In recent years, the low-altitude economy has emerged as a transformative force, driven by advances in drone technology, digital infrastructure, and evolving airspace regulations. Unmanned aerial vehicles (UAVs) have transitioned from specialized military applications to ubiquitous civilian services, including logistics, agriculture, surveillance, and emergency response. Among these, drone delivery stands out as a cornerstone application, promising to revolutionize last-mile logistics by reducing transit times, operational costs, and environmental impact. The core technical enabler behind efficient and safe drone delivery is path planning — a multi‑objective, strongly constrained, and highly dynamic optimization problem. This article provides a systematic first‑person review of the state-of-the-art in drone delivery path planning. I begin by formulating the problem with explicit objectives and constraints, then categorize existing methods into model‑driven, data‑driven, and rule‑driven approaches, presenting detailed tables and mathematical formulations. Real‑world application cases are discussed, followed by an analysis of key technical challenges. Finally, I outline future trends that will shape the next generation of drone technology for low‑altitude logistics.
The rapid growth of the low‑altitude economy is underpinned by supportive policies. In China, the 15th Five‑Year Plan identifies low‑altitude economy as a strategic emerging industry, and multiple ministries have issued guidelines to accelerate infrastructure development, including a national standard system for low‑altitude transport. By 2027, the coverage of 5G networks for low‑altitude routes is expected to exceed 90%, providing a solid digital foundation for drone technology. Cities such as Shenzhen, Chengdu, and Hefei have been designated as eVTOL pilot cities, driving innovation in drone delivery. Against this backdrop, path planning becomes the linchpin for scalability, safety, and economic viability. In the following sections, I elaborate on how we, as researchers, model and solve this problem.
1. Problem Formulation for Drone Delivery Path Planning
Drone delivery path planning aims to find a feasible trajectory from origin to destination while optimizing multiple, often conflicting, objectives under a set of hard constraints. The primary objectives can be summarized as:
- Economic efficiency: Minimize total flight time or distance to reduce direct operational costs. This can be formalized as minimizing the total travel cost $$C_{\text{travel}} = \sum_{i} c(d_i)$$ where \(c(d_i)\) is the cost per unit distance and \(d_i\) is the segment length.
- Service quality: Ensure on‑time delivery within customer‑specified time windows, which is often modeled as a soft or hard time window constraint.
- Safety and robustness: Maximize the minimum distance from known risks (e.g., obstacles, no‑fly zones) and incorporate risk‑quantification terms in the objective function, e.g., $$\max \min_{p \in \text{path}} \text{distance}(p, \text{obstacle}).$$
The optimization is subject to three categories of constraints:
- Physical performance constraints: Maximum flight range, payload capacity, battery capacity, energy consumption model, and kinematic limits (maximum turn angle, climb rate).
- Task and environment constraints: Customer time windows, task sequence order (e.g., visiting multiple drop‑off points in a specific order), and coordination constraints for multi‑platform systems (e.g., UAV‑truck rendezvous points and time synchronization).
- Airspace and regulation constraints: Static obstacle avoidance (buildings, mountains), dynamic obstacle avoidance (other UAVs, temporary structures), and compliance with no‑fly zones, restricted airspace, and altitude limits.
The interplay of these elements makes drone technology delivery path planning a quintessential multi‑objective, strongly constrained problem. Table 1 summarizes the key dimensions.
| Dimension | Specific Element | Description |
|---|---|---|
| Optimization Objectives | Economic | Minimize total flight time/distance; optimize energy consumption model |
| Service quality | Meet customer‑specified delivery time windows | |
| Safety and robustness | Avoid risks; maximize distance from danger sources | |
| Physical Constraints | Maximum range | Single‑flight distance limit imposed by battery capacity |
| Payload capacity | Maximum weight of carried goods | |
| Battery & energy | Endurance limit from battery physics and flight energy model | |
| Kinematics | Turn angle, climb rate, etc. | |
| Task & Environment | Time windows | Hard or soft constraints for each delivery |
| Task sequence | Order of visiting multiple points must be optimized | |
| Cooperative | Temporal and spatial coupling in multi‑platform systems | |
| Airspace & Regulations | Static obstacles | Buildings, mountains, etc. |
| Dynamic obstacles | Other UAVs, moving cranes, birds | |
| No‑fly zones | Permanent or temporary restricted areas | |
| Regulatory constraints | Altitude limits, speed limits, operational rules |
Mathematically, a generic path planning problem can be written as:
$$
\begin{aligned}
\min_{\mathcal{P}} \; & J(\mathcal{P}) = w_1 \cdot \text{Time}(\mathcal{P}) + w_2 \cdot \text{Energy}(\mathcal{P}) + w_3 \cdot \text{SafetyPenalty}(\mathcal{P}) \\
\text{s.t.} \; & \text{Range}(\mathcal{P}) \leq R_{\max},\quad \text{Payload} \leq P_{\max}, \\
& \text{ObstacleClearance}(\mathcal{P}) \geq \epsilon_{\text{safe}},\quad \text{NoFlyZone}(\mathcal{P}) = 0, \\
& \text{TimeWindow}_k \leq t_k \leq \overline{\text{TimeWindow}}_k,\quad \forall k \in \text{customers}, \\
& \text{KinematicFeasibility}(\mathcal{P}) = \text{True}.
\end{aligned}
$$
Here, \(\mathcal{P}\) denotes the flight path, and the objective \(J\) is a weighted sum of time, energy, and safety cost. The constraints encode physical limits, spatial regulations, and service requirements. This formulation serves as the backbone for all solution methods discussed next.
2. Classification of Path Planning Methods
Existing path planning methods for drone delivery can be broadly classified into three paradigms based on their core decision‑making logic and reliance on prior knowledge: model‑driven, data‑driven, and rule‑driven. In the following subsections, I provide a detailed survey of each category, highlighting representative algorithms, their strengths, and limitations.
2.1 Model‑Driven Optimization Methods
Model‑driven methods rely on explicit mathematical modeling of the environment and the drone’s dynamics. They formulate path planning as an optimization or search problem and solve it using deterministic or heuristic algorithms. These methods excel when environmental information is complete and constraints are well‑defined, but they can struggle with high uncertainty.
2.1.1 Classical Graph‑Search Algorithms
The most foundational approaches model the free space as a graph, where nodes represent waypoints and edges represent feasible paths. The objective is to find the shortest or lowest‑cost route from start to goal. Dijkstra’s algorithm (1959) pioneered this field by greedily expanding the shortest path tree. A* (1968) introduced a heuristic function \(h(n)\) to guide the search, significantly improving efficiency. The total cost for a node is \(f(n)=g(n)+h(n)\), where \(g(n)\) is the actual cost from start. Over time, researchers have enhanced A* with bidirectional search, turning‑angle penalties, and post‑processing for smoothness. For example, an improved bidirectional A* reduces node expansion by up to 57% and turning points by 78% compared to the basic version. Another variant, Theta*, allows any‑angle connections, reducing total path cost by 30–50% while maintaining kinematic feasibility. These advancements are critical for drone technology in urban environments where sharp turns are prohibitive.
2.1.2 Metaheuristic Algorithms
Metaheuristics are population‑based global optimization techniques that simulate natural phenomena. They are well‑suited for multi‑objective, high‑dimensional, and nonlinear problems. Common methods include:
- Genetic Algorithm (GA): Encodes paths as chromosomes and uses crossover, mutation, and selection to evolve better solutions.
- Ant Colony Optimization (ACO): Models drone as ants that deposit pheromones on traversed edges, guiding subsequent agents.
- Particle Swarm Optimization (PSO): Each particle represents a candidate path; particles update velocities based on personal and global best positions.
- Improved versions: For example, an improved PSO with adaptive inertia weights reduces path length by 3.4% and path corners by 21% over standard PSO. A tuna swarm algorithm mimics tuna foraging and can handle up to 50 drones in cooperative delivery scenarios.
Table 2 summarizes key model‑driven algorithms with their performance metrics reported in the literature.
| Algorithm | Technical Feature | Problem Addressed | Advantage | Limitation | Scenario Scale | Solution Quality | Convergence Speed |
|---|---|---|---|---|---|---|---|
| Improved Dragonfly | Enhanced convergence precision for complex 3D problems | Urban 3D path planning | Better smoothness and obstacle avoidance | New algorithm, limited validation | Medium (≤50 UAVs) | Shorter path, high safety | Faster than original dragonfly |
| Bidirectional A* | Simultaneous search from start and goal | Real‑time single UAV path planning | Reduces node expansion, fast | Path quality may suffer | Single / small | Near‑optimal | <1s real-time |
| Improved Theta* | Allows any‑angle connections | Smooth path on grid maps | Shorter, smoother trajectories | Higher computational cost | Single / small | Shorter path, fewer turns | Moderate |
| Improved A* (penalty) | Penalty for turning angles | Static single‑UAV, smoothness | Enhanced flyability | Weak in dynamic environments | Single | High smoothness | Millisecond level |
| Improved PSO | Dynamic parameter adjustment | Single / multi‑UAV | Faster convergence, better quality | May trap in local optima in high‑dim | Small‑medium (≤30 UAVs) | 3.4% shorter, 21% fewer corners | ~150 iterations |
| Tuna Swarm | Simulates tuna schooling behavior | Large‑scale swarm delivery | High global search ability | Needs more comparative tests | Large (≥50 UAVs) | Small path cost | Fast, outperforms PSO |
| A*‑APF hybrid | Combines A* (global) with artificial potential field (local) | Dynamic obstacle avoidance | Global optimal + local reactive | Complex hybrid architecture | Single / small dynamic | Global optimum, local agility | Global offline fast, local real‑time |
| Mixed‑Integer Programming | Exact mathematical optimization | Multi‑UAV cooperative task assignment & planning | Theoretically optimal, rigorous | NP‑hard, small scale only | Small (≤10 UAVs) | Optimal given model | Hours (solver‑dependent) |
| Improved ACO | Optimized pheromone update and selection | Single‑UAV trajectory planning | Explores global optimum via positive feedback | Slow initial search | Small‑medium | High quality, good obstacle avoidance | ~300 iterations |
2.1.3 Artificial Potential Field (APF)
APF models the goal as an attractive force and obstacles as repulsive forces. The drone moves along the resultant force vector. While simple and effective for real‑time local planning, classic APF suffers from local minima and oscillations in narrow corridors. Modern approaches combine APF with a global planner (e.g., A* or RRT) to create a hierarchical framework: the global planner provides a coarse reference path, and APF performs local refinements. This hybrid design significantly improves robustness in complex urban settings.
2.1.4 Mathematical Programming and Control
For problems requiring provably optimal solutions, researchers formulate the path planning as a mixed‑integer linear or nonlinear program. The objective and constraints (energy, time, obstacle avoidance, etc.) are expressed in precise algebraic form. This approach is particularly valuable for vehicle‑drone cooperative delivery, where tight temporal and spatial coupling must be resolved. For instance, a mixed‑integer programming model for emergency supply distribution optimizes truck‑UAV rendezvous points and timing, achieving up to 41.4% reduction in service time compared to conventional methods. However, the computational complexity limits its applicability to small instances (e.g., fewer than 10 UAVs).
2.2 Data‑Driven Methods
Instead of relying on handcrafted models, data‑driven methods learn decision strategies directly from interaction data or historical examples. They excel in highly dynamic and uncertain environments where physical models are difficult to derive. The evolution has progressed from reinforcement learning (RL) to graph neural networks (GNNs) and end‑to‑end learning.
2.2.1 Reinforcement Learning (RL)
In RL, the drone acts as an agent that observes the environment state (position, battery, nearby obstacles) and chooses an action (e.g., velocity vector) to maximize cumulative reward. The reward function typically encourages goal reaching and penalizes collisions or delays. The problem is modeled as a Markov decision process (MDP). For drone delivery, researchers have used deep Q‑networks (DQN) and proximal policy optimization (PPO) to learn policies that handle delivery time windows, dynamic obstacles, and multi‑UAV coordination. A typical reward function is:
$$
R(s,a) = \alpha \cdot \text{GoalReward} – \beta \cdot \text{ObstaclePenalty} – \gamma \cdot \text{TimePenalty}
$$
Recent work extends RL to “informative path planning,” where the drone explores unknown environments to gather data efficiently. The advantage of RL is its adaptiveness; however, it suffers from sample inefficiency, poor interpretability, and sensitivity to reward design.
2.2.2 Graph Neural Networks (GNNs)
GNNs provide a natural way to model interactions among multiple drones or between drones and infrastructure. In the urban delivery network, each node (drone, warehouse, customer) is associated with features (position, load, time window), and edges represent spatial relationships or communication links. Through message passing and aggregation, GNNs learn a representation that captures the global structure. A multi‑agent GNN framework can then assign tasks and plan conflict‑free routes in a centralized‑training, decentralized‑execution manner. Experiments show that GNN‑based coordination significantly improves total throughput and reduces conflicts compared to traditional heuristics. However, the communication overhead during training and deployment remains a challenge, especially for large swarms.
2.2.3 End‑to‑End Learning
End‑to‑end methods aim to map raw sensor inputs (e.g., camera images, LIDAR point clouds) directly to control commands (throttle, yaw, pitch). A recent breakthrough uses a differentiable physics simulation framework to train a lightweight neural network that, given only low‑resolution depth images, can fly at 20 m/s through dense dynamic obstacles. This approach also enables multi‑drone cooperative flight without explicit communication. Nevertheless, pure end‑to‑end models are black boxes and require massive, high‑fidelity training data. To improve trustworthiness, hybrid models merge deep learning (for perception and context understanding) with classical planning algorithms like RRT* for safety guarantees. For example, a CNN‑LSTM‑Attention architecture extracts features and feeds them into an RRT* planner to bias sampling, achieving robust navigation even in unfamiliar environments.
Table 3 summarizes the data‑driven approaches.
| Algorithm | Technical Feature | Problem Addressed | Advantage | Limitation |
|---|---|---|---|---|
| Deep Reinforcement Learning | MDP formulation; reward‑driven policy learning | Online decision in dynamic, uncertain environments | Adaptive, handles complex dynamics | Poor interpretability, high training cost |
| Multi‑agent GNN | Message passing on graph; captures spatio‑temporal dependencies | Multi‑drone cooperative conflict resolution | Effective complex interaction modeling | Training complexity, communication demands |
| Differentiable Physics Framework | End‑to‑end from depth image to control | High‑speed dynamic obstacle avoidance and swarm coordination | Record agility (20 m/s), no‑communication cooperation | Difficult safety verification, high‑fidelity simulation required |
| Hybrid DL + RRT* | CNN‑LSTM‑Attention for perception + RRT* for safety | Safe adaptive path planning in dynamic environments | Interpretable, robust | Compromise between autonomy and security |
2.3 Rule‑Driven Methods
Rule‑driven methods incorporate operational regulations, airspace laws, and coordination protocols directly into the algorithmic framework as hard constraints or optimization criteria. They are essential for ensuring compliance, safety, and interoperability in real‑world operations.
2.3.1 Airspace Rule Embedding
The most basic rule is the absolute prohibition of entering no‑fly zones. Modern planning systems integrate digital maps of permanent and temporary restricted areas (e.g., over government buildings, military zones, or event‑based flight bans). These zones are represented as polygon constraints in the path optimization. For instance, the hybrid A*‑APF planner explicitly enforces that all waypoints lie outside static no‑fly zones and that the local replanning layer avoids any dynamic restrictions.
2.3.2 Multi‑Agent Cooperative Planning
When multiple drones operate in shared airspace, a set of “air traffic rules” must be enforced to prevent collisions. Two complementary paradigms exist:
- Centralized optimization: A mixed‑integer program formulates constraints like “no two UAVs may occupy the same airspace at the same time” and solves for a globally conflict‑free schedule. This gives the theoretical optimum but scales poorly.
- Distributed learning: Using a centralized‑training, decentralized‑execution RL framework, each drone learns a local policy that implicitly respects collision‑avoidance rules without real‑time communication. This approach is more scalable but may not guarantee absolute safety. Recent advances combine game theory with RL: self‑play reinforcement learning enables drones to autonomously evolve effective negotiation strategies for right‑of‑way at intersections.
2.3.3 Heterogeneous Platform Coordination
Real‑world delivery often involves trucks, ground robots, and drones working together. The coordination rules must define task assignment, rendezvous points, timing windows, and handover protocols. For example, in a truck‑drone cooperative model, the truck acts as a mobile base carrying multiple drones. The rules specify that the truck stops at predetermined locations to launch and recover drones, while drones serve nearby customers within their flight range. A mixed‑integer programming model for emergency logistics can reduce total weighted cost by 0.8–2.96% over fixed‑point strategies. In rural settings, the rules adapt to low population density and sparse road networks, allowing flexible rendezvous. In dense urban “drone‑drone‑station” networks, the rules become more complex, involving multi‑modal relay, hub transfer, and dynamic resource allocation. Table 4 highlights representative rule‑driven algorithms.
| Algorithm | Technical Feature | Problem Addressed | Advantage | Limitation |
|---|---|---|---|---|
| Airspace rule embedding | Integrate static/dynamic no‑fly zone data | Compliance with airspace laws | Ensures safety from source | Database must be maintained frequently |
| Centralized cooperative optimization | Mathematical modeling of task and collision rules | Multi‑UAV centralized planning | Explicit, verifiable, globally optimal | Single point of failure, limited scalability |
| Heterogeneous coordination rules | Task assignment, rendezvous, time synchronization | Truck‑UAV or multi‑platform delivery | Converts operational knowledge into computational rules | Highly context‑specific, needs custom design |
| Self‑play reinforcement learning | Game‑theoretic RL for autonomous conflict resolution | Cooperative adversarial control for drone swarms | Evolves effective rules without human pre‑programming | Slow equilibrium convergence, validation needed in realistic environments |
3. Real‑World Applications
Drone delivery path planning algorithms have been deployed in several operational contexts, each imposing unique demands on drone technology. I highlight four representative scenarios.
3.1 Urban On‑Demand Delivery
In dense cities, the challenges include complex static obstacles, dynamic airspace, and high order volumes. Companies like Meituan operate “drone + smart pickup locker” systems, achieving an average delivery time of 12 minutes. Their intelligent scheduling system blends rule‑driven airspace constraints with data‑driven order allocation and reinforcement‑learning‑based fleet coordination. This hybrid approach effectively handles the high‑frequency, highly dynamic urban environment. The success demonstrates that no single method suffices; rather, an integrated system combining perception, optimization, and regulatory compliance is required.
3.2 Medical Emergency Delivery
Medical logistics (blood, vaccines, emergency drugs) demands near‑100% reliability and strict time windows. Rule‑driven methods dominate: a “green channel” priority is established in airspace, and mathematical programming precisely models takeoff, flight, and handover times. For example, SF Express’s Fengyi drone service for hospital‑to‑hospital blood transport has cut transit time by up to 80%. The underlying system relies on a deterministic model of the physical environment and rigid rule enforcement.
3.3 Rural and Remote Delivery
In rural areas with weak ground infrastructure and sparse settlement, model‑driven methods excel because they do not rely on real‑time data networks. Metaheuristics like GA and ACO optimize long‑distance, multi‑stop routes, while energy‑consumption models are tightly integrated to maximize range given battery limits. JD Logistics’ “drone + rural agent” model uses such offline optimization to reduce last‑mile costs significantly, making delivery economically viable even in low‑density regions.
3.4 Special Scenarios and Networked Operations
Island, logistical park, and cross‑border corridors each have unique physical and regulatory constraints. Here, rule‑driven methods first define the operational framework (e.g., cross‑administrative airspace coordination, environmental protection rules). Then, model‑driven or data‑driven algorithms compute specific paths under those boundaries. The Yangtze River Delta island low‑altitude logistics tests, for instance, succeeded only after harmonizing airspace rules across multiple provinces. In large logistics parks, automated drone material handling follows pre‑defined internal traffic rules, with path planning being a controlled combination of deterministic models and local sensing.
4. Technical Challenges
Despite impressive progress, several fundamental challenges prevent the widespread deployment of drone delivery path planning. I analyze five key hurdles.
4.1 Real‑Time Perception and Robust Decision‑Making in Complex Dynamic Environments
GNSS signals are often blocked or degraded in urban canyons, causing localization drift that invalidates model‑based planners. Algorithms must fuse multi‑source information (e.g., visual odometry, IMU, cellular signals) to maintain accuracy. Additionally, sudden dynamic obstacles (birds, other UAVs, temporary structures) require fast online replanning and trajectory prediction. Current data‑driven methods show promising adaptiveness but still suffer from generalization gaps when deploying from simulation to reality.
4.2 Scalable Coordination for Large‑Scale Swarms
As the number of drones increases, the computational complexity of joint task allocation and path planning grows exponentially. Decentralized methods reduce the burden but introduce communication delays, bandwidth limitations, and topology changes. GNN‑based approaches can capture interactions, but training and executing them at scale remains non‑trivial. Designing a scalable, fault‑tolerant coordination architecture is an open problem.
4.3 Generalization and Reliability of Models and Algorithms
Many algorithms that perform well in specific simulated benchmarks suffer drastic performance drops in real‑world tests. The gap stems from unmodeled factors like wind turbulence, sensor noise, and variations in target characteristics. Future research must move beyond benchmark chasing and focus on robustness and lifelong adaptation.
4.4 Energy and Payload Endurance Contradiction
The fundamental trade‑off between battery capacity, payload weight, and flight range directly limits the commercial viability of drone delivery. While energy‑aware planning can mitigate this by optimizing speed and altitude, existing energy models are too idealized. Real‑time effects of temperature, aging, and wind need to be incorporated into the cost function.
4.5 Lagging Regulatory and Standards Frameworks
Technology outpaces regulation. Clear, digitized airspace management rules, UAV airworthiness standards, liability allocation, and privacy protections are still under development. Without stable regulatory inputs, an automated path planning system cannot guarantee compliance at scale. Hence, technical innovation must go hand‑in‑hand with policy evolution.
5. Future Trends
Driven by the challenges above, the field of drone delivery path planning is evolving along several promising directions. I highlight five key trends.
5.1 Deep Integration of Artificial Intelligence
The future lies in hybrid models that combine the strengths of data‑driven learning and model‑driven optimization. For example, a GNN can learn complex spatial dependencies and then guide a metaheuristic optimizer (e.g., improved artificial bee colony) to faster convergence. Reinforcement learning can adaptively tune the parameters of a classical planner in real time. This fusion marks a shift from isolated algorithm improvement to complementary intelligence.
5.2 Full‑Chain System‑Level Cooperative Optimization
Path planning will no longer be a standalone module but a core component of a unified cyber‑physical system that includes ground transportation, warehouse scheduling, and air traffic management. The optimization will span the entire delivery chain, from order assignment to final drop‑off, leveraging real‑time data and high‑reliability low‑latency communication networks such as 5G/6G. The result will be a global resource allocation that minimizes total cost and maximizes throughput.
5.3 High‑Reliability and Resilient System Design
Safety is paramount. Future planners will prioritize resilience over pure optimality. This involves integrating risk‑aware metrics, fail‑safe contingency plans, and dynamic energy management that adapts to battery degradation and wind conditions. Hierarchical frameworks like A*‑APF will evolve to predict and respond to sensor failures, communication blackouts, and other anomalies, ensuring that the drone can always return to a safe state.
5.4 Standardization and Platform‑Based Operations
Large‑scale commercialization requires standardized, shared infrastructure. Future urban air mobility will likely feature a unified low‑altitude logistics platform that provides dynamic airspace management, conflict resolution, and interoperability across operators. Such a platform will digitize all airspace rules, integrate them with path planning APIs, and enable on‑demand allocation of airspace resources. This will drastically reduce the barrier for individual operators and pave the way for drone technology to become as ubiquitous as ground traffic.
5.5 Expansion and Deepening of Application Scenarios
As path planning algorithms mature, they will unlock new high‑value applications: cross‑border e‑commerce, industrial supply chains, disaster response in remote regions, and even inter‑city parcel delivery. Each new scenario will challenge the algorithm with longer endurance, extreme weather, unfamiliar terrains, and diverse regulatory regimes. This virtuous cycle of “application pulling innovation, innovation enabling application” will drive drone technology to become the backbone of the next‑generation global logistics network.
6. Conclusion
In this first‑person review, I have systematically surveyed the landscape of drone delivery path planning from problem formulation to solution methods, real‑world applications, technical challenges, and future trends. The field is at a critical inflection point where algorithms are evolving from single‑domain solvers to integrated, intelligent systems. Model‑driven methods provide rigorous foundations; data‑driven methods offer adaptability; rule‑driven methods ensure compliance and safety. Their convergence will define the next generation of drone technology. By addressing the five major challenges—real‑time perception, scalable coordination, model generalization, energy constraints, and regulatory alignment—researchers and practitioners can unlock the full potential of low‑altitude logistics. The road ahead is demanding, but the rewards—faster, greener, and more accessible delivery for all—are well worth the effort.