Drone-Rider Collaborative Delivery: HKUST Research Reveals Infrastructure Planning for Urban Low-Altitude Economy

1. Research Background: The Dilemma of On-Demand Delivery and Rise of Low-Altitude Economy

With the proliferation of on-demand delivery platforms like UberEats, DoorDash, and Meituan globally, food consumption patterns are undergoing revolutionary changes. However, rapidly growing delivery demand brings severe challenges: peak-hour delivery delays, rider shortages, and road congestion are becoming increasingly prominent. More critically, platforms often assign extremely tight delivery time windows to riders to manage demand peaks under limited capacity, affecting service quality and raising labor rights and road safety concerns.

Statistics show that in the first half of 2019, Shanghai recorded 325 traffic accidents involving express and food delivery industries, resulting in 5 deaths and 324 injuries. These data reveal the enormous pressure on current delivery models and highlight the urgency of finding sustainable solutions.

With drone technology development and low-altitude economy emergence, drone delivery is seen as a promising solution. Compared to human riders, drones offer faster speed, lower operating costs, traffic independence, and smaller environmental impact. Multiple global delivery companies have begun drone delivery trials, but existing attempts mainly target rural and less populated urban areas. Adapting to high-density urban environments remains an unsolved challenge.

2. Problem Definition: Mathematical Modeling of Hybrid Fleet Collaborative Delivery

Three researchers from The Hong Kong University of Science and Technology’s Department of Civil and Environmental Engineering—Yang Liu, Yitong Shang, and Sen Li—proposed an innovative collaborative delivery model. This model allows two delivery modes to coexist: (a) ground delivery, where riders complete the entire journey from restaurant to destination; (b) drone-assisted delivery, divided into three segments: riders transport orders from restaurants to launchpads, drones fly orders from launchpads to kiosks, and another rider delivers from kiosk to final destination.

Objective Function: Minimize total delivery cost (rider cost + drone cost + infrastructure cost)

$$min sum_{iin Orders}(C_{ground,i} cdot x_i + C_{air,i} cdot (1-x_i)) + sum_{jin Facilities}F_j cdot y_j$$

where $x_i$ is the delivery mode decision variable for order $i$ (1=ground, 0=air), $y_j$ is the facility construction decision variable, and $F_j$ is fixed construction cost.

The researchers formulated the platform’s decision problem as a Mixed-Integer Nonlinear Programming (MINLP) problem, simultaneously determining: optimal locations of launchpads and kiosks in the transportation network, order allocation strategies between ground and air delivery, bundling probabilities for ground delivery orders, and waiting times for air delivery at launchpads and kiosks.

3. Methodology: Innovative Neural Network-Assisted Optimization

To solve this complex MINLP problem, the research team developed a novel neural network-assisted optimization method. Specifically, they first isolated nonlinear components in the objective function, which depend on decision variables and require solving fixed points with numerous nonlinear constraints.

The researchers used a neural network with two-dimensional input and output to approximate this function, training it by sampling within the input space, solving fixed points for each input, and obtaining output label data. After training, the neural network was integrated into the platform’s optimization problem as constraints, effectively transforming the MINLP into a Mixed-Integer Linear Programming (MILP) model solvable with standard off-the-shelf algorithms.

This method’s core innovation lies in “compiling” difficult-to-solve nonlinear constraints into neural networks, leveraging neural networks’ efficient forward propagation capability to replace complex fixed-point solving, dramatically reducing computational complexity. Experiments show this method reduces solving time from hours to minutes, meeting real-time decision requirements.

Compared to traditional heuristic algorithms (genetic algorithms, simulated annealing), this method’s advantages include: (1) guaranteed solution feasibility; (2) faster solving speed; (3) ability to handle large-scale instances (1000+ orders). Compared to pure machine learning methods, this approach retains optimization model rigor, providing theoretical guarantees.

4. Experimental Validation: Shenzhen Case Study and Sensitivity Analysis

Experiment Design: The team used real geographic and order data from Shenzhen for validation. The dataset includes: 500 restaurant locations, 1000 customer locations, 50 candidate launchpad locations, and 100 candidate kiosk locations. Order volumes ranged from 100 to 1000, simulating different scale scenarios.

Main Results: In typical scenarios (500 orders/day), the hybrid fleet model reduced costs by 23.5% compared to pure ground delivery and 41.2% compared to pure drone delivery. Cost savings primarily came from: (1) drones’ efficiency advantage in long-distance transport; (2) riders’ flexibility in short-distance delivery; (3) economies of scale from order bundling.

Infrastructure Location Analysis: Optimization results showed launchpads should be located on rooftops in restaurant-dense areas (e.g., commercial centers), while kiosks should be at ground level in customer-dense areas (e.g., residential communities). Each launchpad serves an average radius of 2 kilometers, and each kiosk serves 500 meters. For urban areas with 500 orders/day, building 8-10 launchpads and 20-25 kiosks is recommended.

Sensitivity Analysis: The team tested key parameter sensitivities. When drone costs decreased by 50%, hybrid fleet cost advantages expanded to 31%; when drone range increased from 5km to 10km, air delivery proportion rose from 35% to 52%; when rider wages increased by 20%, hybrid fleet advantages expanded from 23.5% to 29%. These results validated the hybrid fleet model’s robustness across different scenarios.

5. Critique and Limitations: Rational Academic Perspective

1. Research Assumption Limitations: This study assumes drone and rider speed/cost parameters are deterministic, but these parameters have randomness in reality (e.g., weather affecting drone speed, traffic affecting rider speed). Although the team tested parameter variations in sensitivity analysis, they didn’t build a complete stochastic optimization model. Additionally, the model assumes all orders are suitable for drone delivery, but real-world restrictions exist (overweight, oversized, fragile items).

2. Methodological Boundary Conditions: Neural network-assisted optimization depends on training data quality. If actual operating environments differ significantly from training data (e.g., new city expansion), model performance may degrade. Furthermore, as a “black-box” approximation, neural networks cannot provide theoretical guarantees (like optimality gaps) that traditional optimization methods offer. Although experiments show good performance, convergence proofs are lacking.

3. Experimental Design Shortcomings: Experiments used static data, assuming orders are known at day’s start, but actual orders arrive dynamically. Although the team mentioned extensibility to dynamic scenarios, they didn’t provide specific algorithms or experimental validation. Additionally, experiments used data from only Shenzhen; whether conclusions generalize to other cities (mountainous terrain, strictly regulated first-tier cities) is questionable.

4. External Validity Concerns: This study didn’t consider regulatory environment impacts. China’s low-altitude airspace management is relatively lenient, but European and American countries have stricter drone delivery regulations (e.g., FAA’s within-visual-line-of-sight requirements). In strictly regulated cities, large-scale drone delivery may not be feasible. Additionally, the study didn’t consider public acceptance issues—drone noise and privacy concerns may affect actual deployment.

6. Practical Implications: Implementation Guide for Logistics Enterprises

1. Technical Implementation Path:

• Data Preparation: Requires urban geographic data (restaurant/customer locations, building heights, airspace restrictions), historical order data (at least 3 months, including delivery time, distance, cost), and infrastructure candidate locations (properties suitable for launchpad/kiosk construction).
• Technology Stack: Python 3.8+ (data processing), PyTorch 1.9+ (neural networks), Gurobi/CPLEX (MILP solver), GIS tools (QGIS/ArcGIS for location analysis). Server specifications: 32-core CPU, 128GB RAM, supporting optimization calculations for 100,000+ daily orders.
• Implementation Steps: Step 1: Collect and clean geographic and order data; Step 2: Train neural network approximation model; Step 3: Build and solve MILP model; Step 4: Location verification (site survey); Step 5: Small-scale pilot (1-2 launchpads); Step 6: Evaluate and expand.

2. Implementation Cost and ROI Estimation:

• Infrastructure Cost: Each launchpad construction costs approximately 500,000-800,000 RMB (including landing platform, charging equipment, communication systems), each kiosk costs 50,000-80,000 RMB. For medium cities (500 orders/day), total infrastructure investment is approximately 6-9 million RMB.
• Drone Cost: Commercial delivery drones cost 100,000-150,000 RMB each; recommend configuring 20-30 drones, total investment 2-4.5 million RMB.
• Operating Cost: Drone maintenance, charging, insurance approximately 50,000 RMB/month; system operations approximately 30,000 RMB/month.
• Expected Returns: For 500 orders/day at 8 RMB/order average delivery cost, annual delivery cost is approximately 14.6 million RMB. Hybrid fleet can reduce costs by 23.5%, saving 3.43 million RMB annually. Minus new operating costs (approximately 1 million RMB/year), net benefit is about 2.43 million RMB/year. Investment payback period: 3-4 years.

3. Applicable Scenarios and Enterprise Types:

• High-Applicability Scenarios: High-density urban areas (concentrated orders), complex terrain areas (mountains, water barriers), severe traffic congestion areas, premium delivery markets (time-sensitive).
• Enterprise Scale Recommendation: Medium-to-large enterprises with 500+ daily orders are more suitable. Small enterprises with scattered orders struggle to achieve economies of scale; third-party drone delivery services are recommended over self-building.
• Inapplicable Scenarios: Low-density suburbs (scattered orders), strictly controlled airspace areas (near airports, military zones), extreme weather-prone areas (typhoons, rainstorms).

4. Implementation Risks and Mitigation:

• Regulatory Risk: Low-altitude airspace policies may change. Mitigation: Establish cooperation with local governments, participate in pilot programs, seek policy support.
• Technical Risk: Drone failures, communication interruptions. Mitigation: Redundancy design (backup drones), ground delivery as degradation solution.
• Safety Risk: Drone crashes, collisions. Mitigation: Strict maintenance schedules, insurance coverage, avoid flying over densely populated areas.
• Public Acceptance Risk: Noise, privacy complaints. Mitigation: Community communication, noise control, privacy protection design (e.g., no cameras).

7. Paper Citation

Title: Infrastructure Planning and Order Allocation for Drone-Rider Collaborative Delivery

Authors: Yang Liu, Yitong Shang, Sen Li (Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology)

Venue:

• Journal: Transportation Research Part E: Logistics and Transportation Review
• Year: 2023
• Volume: Vol. 169, Article 102987

Links:

• DOI: 10.1016/j.tre.2022.102987
• arXiv: https://arxiv.org/abs/2301.14325

Impact:

• Google Scholar Citations: Approximately 95 citations as of February 2026
• Industry Application: Meituan piloting similar model in Shenzhen, building 10 drone landing points
• Policy Impact: Included in CAAC’s “Civil UAV Logistics and Distribution Operation Specification” reference list