The Future of Drone Deliveries
Evaluating fair usage of airspace for drone delivery operations
Imagine opening your front door to see a drone gently lower your favorite fast food meal, just minutes after you placed your order.
This reality is getting closer each day. The future integration of uncrewed aircraft systems (UAS), commonly referred to as drones, into the US airspace is a priority issue for the Federal Aviation Administration (FAA) and the National Aeronautics and Space Administration (NASA).
Over the last two decades, the FAA and NASA have been developing concepts of operations to support future UAS operations. The envisioned UAS traffic management (UTM) and control system will use a community-based approach. Traditional air traffic management is centrally controlled by the FAA in a top-down fashion. In contrast, UTM will rely on negotiation and cooperation between operators themselves in a bottom-up design. Therefore, ensuring that all operators have fair access to the UAS airspace is a challenging (and unresolved) issue as UAS integration and commercialization becomes more widespread.
UAS Cooperative Airspace Simulation (UCATS™)
CNA developed UAS Cooperative Airspace Simulation (UCATS™), an agent-based modeling tool that simulates UAS flight planning in various scenarios to provide insights into airspace fairness for industry and government stakeholders.
What is Agent-Based Modeling?
Agent-Based Modeling (ABM) is a computational approach used to analyze the impact that agents have on a system. The agents in the system are autonomous and interact with each other. Each is defined by their own behavior, one that may lead to system conflicts later on. In the below visual, the agents are UAS operators with behaviors that are defined by their flight plans, trajectories, and conflict resolution methods. For UCATS™, the system impact is viewed through the lens of fairness.
Agent-Based Modeling is a computational approach used to analyze the impact that agents have on a system.
Defining Fairness
We define fairness as the relative extent to which an UAS operator is granted their desired flight plan when compared to other operators. Thus, fairness in our study is measured by the difference between the desired flight plan and the final filed flight plan. We consider three flight planning outcomes:
- As-planned = the filed flight plan is the same as the desired flight plan
- Re-planned = the filed flight plan differs from the desired flight plan
- Canceled flights = no flight plan can be filed
The flight plans for each delivery operation include the departure time and trajectory. If a flight plan is re-planned, the operation is delayed and has a later departure time than expected. In the future, UCATS™ may also consider changes to trajectory paths as another method of re-planning.
Our Research Questions
UCATS™ was developed to simulate the planning of UAS delivery operations and evaluate airspace usage fairness among operators. In particular, we aimed to answer the following questions using a series of scenarios:
- How can airspace fairness be measured for planning small UAS delivery operations?
- How will prioritization using a first-filed-first-served (FFFS) method affect airspace fairness?
Methodology
Our overall UCATS™ modeling approach followed five steps:
- Import ArcGIS Data for each operator, including address and sector coverage and pre-generated shortest path trajectories and trajectory lengths.
- Define Parameters for each operator, including number of flights, file-ahead choices, delay tolerance, and start and end times for operations.
- Generate Operator Flights by randomly selecting delivery address, departure time, and file-ahead time based on parameters. Then, order the flight operations by filed time to follow first-filed-first-served.
- Simulate Scenarios by running through each operation, attempting to file each flight plan into a tracking matrix, and recording the flight outcomes (i.e., as-planned, re-planned, canceled).
- Analyze Results by recompiling all results from parallel processing and calculating descriptive statistics.
We made a number of assumptions regarding our UCATS™ case study.
Location: Howard County, Maryland
We chose Howard County as our region of interest because of its dense suburban population. Howard County has a total of 102,029 addresses (residential and commercial) that we assume are eligible for delivery pending distribution location. For simplicity, we do not consider the effect of flight restricted areas such as those near hospitals or airports.
In addition, we assumed the year 2035 to calculate the expected population in Howard County and the corresponding number of deliveries for the simulations.
Two Package Delivery Warehouses
In this study, we simulated two fictitious package delivery warehouses. The "Elkridge Warehouse" is depicted in blue on the map near the southwest corner of the county. The "Ellicott City Warehouse" is depicted in orange on the map towards the northeast of the county. Both warehouses are located near major transportation hubs at the edge of the county border.
Warehouse Delivery Radii
Based on current drone capabilities, we assumed that each warehouse would serve addresses within a 10-mile delivery radius from the warehouse. As seen in the map, the delivery areas of the two distribution locations overlap significantly.
The two warehouses operated a total of 12,000 package deliveries over a 12-hour period. The number of baseline deliveries was calculated based on a conservative forecast for the year 2035. The deliveries are scheduled throughout the day with randomly determined file-ahead times (i.e., next day, same day, and one hour).
We assumed each warehouse location had an unlimited number of drones at their disposal to use.
Food Delivery Location
In addition to package delivery operations, we also looked at the effect of adding new types of delivery operations. We added three distribution locations for fast food delivery over lunchtime hours. These locations can deliver in a 3-mile radius and receive their orders 30 minutes ahead of time. Because of the time sensitivity of the orders, flights are canceled if delayed more than an hour.
We purposely chose the locations to overlap with each other and with the two existing package delivery services. The locations also correspond with actual food establishment hubs that serve surrounding residents in Howard County.
Based on existing literature, we assume that 47 percent of the eligible addresses order food at least once per week. We divide this percentage by 7 days to obtain the assumption that 6.7 percent of addresses order once per day. This meant that each fast food location deliveries 700 - 1,500 orders over lunchtime hours. We assume that the food deliveries only occur during lunchtime hours from 11 am to 2 pm.
Sector Approach
We divided the geographical region of interest into 300-by-300-foot sectors to represent the spatial aspect of the environment in our model. We assumed it would take 5 seconds for the drone to traverse a sector at 50 miles per hour.
Each sector contained zero or more addresses. The approximate housing density per sector was 5 to 10 single family homes, 10 to 15 townhomes, or 20-plus apartments. You can see an example of the address density of the sectors in the image to the right.
To reduce the computational effort of the model, we pre-generated one trajectory from each delivery address sector to each warehouse sector. The trajectories are calculated using a shortest path approach, where the drones can travel diagonally or orthogonally across the sectors.
Flight Plan Conflicts
A conflict occurs when an operator attempts to propose a plan that includes one or more instances in which the UAS will be in the same sector at the same time as another UAS, whose plan has already been submitted and filed. This conflict may occur at any point during the proposed operation.
This definition assumes that all UAS are operating at a single altitude. To allow for multiple launches and maneuvering from the distribution location, no conflicts are considered within a four-sector buffer around the distribution location sector.
Deconflicting Flight Plans
If a conflict occurs, the operator who files second is responsible for resolving the conflict. Our deconfliction rules implement a first-filed-first-served approach because flights are considered in the order that they are filed. Operators must deconflict with any filed flight plans before their proposed flight plan can be successfully filed.
Thus, deconfliction is managed at the preflight or planning stage and all deconfliction occurs at the ground (not air) level. The only acceptable method of deconfliction is to ground (delay) the flight plan until no conflicts remain with any other filed flight plans. Directional changes to flight trajectories and airborne holdings are not considered as deconfliction strategies.
The animation to the right shows how proposed flight plans are accepted as-is (green) if no conflicts occur or delayed (yellow) until no conflicts occur. Flights are canceled (see Flight #29 in the animation) if they are delayed past the last available departure time or past an operator’s defined tolerance for time-sensitive operations.
Baseline Scenario
We first used UCATS™ to simulate a basic scenario with the two package delivery operators at Elkridge Warehouse and Ellicott City Warehouse. Each operator delivered 6,000 packages over the course of 12 hours. After running 1,000 iterations of the simulation using randomized addresses, departure times, and file ahead times, we found that, on average, almost 70% of flights were delayed with an average of 60 minutes of delay, and that nearly 9% of flights were canceled.
Because of the stochastic nature of the model, it is also useful to look at the range of possible outcomes, including best and worst case outcomes. The worst case outcomes would represent instances when conflicts are more likely to occur (e.g., when a surge of deliveries are requested within a short amount of time or by addresses along the same delivery route). In the graphs shown below, we see that the range of re-planned flights (5.2%) is more consistent than the range of canceled flights (7.5%), which fluctuated more in the simulations. In extreme cases, the number of re-planned flights can be as high as 8,505 and the number of canceled flights can be as high as 1,681.
We can also look at the flight outcomes as they are filed over time. In the example below, the flight outcomes of one simulation iteration are animated in "real time," We can see that when the earlier flights are filed, a larger portion are filed as-planned. The portion of as-planned decreases as time goes on, so that eventually the majority of flights are delayed. Finally, at the end of the day, some flights are canceled.
We can also view the average flight outcomes based on the original desired departure time of the flight plan. A higher cancellation rate is observed for later departing flights due to the domino effect that is created by the system's compounding delays. Note the percentage of as-planned flights is similar for all departure times (~ 20%). This occurs because flights planned far in advance (e.g., 12 or 24 hours) are distributed throughout the day. Thus, operators that support later flights will have more difficulty filing their flight plans unless they are able to file far in advance.
Flights that depart at a later time are canceled more often than flights that depart earlier in the day.
Looking at the flight outcomes based on file ahead time, we can confirm that flights planned farther in advance have a clear advantage over flights planned closer to their departure time. Note that no flights planned 12 or 24 hours ahead were canceled in any of the 1,000 iterations. Thus, operators that support time-sensitive flights will experience more difficult in obtaining their desired flight plans.
Flights planned farther ahead are at an advantage over those that cannot be filed in advance.
Finally, we can also break down the results based on which sectors are experiencing the most conflicts. UCATS™ processes the trajectory for each flight plan in the order that it was filed. If an attempt to file a plan fails because an earlier filed flight already occupies the same sector at the same time, a conflict is recorded for that sector. The map shows the average number of conflicts per sector for all iterations. We found that the most conflicts occurred near the warehouses due to congestion at the landing and takeoff areas. Conflicts also occurred in “corridors” leaving the warehouse due to the shortest path algorithm used to generate the flight trajectories.
Most conflicts occurred near the warehouses and along popular corridors leading to or from the warehouses.
Other Scenarios
We've discussed the baseline scenario and its results, but, in the future, delivery traffic may vary depending on the day, drones may be equipped to fly with less separation (for better airspace utilization), and different types of operators may be sharing the same airspace. To evaluate these scenarios, we changed the parameters of our baseline scenario and used UCATS™ to investigate.
1. What happens if the volume of drone deliveries decreases or increases significantly?
Similar to traditional air traffic, the number of drone operations on a given day can vary for many reasons (e.g., day of the week, weather, seasonal impacts). Therefore, we ran our baseline scenario with total daily flights ranging from 6,000 to 24,000 and evaluated how the system responded to these changes. Prior to 9,000 flights, we found that there were no cancellations, due to the operators' ability to resolve a conflict with a delay instead. However, after 9,000 flights, the system reaches its capacity to manage delays before flight cancellations begin. The average minutes delayed and cancellations grow, until at 24,000 flights, the number of cancelled exceed those that are completed (with a delay). While the exact threshold numbers may vary based on assumptions, this type of analysis can help identify chokepoints where the overall system is stressed.
The overall system exhibits signs of stress at various chokepoints based on increasing levels of traffic.
2. What happens if drones are able to fly closer together?
In the future, advanced tactical deconfliction technology may allow drones to fly closer together without heightened safety risks. To evaluate this scenario, we analyze the outcome of allowing multiple drones per sector. This may be implemented by allowing drones to fly in different altitude layers in the same sector, and is akin to opening a new lane on the highway to increase the flow of traffic. The results of the simulations showed that flight outcomes improve as the number of drones per sector increases, because the number of conflicts is directly reduced. The portion of re-planned flights decreases linearly, while the average delay and portion of canceled flights are drastically reduced.
Increasing the allowed number of drones in the air directly improves the ability of the system to support flights.
3. What happens if other types of operators enter the airspace?
In the previous scenarios, we only considered operators of the same type: small package delivery entities. In the future, various types of operators will share the airspace and have different requirements and preferences. In this scenario, we added fast food delivery operators that behave differently. These operators only file 30 minutes before departure, service residents in a 3-mile radius, and are restricted to lunchtime operating hours (11 am - 2 pm). To see the affect of having the package and food delivery operators co-existing in the same airspace, we looked at the results with both types of operations occurring, and then with only the food delivery operations occurring.
In the bar charts above, we see that package deliveries take priority over food deliveries. In simulations without package deliveries, the food operators experience fewer delayed flights, almost no canceled flights, and less delay. Thus, a segregated airspace may help improve equity among operators that use short file-ahead times, such as fast food delivery services.
Fast food delivery operations (and other time-sensitive operations) are underprioritized when co-existing with operations that can file further in advance.
What's Next
We believe there are many applications for UCATS™. Because of its agent-based design, UCATS™ is highly customizable and can be used to model different drone operations and environments. The scenarios developed in this proof of concept demonstrate initial insights that can be further investigated. Our tool can be used in numerous possible applications to provide decision-makers with data-driven information. As the U.S. plans for wide-scale integration of drones into the National Airspace System, tools like UCATS™ can help stakeholders assess various scenarios and parameters.
Possible applications of UCATS™ for industry and government stakeholders
As the U.S. plans for wide-scale integration of drones into the National Airspace System, tools like UCATS™ can help stakeholders assess various scenarios and parameters.
If you enjoyed learning about UCATS™, click the button below to read about SAFER-C, another agent-based modeling tool that CNA developed to simulate the spread of a virus in correctional facilities.