorithms solve an integer or mixed-integer programming model. Exact methods include branch-and-bound, dynamic programming, and integer linear programming (ILP). However, exact solutions become intractable for large \( N_t \) and \( N_u \). Therefore, metaheuristics are widely used.
Genetic Algorithm (GA): GA encodes a task sequence as chromosomes. In China UAV research, GA variants with tailored encoding and repair mechanisms are common. For example:
Let the chromosome represent a permutation of tasks. Crossover and mutation operators are applied. The fitness function is the reciprocal of the total cost.
Particle Swarm Optimization (PSO): Each particle represents a candidate assignment with position and velocity. We use a discrete version with swap operators.
Ant Colony Optimization (ACO): Ants construct solutions by moving through a graph of tasks, with pheromone trails guiding decisions.
| Algorithm | Encoding | Key Modifications for China UAV Logistics |
|---|---|---|
| GA | Task permutation | Double-chromosome for heterogeneous UAVs; repair mechanism for constraints |
| PSO | Task-to-UAV matrix | Multi-objective variant (KnCMPSO); quantum-behaved PSO for faster convergence |
| ACO | Task sequence | Roulette wheel selection with elite strategy; dynamic pheromone update |
| Simulated Annealing | Permutation | Hybrid with GA to escape local optima |
5.2 Distributed Algorithms
Distributed algorithms allow each China UAV to negotiate with peers. The most popular method is the Consensus-Based Bundle Algorithm (CBBA) and its variants. CBBA works in two phases: bundle construction and consensus. Each UAV maintains a list of tasks it intends to perform (bundle) and communicates with neighbors to resolve conflicts. Auction-based algorithms also fall under this category, where tasks are bid upon and the highest bidder wins.
Formulation of Auction Process:
Let \( b_{ij} \) be the bid of UAV \( i \) for task \( j \). The bid is typically the marginal cost savings from adding task \( j \) to the current bundle. The auctioneer (possibly a designated UAV) collects bids and assigns task \( j \) to the UAV with the highest bid. Consensus algorithms ensure global consistency without a central coordinator.
5.3 Hybrid Task Allocation
Hybrid methods combine initial centralized planning with online distributed adjustment. For instance, a genetic algorithm may perform initial assignment, and subsequently a contract net protocol handles dynamic task insertions. In China UAV scenarios, this is especially useful for emergency deliveries where new tasks appear in real time.
6. Trajectory Planning Methods
Trajectory planning (also called path planning) generates a collision-free path from the UAV’s depot to all assigned task points and back. The problem is typically solved in continuous 3D space, with constraints on altitude, climb angle, and avoidance of obstacles.
6.1 Environment Modeling
We commonly use three approaches for representing the environment:
- Grid-based method: The space is discretized into cells (2D or 3D). A* or Dijkstra algorithms search for the lowest-cost path. Cost values represent risk, distance, or fuel. In China UAV urban logistics, grid resolution must balance computational load and path quality.
- Graph-based methods: Voronoi diagrams, visibility graphs, and Rapidly-exploring Random Trees (RRT) represent the environment as nodes and edges. Voronoi diagrams ensure paths stay away from obstacles; RRT is efficient for high-dimensional spaces.
- Artificial Potential Field (APF): Attractive forces from the goal and repulsive forces from obstacles guide the UAV. APF is fast but prone to local minima.
We summarize the characteristics in the table below.
| Method | Complexity | Path Quality | Suitability for Urban China UAV |
|---|---|---|---|
| Grid (2D/3D) | O(n³) | Medium (grid effect) | Widely used; easy to incorporate risk maps |
| Voronoi Diagram | O(n log n) | Good (safe distance) | Suitable when obstacle polygons known |
| Visibility Graph | O(n²) | Optimal for polygonal obstacles | High computation for many obstacles |
| RRT | O(m log m) | Probabilistically complete | Good for dynamic unknown environments |
| APF | O(n) | Low (local minima) | Only with hybrid or improved versions |
6.2 Trajectory Planning Algorithms
We categorize trajectory planning algorithms into classical search and metaheuristic methods, with recent trends toward 4D (space + time) planning for multi-UAV coordination.
A* Algorithm: A grid-based algorithm that uses heuristic estimates to guide search. Improved versions include dynamic weighting, bidirectional search, and safety-aware cost functions. For example, in urban China UAV logistics, we introduce a risk term \( R(\text{cell}) \) into the cost function:
$$
f(n) = g(n) + h(n) + \lambda \cdot R(n)
$$
where \( g(n) \) is the traveled cost from start, \( h(n) \) is the heuristic distance to goal, and \( \lambda \) weights risk.
Particle Swarm Optimization (PSO): In continuous path planning, each particle represents a sequence of waypoints. The velocity update formula is modified to handle constraints. In China UAV studies, we often combine PSO with cubic spline interpolation to smooth the path.
Ant Colony Optimization (ACO): ACO is applied to find the shortest path in a graph. For 3D trajectory planning, we discretize the space into layers and allow ants to move between layers.
Genetic Algorithm (GA): GA evolves a set of waypoints. Chromosome encoding can be the coordinates of each waypoint. Constraints are handled via penalty functions.
Improved Grey Wolf Optimizer (IGWO): We adjust the hunting mechanism to adapt to path constraints. An example modification includes introducing a repulsion operator to avoid obstacles.
In practice, hybrid algorithms combining multiple methods are common. For instance, A* may generate an initial path, and then a PSO refines it locally.
6.3 4D Trajectory Planning for Multi-UAV Coordination
For China UAV swarms operating in the same airspace, 4D planning (3D space + time) is essential to avoid conflicts and ensure temporal coordination (e.g., simultaneous arrival or sequential delivery). The problem can be formulated as:
Let \( \mathcal{P}_i(t) \) be the position of UAV \( i \) at time \( t \). The constraints are:
\[
\| \mathcal{P}_i(t) – \mathcal{P}_j(t) \| \geq d_{\text{safe}}, \quad \forall i \neq j, \forall t
\]
and for time-critical tasks:
\[
t_i^{\text{arrival}} = t_j^{\text{arrival}}, \quad \text{or} \quad t_i^{\text{arrival}} \leq t_j^{\text{arrival}} \text{ (order constraint)}
\]
Approaches include velocity adjustment, altitude separation, and time-shifting of takeoff times. In the literature related to China UAV, a popular method is to first plan a spatial path, then perform velocity scheduling using a consensus algorithm to synchronize arrival times.
7. Integrated Task Planning Approaches
Since task allocation and trajectory planning are tightly coupled, solving them separately may lead to suboptimal solutions. We discuss two main strategies to handle this coupling: centralized (simultaneous) and layered decoupling.
7.1 Centralized (Integrated) Approach
In the centralized approach, we embed trajectory cost estimation directly into the task allocation. This can be done by encoding both the task sequence and the path waypoints in a single chromosome. For example, a GA may represent each UAV’s route as a sequence of task IDs, and the fitness function calls a low-level path planner to compute the actual flight distance. This approach ensures that allocation decisions are based on realistic path costs, but it is computationally expensive. Another method is to use a multi-agent reinforcement learning (MARL) framework where centralized training with decentralized execution allows the system to learn both allocation and planning jointly. In China UAV applications, this is promising for large-scale dynamic logistics.
7.2 Layered Decoupling
Layered decoupling separates the problem into two levels: an upper level handles task allocation, and a lower level handles trajectory planning. Information flows between levels either in a one-way (loose coupling) or iterative manner.
Loose coupling: The upper level uses approximate distances (e.g., Euclidean distance) to allocate tasks. The lower level then plans precise paths. There is no feedback. This is efficient but may produce infeasible paths for certain assignments.
Iterative coupling: The lower level sends back actual path costs or feasibility flags to the upper level. The upper level revises the allocation accordingly. This iteration continues until convergence or a maximum number of iterations is reached. A common implementation is a bilevel optimization:
\[
\min_{\mathbf{x}} \left( \text{Task allocation cost} + \text{Path cost from lower level} \right)
\]
where the path cost is defined by the solution of the lower-level problem:
\[
\min_{\mathbf{p}_i} \text{Path cost}(\mathbf{p}_i) \quad \text{subject to dynamic and obstacle constraints}
\]
This method achieves better overall quality at the expense of increased computation time.
8. Challenges and Future Directions for China UAV Logistics
Despite significant progress, several challenges remain for multi-UAV task planning in the China UAV logistics ecosystem:
- Heterogeneous goods: Current models often assume uniform cargo, but real logistics requires handling diverse items (e.g., food, electronics, pharmaceuticals). Future research should incorporate compatibility constraints and compartmentalized loads.
- Real-time dynamic tasks: Most studies focus on static batch assignments. In practice, orders arrive continuously. We need online algorithms that can insert new tasks without fully recomputing the plan, using techniques like rolling horizon or dynamic auction.
- 4D trajectory coordination: As China UAV traffic increases, airspace management becomes critical. We need robust, scalable algorithms for conflict detection and resolution, possibly integrated with future urban air traffic management systems.
- Machine learning integration: Deep reinforcement learning (DRL) can learn complex policies for both task allocation and path planning. Especially, multi-agent DRL (MADRL) can handle decentralized execution with centralized training, suitable for large-scale China UAV swarms.
- Resilience and fault tolerance: In case of UAV failure or communication loss, the system must reallocate tasks and reroute quickly. Distributed consensus algorithms with fallback mechanisms are essential.
- Regulatory compliance: China’s evolving low-altitude aviation regulations impose constraints on flight altitudes, no-fly zones, and noise levels. Task planners must integrate these rules as hard constraints.
9. Conclusion
In this survey, we have provided a comprehensive overview of multi-UAV task planning for logistics, with a particular focus on the China UAV ecosystem. We discussed problem formulations, constraints, evaluation criteria, control structures, and the key subproblems of task allocation and trajectory planning. Both centralized and distributed solution methods were reviewed, along with integrated and layered approaches for handling coupling. Future research should focus on real-time adaptability, heterogeneous goods, AI-driven methods, and robust coordination to fully unlock the potential of China UAV in the low-altitude economy.