With the rapid rise of the low-altitude economy, unmanned aerial vehicle (UAV) delivery has undergone transformative changes in both scenario expansion and technological deepening. In China, the low-altitude economy has become a strategic new growth engine, with the market reaching 505.95 billion RMB in 2023 and projected to exceed 1.5 trillion RMB by 2025. The “15th Five-Year Plan” explicitly lists low-altitude aviation as a strategic emerging industry, providing strong policy support for the development of China drone logistics. Path planning is the key technology enabling intelligent and large-scale UAV delivery, directly impacting economic efficiency and operational safety. This article systematically reviews the current research status of UAV delivery path planning from a first-person perspective, summarizing our work on problem modeling, method classification, practical applications, technical challenges, and future trends.
1. Problem Modeling for UAV Delivery Path Planning
We formulate UAV delivery path planning as a multi-objective, strongly constrained optimization problem. The primary objectives are economic efficiency, service quality, and safety. Economic objectives aim to minimize total flight time or energy consumption. Service quality objectives focus on meeting customer time windows and ensuring delivery reliability. Safety objectives maximize the distance from risk sources and ensure collision avoidance. The constraints can be categorized into three groups: physical performance constraints, mission and environment constraints, and airspace and regulatory constraints.
Physical constraints include maximum range, payload capacity, battery capacity, and dynamics (e.g., maximum turning angle, climb rate). Mission and environment constraints involve customer time windows, task sequence precedence, and cooperative synchronization (e.g., truck-drone rendezvous). Airspace and regulatory constraints require avoidance of static obstacles (buildings, mountains), no-fly zones, and dynamic obstacles. These constraints are mathematically expressed as:
$$ \text{Minimize } J = \alpha_1 \cdot \text{Cost}_{\text{time}} + \alpha_2 \cdot \text{Cost}_{\text{energy}} + \alpha_3 \cdot \text{Penalty}_{\text{safety}} $$
subject to:
$$ \begin{aligned}
& \text{Range: } \sum_{i} d_i \leq D_{\max} \\
& \text{Time window: } t_i \in [e_i, l_i] \\
& \text{Dynamics: } |\Delta \theta_i| \leq \Theta_{\max}, |\Delta \phi_i| \leq \Phi_{\max} \\
& \text{Obstacle avoidance: } \text{dist}(p_i, O_j) \geq d_{\text{safe}} \\
& \text{No-fly zone: } p_i \notin \mathcal{F}
\end{aligned} $$
The following table summarizes the key modeling elements.
| Dimension | Specific Element | Description |
|---|---|---|
| Optimization Objectives | Economic | Minimize total flight time/distance or energy consumption to reduce direct operating cost |
| Service Quality | Meet customer time windows; improve delivery reliability and punctuality | |
| Safety & Robustness | Maximize distance from risk sources; quantifiable safety metric | |
| Physical Constraints | Maximum Range | Limit on single-flight distance |
| Payload | Maximum weight the UAV can carry | |
| Battery & Energy | Energy consumption model and battery capacity limit | |
| Dynamics | Turning angle, climb rate, etc. | |
| Mission & Environment | Customer Time Windows | Hard window constraints |
| Task Sequence | Precedence order among multiple delivery points | |
| Cooperative Operation | Truck-drone rendezvous timing and location synchronization | |
| Airspace & Regulations | Static Obstacle Avoidance | Avoid buildings, mountains, etc. |
| No-fly Zones | Comply with airspace regulations | |
| Dynamic Risk | Quantify risk from moving obstacles |
2. Classification of Path Planning Methods
According to the core decision-making logic, we classify existing path planning methods into three categories: model-driven, data-driven, and rule-driven. Each category is further divided into sub-categories with representative algorithms.
2.1 Model-Driven Optimization Methods
Model-driven methods rely on explicit mathematical modeling of the environment and UAV dynamics, solved by optimization algorithms. They include classical graph search algorithms, metaheuristics, artificial potential fields, and mathematical programming. For example, the A* algorithm uses a heuristic function to guide the search; bidirectional A* reduces node expansion. Improved Theta* allows arbitrary-angle paths, reducing redundancy. Metaheuristics like particle swarm optimization (PSO) and genetic algorithm (GA) are applied to multi-UAV coordination. The artificial potential field method combines attractive and repulsive forces for local real-time obstacle avoidance but often hybridizes with global planners. Mathematical programming (e.g., mixed-integer linear programming) provides provably optimal solutions for small-to-medium sized problems.
| Algorithm | Technical Features | Problem Solved | Advantages | Limitations | Scale |
|---|---|---|---|---|---|
| Bidirectional A* | Simultaneous search from start and goal | Real-time 3D path planning | Reduced node expansion, faster planning | Path quality depends on meeting strategy | Single / small |
| Improved Theta* | Allows line-of-sight connections | Smooth path on grid | Shorter, more natural trajectories | Higher computational cost | Single |
| Improved PSO | Dynamic parameter adjustment | Multi-UAV path optimization | Faster convergence, better solution quality | May fall into local optimum | Small-medium (≤30 UAVs) |
| Tuna Swarm Optim. | Mimics tuna foraging | Large-scale urban logistics | Excellent global search | Needs more validation | Large cluster (≥50) |
| A* + Artificial Potential Field | Hybrid global+local | Dynamic obstacle avoidance | Global optimization + real-time response | Complex architecture, local minima | Single dynamic |
| Mixed-Integer Programming | Exact modeling | Multi-UAV cooperative task allocation | Provably optimal, rigorous logic | High complexity, small scale only | Small (≤10) |
2.2 Data-Driven Intelligent Methods
Data-driven methods learn decision policies directly from historical data or environment interactions without relying on predefined physical models. Reinforcement learning (RL) models the problem as a Markov decision process (MDP) and uses reward functions to guide UAV learning in dynamic environments. Graph neural networks (GNNs) naturally capture interactions among multiple agents by representing each UAV or node as part of a graph. End-to-end learning directly maps sensor inputs (e.g., camera images, LiDAR) to control commands, achieving agile flight at speeds up to 20 m/s in dynamic obstacle avoidance. Hybrid approaches combine deep learning with traditional planners for better interpretability and robustness. The following table summarizes key data-driven methods.
| Method | Technical Characteristics | Problem Addressed | Advantages | Limitations |
|---|---|---|---|---|
| Deep Reinforcement Learning | MDP modeling, trial-and-error learning | Online decision in uncertain environments | Adapts to complex dynamics | Black-box, high training cost |
| Multi-Agent GNN | Graph representation for interactions | Multi-UAV coordination & conflict resolution | Captures spatio-temporal dependencies | Complex training, communication overhead |
| Differentiable Physics Simulation | End-to-end: perception to control | High-speed agile flight, uncoordinated swarms | Achieves 20 m/s obstacle avoidance, 90% success rate | Safety verification difficult, simulator-dependent |
| Hybrid DL + RRT* | CNN/LSTM + traditional planner | Safe adaptive path planning | Better interpretability, robustness | Compromise between autonomy and safety |
| Sequence-aware Transformer (SAT) | Transformer in multi-agent RL | Clustering cooperative decision-making | Improves task completion by 21–33% | High resource demand |
2.3 Rule-Driven Optimization Methods
Rule-driven methods incorporate operational rules, airspace regulations, and coordination protocols as hard constraints or optimization objectives within the planning framework. They ensure compliance and safety in real-world deployment. Key sub-categories include airspace rule embedding, centralized multi-agent coordination, and heterogeneous platform cooperation (e.g., truck-drone, drone-drone-station). The rules are formalized as constraints in mathematical models or as behavioral protocols learned through self-play reinforcement learning. For example, centralized coordination uses mixed-integer programming to encode collision-avoidance intervals and task exclusivity. Self-play RL enables UAVs to evolve cooperative strategies without explicit communication.
| Method | Technical Features | Problem Solved | Advantages | Limitations |
|---|---|---|---|---|
| Airspace Rule Embedding | Integrate no-fly zones, altitude limits | Compliance-ensured planning | Guarantees legal operation | Needs continuously updated database |
| Centralized Coordination | Mathematical modeling of collaborative rules | Multi-UAV conflict-free scheduling | Explicit, verifiable, globally optimal | Single point of failure, scalability |
| Heterogeneous Platform Rules | Task assignment, rendezvous, time synchronization | Truck-drone joint delivery | Exploits complementary strengths | Highly scenario-specific |
| Self-Play Reinforcement Learning | Game theory + RL | Cooperative adversarial decision-making | Evolves effective rules without manual design | Slow convergence, deployment validation needed |
3. Practical Applications
We have observed that UAV path planning techniques have been deployed in various real-world scenarios across China. In city instant delivery, Meituan’s “UAV + smart cabinet” model achieves an average delivery time of 12 minutes by integrating rule-driven airspace constraints with data-driven dynamic scheduling. For medical emergency delivery, SF Express’s Fengyi technology operates UAV blood transport between hospitals, reducing transport time by up to 80% through rule-based priority corridors and exact mathematical optimization. In rural areas, JD Logistics employs model-driven GA algorithms to optimize round-trip routes from town hubs to village stations, significantly lowering last-mile costs. Special scenarios such as island logistics and cross-border e-commerce benefit from rule-driven frameworks that harmonize disparate airspace regulations. These applications demonstrate the necessity of selecting the appropriate method—or hybrid combination—according to the specific scenario.
4. Technical Challenges
Despite promising results in simulations and pilot projects, several technical challenges hinder the large-scale deployment of UAV delivery path planning. First, real-time perception and robust decision-making in complex dynamic environments remain difficult due to GNSS signal loss in urban canyons and unpredictable moving obstacles. Second, efficient coordination of large-scale unmanned systems faces exponentially increasing computational complexity and communication delays. Third, model generalization and reliability suffer from the sim-to-real gap: algorithms trained in idealized simulations often degrade in real-world conditions with turbulence, sensor noise, and battery aging. Fourth, the fundamental trade-off between energy, payload, and range constrains commercial viability. Finally, regulatory and standardization gaps create uncertainty in operating procedures and liability assignment, slowing industry adoption.
5. Future Trends
Based on the current challenges, we identify five major development trends for UAV delivery path planning.
- Deep Integration of Artificial Intelligence: Data-driven methods (RL, GNN, Transformers) will increasingly fuse with model-driven optimization to create hybrid intelligent systems. For instance, GNN outputs can guide metaheuristic initialization, while RL can adaptively select operators.
- Full-Chain System-Level Collaborative Optimization: Path planning will evolve from isolated trajectory optimization to integrated planning with ground transport, warehousing, and communication networks. This requires high-reliability, low-latency communication infrastructure—exactly the focus of China drone development.
- High-Reliability and Resilient System Design: Future algorithms must guarantee safety under extreme disturbances through predictive fault detection and adaptive response, embedding energy-aware constraints.
- Standardization and Platform-Based Operation: Unified low-altitude logistics management platforms will digitalize airspace rules and provide shared interfaces for multi-operator fleet coordination, lowering entry barriers.
- Continuous Expansion of Application Scenarios: As technology matures, UAV delivery will penetrate cross-border e-commerce, industrial parks, and emergency humanitarian aid globally, creating a virtuous cycle of innovation and deployment.
6. Conclusion
In this review, we have systematically examined the state-of-the-art in UAV delivery path planning for the low-altitude economy, with a particular focus on China drone applications. We formulated the problem as a multi-objective, strongly constrained optimization, classified existing methods into model-driven, data-driven, and rule-driven categories, and provided quantitative comparisons via tables. Practical deployments by leading Chinese companies illustrate the real-world relevance. We also highlighted key technical challenges and outlined future directions toward more intelligent, collaborative, and resilient planning systems. As China drone technology continues to advance, the convergence of AI, communications, and aviation will unlock unprecedented logistics efficiency and societal benefits.