Decentralized deconfliction of aerial robots in high intensity traffic structures

Projections for future air mobility envisage intensely utilized airspace that does not simply scale up from existing systems with centralized air traffic control. This paper considers the implementation and test of a software and hardware framework for decentralized control of aerial vehicles within intensely used airspace. Up to 10 rotary wing vehicles of maximum all up mass of 1   kg 1\unicode{x0200A}kg are flown in an outdoor volume with length scale of 100   m 100\unicode{x0200A}m with GPS and WiFi connectivity. Flight control is implemented using a Pixhawk 4 flight controller running the PX4 firmware with guidance algorithms run on a separate onboard companion computer. Deconfliction is implemented using a simple elastic repulsion model with a guidance update rate of 10   Hz 10\unicode{x0200A}Hz . Traffic structures are constructed from a path of directed waypoints and associated cross sectional geometry. Junctions are implemented when two paths converge into one or when one path diverges into two. Agents engage with structures through execution of flow, merge and swirl velocity rules. Calibration experiments showed that the worst case latency in agents sharing position information was of the order of 0.5   s 0.5\unicode{x0200A}s made up from delays due to finite guidance update rate, WiFi processing and centralized message processing. A choice of vehicle cruise speed of 2   m ∕ s 2\unicode{x0200A}m\unicode{x02215}s and conflict radius of 2.5   m 2.5\unicode{x0200A}m provided an acceptable compromise between experiment time efficiency (speed) and spatial efficiency (resolution) within the test volume. Results from recirculating junction experiments show that peak deconfliction activity occurs at the junction node, however biased distribution of agents within a corridor means the peak intensity is pushed ahead of the node. Use of meshed helical junction structures significantly reduces the intensity of conflict at the expense of reduced junction time efficiency.


| INTRODUCTION
Anticipated development of urban air mobility and aerial drone delivery services benefits from intensely utilized airspace that does not simply scale up from existing transport systems with centralized air traffic control (Airbus, 2018;Bauranov & Rakas, 2021;Federal Aviation Administration, 2020;Federal Aviation Administration & NASA, 2020).
Current road transport systems use both centralized and decentralized management, for example, traffic lights and delegated driver autonomy, respectively, and vehicles flow through predefined networks of various classes of roads connected via junctions.Similar concepts are proposed for three-dimensional aerial traffic networks as in Jang et al. (2017), Nguyen et al. (2021), Quan et al. (2021), Sunil et al. (2015), and Tony et al. (2021).High intensity traffic is understood to be traffic in which the operational volume available to a vehicle is of the same order as the volume within which active deconfliction is required, that is, traffic in which the control effort needed to avoid collision becomes large compared to the control effort required for navigation in the absence of other vehicles.Traffic intersections are implicitly high intensity, however, any traffic flow may be considered high intensity if the vehicle density is sufficiently high.
The present work addresses the problem of developing and testing algorithms for automated following of aerial traffic structures whilst maintaining safe separation between agents (deconfliction).The work is based on development of a practical outdoor test environment using commercial off the shelf drones operated within line of sight regulations.
In this work control is decentralized in the sense that deconfliction algorithms run locally on the vehicles, while communications are centralized through use of a hub and spoke message broker.The aim is to progress the state of the art in deployment of laboratory-based aerial robotics techniques for multiagent aerial traffic flow experiments.
The main contribution of the paper is the documentation of the engineering methods required for successful field validation trials using decentralized aerial robots.A secondary contribution is documentation of development and application of novel structures for traffic intersection design.The following literature review is separated into three parts: the first considers frameworks for implementing a practical multiaerial robot experiment, the second uncrewed traffic management (UTM) simulation, and the third latency in multiagent communications.
A summary of existing frameworks developed to support practical aerial robotic applications is shown in Table 1.A framework is defined as a set of standards, libraries, and tools that integrates with one or more flight control hardware families.Telekyb (Grabe et al., 2013) provides a controller capable of implementing decentralized flight formations and supports trajectory planning, state estimation, and tracking.Mavwork (Mellado-Bataller et al., 2013) focuses on visual control of multiple micro aerial vehicles for indoor applications.Twirre (Van De Loosdrecht et al., 2014), was developed for vision-based autonomous flight of mini UAVs in both GPS-enabled and GPS-deprived environments.Paparazzi (Hattenberger et al., 2014) is an open-source drone hardware and software project for rotary and fixed-wing drones that supports popular autopilot firmware.In Preiss et al. (2017), an indoor practical demonstration of a swarm of 49 Crazyflie 2.0 drones with a Vicon positioning system was conducted.This work was extended with a software in the loop (SITL) framework enabling simulation and integration with robot operating system (ROS) and Gazebo (Silano & Iannelli, 2020).Aerostack (Molina et al., 2020;Sanchez-Lopez et al., Features Aerostack (Sanchez-Lopez et al., 2016) Paparazzi (Remes et al., 2013) Telekyb (Grabe et al., 2013)   apply a mathematical model to similar structures and introduce the concept of a rotary island which behaves similarly to a road traffic roundabout (Quan et al., 2021).When considering the lane itself, Mao and Quan provide a formal definition for the virtual tube (Mao & Quan, 2022) and Quan et al. present a method for containing agents within a virtual tube based on artificial potential fields (APF) (Quan et al., 2022(Quan et al., , 2023)).APF was originally described in Khatib (1985) where the goal produces an attractive potential and obstacles produce a repulsive potential, with the force applied to the agent being derived from the gradient of the field.
In this work, we define an aerial routing network comprised of paths (edges) and junctions (nodes).A path is defined both by its geometry and the guidance rules used to follow it.The geometry of a path is defined by a directed series of waypoints that form a centreline.A cross-sectional shape is swept along the centerline to form an envelope.A full theoretical description of these concepts is provided in Knox (2023).

| Guidance rules
A minimal implementation of a rules-based algorithm for which the output is a total goal velocity is chosen to demonstrate the proposed experimental framework.This method is similar to those employed in Crowther (2004), Khatib (1985), and Vásárhelyi et al. (2018) where velocity rules are used to produce flocking behavior.These methods can be considered as specific implementations of the more general APFs method.In the present case, rules are mathematical functions operating on locally derived field information for each agent.The principal distinguishing feature of these methods is that velocity behavior is obtained from weighted summation of responses from multiple sources (environment and other vehicles).This is in contrast to control centered approach in which the dynamics are solved explicitly.
Path guidance is provided by flow that governs the agent velocity parallel with the path centreline, merge which governs velocity normal to the centreline, and swirl that governs velocity normal to flow and merge (Figure 1).Rules outlined in this work assume the vehicles utilize F I G U R E 1 Definition of flow, merge, and swirl guidance velocities for agent path following.
| 1543 powered lift, for example, via rotary wing, that allows vehicle acceleration control independent of vehicle velocity vector.The rules would require some modification to work with fixed wing aircraft, which have inherently reduced controllabilty compared to rotary wing vehicles.
The total agent velocity demand v → t i is the sum of the velocities generated by these rules: where N is the number of agents, v → f i is the flow velocity, v → m i is the merge velocity, v → d i total is the total deconfliction velocity due to all other agents, and v → s i is the swirl velocity.The velocity demand v → t i is saturated with the following to produce the velocity command v which is output to the flight controller: where v max is the velocity limit.

| Flow rule
The flow rule is the same for all path types and is calculated using the unit direction vector d → i from the current waypoint to the next as shown in Figure 2.
Flow velocity is given by: where k f is the Flow velocity gain which controls the cruise speed.
This method differs from APF as implemented by Khatib (1985) in that the agent is not attracted to a goal location and instead moves along the path with a constant cruise speed.

| Merge rule
The merge rule serves to keep agents within the path geometry, performing a similar purpose to the barrier function used by Quan et al. (2022Quan et al. ( , 2023) ) to constrain agents within their virtual tube.The formulation of the merge velocity for each path type is given in Supporting Information S1: Appendix A. Formulas for the two path types implemented experimentally in this work, the cylinder and the helical tube, are given by ( 4) and ( 5), respectively: (5) The merge vector m → i is simply the shortest vector from the agent to the centreline of the path as shown in Figure 1.v mmax is the maximum per- mitted merge velocity, r cyl is the cylinder radius and r tube is the tube radius.

| Swirl rule
The swirl rule is used to produce helical motion for the helical tube path type.The magnitude of the swirl velocity is maximum when the agent is at the surface of the helical tube and gets smaller the further the agent is from the surface.The swirl velocity is given as follows: where k s is the swirl velocity gain.

| Deconfliction rule
Deconfliction is provided by a sphere around each agent defined by a conflict radius, Figure 3.This sphere behaves as a virtual contact surface that provides a repulsive velocity demand proportional to penetration with conflict spheres on other agents.The stiffness of the response is controlled by the deconfliction gain parameter k d .The maximum deconfliction demand is limited to v dmax .This implementation is consistent with the local repulsion described in Vásárhelyi et al. (2018).
The deconfliction velocity of agent i which separates agents i and j is given by: where x → ji is the distance vector from agent j to i. Additionally, the conflict distance, x → conflict ij shown in Figure 3, is equal Waypoints and direction vectors (Knox, 2023).

| Path types
Different types of paths are obtained by varying the topology of the path cross-section.A full description of available path geometries and respective rule implementation for these paths is included as Supporting Information S1: Appendix A. The two path types implemented experimentally in this work are the cylinder and the helical tube, the cylinder is consistent with the definition of a virtual tube provided in Mao and Quan (2022).
A binary junction is formed when two paths combine into one path (a converging junction) or one path splits into two (diverging junction).Junctions are not reversible, that is, the behavior of a converging junction is not the same as that of a diverging junction with a simple sign change.Arbitrarily complex intersections can be formed by the assembly of multiple binary junctions.

| Hardware setup
The system architecture used for this work is shown in Figure 4. • 868 MHz Receiver: TBS Crossfire Nano RX, provides a reserve data link through which agents can be controlled manually in an emergency, or a kill command can be issued.

| Ground control hardware
The ground control station (GCS) hardware includes: • GCS Computer: Runs the MQTT broker software, the experiment control program and QGroundControl • Router: Manages the passing of packets between the access point and the GCS Computer, also handles assigning IP addresses to devices on the network.
• Access point: Used to improve the range of the router network using high-gain directional antennas.

| Software setup
The software architecture used for experimental work is shown in

| Control method
The control algorithm in this experiment is a cascaded control method.The outer loop is the algorithm which calculates the desired velocities based on the rules outlined in Section 2.1.This loop can be considered a P controller due to the coefficients k k k , , f m s , and k d .The inner loop is the flight controller which uses a cascaded PID architecture to minimize the error between the current velocity of the drone and the velocity output from the rules.control system default of 3m∕s 2 .The latency in sharing position information between drones is comprised of a WiFi transmit delay, the MQTT message processing time and a WiFi transmit delay (Figure 11).Experimental evaluation of the system loop-back time (Figure 12), showed that there is a baseline processing delay of around 0.05s due to MQTT message parsing with an additional WiFi delay that increases proportional to the number of drones on the network (Figure 13).The worst case delay in reporting position information for two drones (out of a total of six) on a collision path was estimated based on a delay of 0.2s, due to information being two guidance frames late and a worst case loop-back time also of 0.2s, giving a total latency of 0.4s.At a cruise speed of 2ms −1 , the vehicle will travel 0.8m during this period.Stopping distance from 2m∕s is around 1.4m, verified by experiment, Figure 14  stopping required distance is around 1.4m + 0.8m = 2.2m.The vehicle physical hardware is bounded within a sphere of radius approximately 0.15m giving a minimum required deconfliction radius of 2.2m + 0.15m = 2.35m.On this basis, a conservative value of 2.5m was chosen for the conflict radius.This radius allows a maximum packing of 26 vehicles along the linear dimension of the test volume, which provides adequate spatial resolution for traffic flow experiments to be carried out.Choice of a faster cruise speed would have improved experiment productivity, but with adverse effect on spatial resolution.Choice of conflict radius here is conservative in that the head on collision case, whilst possible, is unlikely in practice because any slight offset in velocity alignment will cause drones to pass either side of each other.

| Experimental method
The maximum practical number of drones in an experiment is limited by increasing latency and/or decreasing bandwidth per drone.
Increasing latency increases the required conflict radius of each drone and hence reduces the physical number of drones that can exist conflict-free in the test volume.Bandwidth per drone has a minimum lower limit based on data packet size and the frame update rate.Figure 15 shows how the required bandwidth changes with number of drones assuming different guidance frame rates.Message rate is assumed to be quadratic with number of drones (all drones share position information).For a guidance frame rate of 10Hz and MQTT packet length of 50 bytes, the theoretical maximum number of drones based on bandwidth is around 116.In practice, the maximum would be less than this due to latency constraints at high drone count.Bandwidth constraints could be significantly reduced using a mesh network approach in which drone position was only transmitted to nearby agents.
The geometry of a typical flight experiment for assessing the performance of a junction is shown in Figure 16 2.
For safe operations, flight testing required a crew of a minimum of two people, with one person responsible for vehicle flight management and experiment control using the GCS, and a second person responsible for manual control of individual drones using dedicated radio control transmitters should an emergency arise that can not be managed by the GCS (Figure 18).Flight tests were undertaken in wind speed conditions from 0 up to 25km∕h (7m∕s), which represents maximum wind speed at which drones could reliably hold position.Tests were conducted at air temperature down to 0°C.However flight at low temperatures reduced experimental productivity due to temperature related reduction in battery capacity. the agent paths but otherwise trajectories are generally similar.Of note is that agents tend to travel on the outside surface of the corridor with respect to the rotation axis of the circuit.This leads to the mean paths from each circuit crossing ahead of the junction node, which is nonideal behavior from a deconfliction point of view.This issue is also related to relatively simple projected distance algorithm for deciding when a waypoint has been achieved.For this particular node, changing to increment upon Recorded deconfliction and merge velocities are significantly higher in the first circuit due to the influence of the use of agent starting locations that were not within the circuit.In subsequent plots, mean integrated velocities are presented based on circuits 2-6 only.
The effect of changing corridor radius on mean integrated deconfliction and merge velocities is shown in Figure 22.The trend is expected in that these quantities reduce with increasing corridor radius, which reduces the density of agents (reduces conflict) and relaxes the lateral manoeuvering required to stay within a corridor cross section.The trend with changing corridor radius is not linear, and in particular there is some feature of the combined geometry of

| Ensemble behavior verification
The correct ensemble behavior of the agents was verified by flying a number of different paths in simulation and in the real world.For the experimental setup used, there was a baseline delay of 0.05 s due to centralized message brokering that was approximately independent of the number of drones (up to 10 drones) and a WiFi delay that increased roughly proportional to the number of drones.
The mean latency with six drones was approximately 0.1 s with a standard deviation of 0.05 s.Further delays are introduced due to asynchronous update of guidance and communication frames at 10Hz giving a total worst case latency of around 0.4 s.The results of this work suggest that meshed helical structures could be used to reduce conflict in high-intensity UTM junctions.
Helical particularly in cases of very high traffic density where thebenefit of reduction in conflict outweighs the reduction in time efficiency.
Capitán et al. (2021) to develop in-flight deconfliction and control services, then implement and test them in customized configurations and scenarios, including use of automated threat management and conflict resolution.InCarramiñana et al. (2021) andBesada et al. (2022), an agentbased UTM simulator platform was presented to simulate the effects of availability of various UTM information sources and sensors on preflight and in-flight stages.A 2D agent-based Python simulation framework for low-altitude UTM systems including variable collision avoidance algorithms and useful definition of safety, capacity and efficiency metrics was introduced in Ramee and Mavris (2021).In Zhao et al. (2019), a multiagent air traffic and resource usage simulation framework was used to evaluate different air traffic management policies and obtain a relationship between policy, environment and resulting traffic patterns.Also of relevance to the present work are studies investigating latency in multiagent communications, which is a key driver of performance.The latency of a multiagent robotic system was measured for up to four agents for different message sizes in Berna-Koes et al. (2004).Communication back channels were proposed to decrease the latency.In Nguyen and Flueck (2011), a stochastic model which can be adjusted by system configuration to predict and simulate latency in multiagent power grids was presented.In Pasandideh et al. (2023) a systematic literature review on aerial robot networking was carried out which presents the limitations of existing flying ad hoc networks and identifies possible future work.While previous work identified the core components required for successful demonstration of distributed aerial control for UTM evaluation, the lack of an off-the-shelf solution for the specific research objectives required the development of a framework with bespoke elements.In particular, there was a need to implement custom on-board guidance algorithms and run efficient data gathering experiments with multiple drones.2| TRAFFIC MANAGEMENT IMPLEMENTATIONAerial routing networks have been proposed by many researchers as useful navigation structures for UTM.InJang et al. (2017) the authors propose structures that extend the conventional road system for aerial traffic management.These structures include lane and tube that lead into various types of intersections.InLow et al. (2014), counter flow corridors are proposed, similar to a dual carriageway, to allow the flow of traffic along main arterials.The corridors are proposed to be stacked vertically or horizontally.In(Airbus, 2018), corridors are proposed to be incorporated into a centralized system, where agents have predeconflicted paths to merge and unmerge from a corridor, but uses distributed control while in the corridor.In Challa et al. (2021), a method is proposed to define corridors to navigate around obstacles.Quan et al.
3.1.1| Aerial robot hardware Flight demonstration hardware was built around a custom five-inch frame racing quadcopter, Figure 5 and Figure 6.A weight breakdown for the vehicle is shown in Figure 7. Power was provided by a 14.8V, 3700mAh lithium polymer battery and propulsion by four 2206 2300kv motors running 5 in.propellers with a 3.5 in.pitch.Typical flight time for a standard configuration was 10 min.The principal avionics system components were: F I G U R E 3 Agent deconfliction model based on a virtual contact sphere.F I G U R E 4 System architecture diagram (Knox et al., 2022).CRANN ET AL. | 1545 • Onboard Computer: Raspberry Pi 4B with 2GB of RAM used to run high-level control software, receive mission commands from the GCS, and send commands to the flight controller through UART.• Flight Controller: Pixhawk 4 running PX4 Firmware used to provide low-level control and interfaces with the Mission computer via UART.• USB WiFi Communication Module: Dynamode 2.4GHz USB WiFi adapter, used to communicate telemetry information between agents and receive commands from the ground control program.Connected to the onboard computer via USB.• GPS Module: Ublox NEO-M8N, used to provide position, velocity and heading information to the Flight Controller.• MAVlink WiFi Bridge Module: Adafruit Huzzah with MavESP8266 firmware, used as a secondary control link to connect the Flight Controller with QGroundControl.

Figure 8 .
Figure 8.The software has three main components: Experiments were conducted at an outdoor flight test site within a working volume 130 × 130 × 110m box, Figure 10.Test volume was constrained by the requirement for line-of-sight operations and a maximum ceiling of 120 m under UK CAA regulations.These dimensions also allowed adequate coverage using a single WiFi access point.Choice of vehicle target cruise speed and conflict radius for the experiments was determined by consideration of the vehicle minimum stopping distance at maximum deceleration, including communication delays.Worst case is two vehicles approaching each other head on at cruise speed.A vehicle guidance update rate of 10Hz was chosen based on suitable compromise between perform- ance and stability.Vehicle maximum deceleration was set to the flight F I G U R E 8 Communication system and framework software (Knox et al., 2022).F I G U R E 9 Flow chart of operations performed on the onboard computer in one time step.CRANN ET AL.| 1547 , thus the total F I G U R E 10 Experiment test volume geometry and operational layout.Center of test volume located at 53.408336, −2.124308.F I G U R E 11 Sequencing diagram illustrating the time delays in transmission of a message between two agents.Diagram drawn to scale with one guidance frame representing 0.1s.

F
I G U R E 12 Drone to base station communication path showing how loop-back time is measured.F I G U R E 13 Average loop-back time results for 1-6 agents.The green line on each subfigure represents the mean, the dotted red line represents the standard deviation.The red bars in the bottom chart represent the results from a separate experiment where the message transmission delay of the MQTT broker was measured with different numbers of publishers and subscribers representing agents.
Figure 17.Experimental results involving velocity are nondimensionalised via the agent cruise speed and results involving length via the agent

F
I G U R E 16 Definition of test geometry for a recirculating junction experiment.Diverging node is due North of converging node.F I G U R E 17 Geometry of recirculating junction experiments to scale with the conflict radius of an agent.A front view of each corridor is shown with the maximum number of conflict radii which can fit in the cross-section.The corridor envelope is shown in yellow and the conflict radius is shown in blue.deconfliction radius.The variables recorded onboard each agent are given in Table

FlightsF
were not undertaken during precipitation.A typical flight operation involve the following steps: (1) Manually position drones in their ground start locations and power up, (2) upload experiment plans to each drone in form of JSON files, (3) Initiate drone take off and go to experiment start position, (4) Initiate flight experiment, and (5) Stop experiment and return drones to land at takeoff position and download log files.To simplify operations, a method was developed to safely position the aerial robots in their experiment start positions without relying on active deconfliction using only goto commands included in PX4, Figure 19.Agents are separated vertically by a pre-defined separation distance before flying to the desired x-y positions and finally moving to the desired altitude.4 | RESULTS 4.1 | Corridor radius A summary of the results from converging junction experiments showing trajectories and computed congestion (mean deconfliction velocity) for different corridor radii with the experimental parameters given in Table 3 is shown in Figure 20.Note that 0.8m∕s is used for the deconfliction velocity gain so that maximum deconfliction velocity is achieved when the agents are one conflict radius away from each other.The most significant region of congestion is around the node of the junction, as expected.Increasing the corridor radius increases the spatial distribution of T A B L E 2 Data acquisition variables recorded onboard each agent.I G U R E 18 Photos captured during experimental testing at Snowdonia Aerospace Centre.F I G U R E 19 Centrally controlled sequence to get drones safely into start positions without using active deconfliction (Knox et al., 2022).(a) Take off, (b) vertical separation, (c) fly to correct XY, and (d) fly to correct Z.
reaching the Easterly coordinate of the waypoint would bring forward the switching point of both left and right hand circuits and reduce the closing velocity at the point of intersection, and hence reduce the intensity of deconfliction required.Measurements of the integral of deconfliction and merge velocities for each of the six circuits in the junction experiments referred to above are shown in Figure 21.Integration is with respect to time over one circuit and is per agent.Velocities are presented in dimensionless form where a value of unity is a guidance demand velocity equal to the vehicle cruise velocity.

F
4 m radius case evident from Figure 20 that introduces additional deconfliction requirements.The cycle-averaged deconfliction velocity per agent as a function of distance through the junction is shown in Figure 23.The trends for the different corridor radii are similar with a small T A B L E 3 Parameters used in recirculating junction experiments.I G U R E 20 Paths of each agent through the recirculating junction experiment for corridor radii of 3, 4, and 5 m overlaid with deconfliction velocity.Deconfliction data only shown for deconfliction velocity greater than 0. Each loop contains three agents (total of six agents in experiment).Experiment is run for six complete circulations through the junction.Agent conflict radius was 5 m, shown to scale in inset.The yellow envelope represents the junction experiment test region.Gray is the return path.F I G U R E 21 Integrated nondimensional deconfliction and merge velocities for each circuit through the 3, 4, and 5 m corridor radii junctions.The first circuit has relatively higher deconfliction and merge velocities due to increased proximity at start and is not included in averaged quantities.peak in congestion around the waypoint at the entrance to the junction and a larger peak at the junction node.There remains evidence of the double peak around the junction node for the 3m radius case after cycle averaging, however this feature is less evident in the 4 and 5 m cases.The cycle-averaged merge velocity per agent as a function of distance through the junction is shown in 24.The merge peaks coincide approximately with the entrance waypoint and the junction node, but with the peak being slightly ahead at the entrance and behind at the node.The peak merge velocity is highest for the smallest corridor radius, consistent with the need to follow a tighter turn radius to stay within the corridor.A direct comparison of deconfliction velocity and merge velocity for the 3m corridor radius case plotted on the same axes is shown in Figure25.The fact that deconfliction peak is substantially ahead of the merge peak at the junction node confirms that it is the crossing streams effect ahead of the node that is causing the high deconfliction rate rather than the funnelling effect of the converged corridors.

A
Figure28, with comparison to an equivalent simple converging junction.The magnitude of separation velocity for the meshed helical junction is significantly less than that for a simple converging junction, as expected.However, for the helical junction, improved deconfliction does come at the cost of decreased transport efficiency as the helicity increases the effective path length and hence transit time through the junction.

Figure 29
Figure 29 shows the MITL simulation output side by side with a long exposure photograph of the flight test.There are some small differences due to real-world effects such as communication latency and sensor measurement error however the ensemble behavior is clearly correct.
A choice was made to operate experiments at cruise speed of 2m∕s, which required a deconfliction radius of 2.5m to satisfy worst case stopping criteria.A higher cruise speed for the same conflict radius would have allowed greater experimental productivity in terms of the amount of useful data that could be obtained between battery changes.With the present constraints, increasing the cruise speed also increases the required conflict radius and hence has neutral effect on experimental productivity.Mean latency could be reduced in future by adopting a mesh networking approach that reduces the centralized communication burden, however this may not address the peak latency, which is what drives the required deconfliction radius.The total number of drones flown simultaneously in the present experiment was a relatively modest six.From an available WiFi bandwidth perspective, it should be possible in theory to fly up to 116 drones with 10Hz communication frame rate in the same experimental setup however larger conflict radii or lower cruise speeds would need to be used due to increased latency.An experiment was successfully conducted to evaluate deconfliction within a converging junction between two traffic corridors with three vehicles continuously circulating on each of two different loops to improve data productivity, with experiments repeated for traffic corridors of different radius.The principal experimental measurements for each drone were position, deconfliction velocity demand and (corridor) merge velocity demand.Aggregate junction performance in terms of congestion was measured based on the integral of deconfliction velocity per drone per cycle through the junction.Integrated merge velocity per drone per cycle provided evidence of the equivalent impact of corridor following on guidance demand.Increasing corridor radius decreased congestion and aggregate merge velocity as expected.Due to the relatively low density of drones in each circuit and that turns were all of the same handedness, drones paths tended to be around the outside edge of the corridor for each circuit, resulting in path crossing just ahead of the junction node.This generated a region where drones had high closing velocity in close proximity and hence the deconfliction activity was high.The experiment could be improved in this respect F I G U R E 29 Long exposure photo of a flight test next to the output from MITL simulation to verify correct ensemble behavior.CRANN ET AL. | 1555 by adapting the merge rule logic to ensure a more uniform distribution across the corridor or by introducing corridor curvature of opposite sign to align the principal flow with the centreline.Evaluation of a more sophisticated junction concept using meshed helical corridors with implicit velocity alignment significantly reduces congestion compared to non velocity aligned junctions, albeit with agents taking more time to transition through the junction.A complete torroidal helical corridor was implemented experimentally as an example of building block for development of composite helical interchanges.
Millan-Romera et al. (2019) andameworks that implement UTM rules that are relevant to the present work.A simulation framework based on ROS and Gazebo was proposed inMillan-Romera et al. (2019) and