See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/384926065

Modelling and Analysis of Autonomous Airport Surface Movement Operations

based on Multi-Agent Planning Explorative Case Study at Amsterdam Airport
Schiphol

Article in European Journal of Transport and Infrastructure Research · January 2025

DOI: 10.59490/ejtir.2025.25.1.7459

CITATIONS READS

0 139

2 authors, including:

Alexei Sharpanskykh
Delft University of Technology
167 PUBLICATIONS 1,715 CITATIONS

SEE PROFILE

All content following this page was uploaded by Alexei Sharpanskykh on 15 October 2024.

The user has requested enhancement of the downloaded file.

 1

Modelling and Analysis of Autonomous Airport Surface

Movement Operations based on Multi-Agent Planning
Explorative Case Study at Amsterdam Airport Schiphol

<author names>
<contact information>

Abstract
Both EASA and SESAR JU define a vision and roadmap towards an autonomous air
traffic management system. Furthermore, past and ongoing SESAR JU projects investigate
how to increase the efficiency and predictability of current operations by means of automa-
tion. In this paper, we explore the operational implications that result from fully-automated
airport surface movement operations modelled with high realism. A hierarchical multi-
agent system model was developed to coordinate and control all movements on the airport
surface. It comprises the Airport Operations Agent to handle the flight schedule and runway
configuration, the Routing Agent to compute conflict-free trajectories, and the Guidance
Agents to instruct and monitor the Aircraft Agents while these execute the planned routes.
The model incorporates the decisive processes and elements of airport surface movement
operations such as pushback, engine-start, inbound and outbound holding, compliance to
CTOT-slots, and wake turbulence separation for takeoffs. To compute conflict-free trajec-
tories for all taxiing agents, we tailored and extended state-of-the-art multi-agent motion
planning algorithms: the two-level routing algorithm combines Priority-Based Search (PBS)
with Safe Interval Path Planning (SIPP). We defined different sizes of aircraft, accounted for
a minimal safety distance between them, and calibrated their speed limits in curves with
historic ADS-B data. Using the real-world flight schedules of two of the busiest days at
Amsterdam Airport Schiphol, including different runway configurations, we examine the
performance of the autonomous taxiing system with respect to the historic operations. For
the considered simulation conditions, we show that the MAS yields 30% lower taxi times
that vary less and are more predictable, and increases runway capacity.

Keywords
multi-agent system; multi-agent motion planning; autonomous airport operations; airport
surface movement operations; automation; air traffic control

I. I NTRODUCTION
The air traffic demand is predicted to exceed 10 billion yearly passengers by 2050 (IATA, 2021),
more than twice the amount of 2019. However, it is expected that infrastructural expansions of
airports are insufficient to facilitate this growth (Eurocontrol, 2018). Therefore, large airports are
facing challenges to improve the efficiency of their operations and, on top of that, to reduce their
environmental footprint to achieve the industry-wide goal of net-zero emissions by 2050 (IATA,
2021).
When the congestion at airports increases, the taxi time of an aircraft, i.e. the time that it travels
over the airport surface from runway to gate or vice versa, becomes harder to predict. This may
affect the respective flight, but may also lead to network-wide knock-on effects (Eurocontrol,
2021). Moreover, when Air Traffic Control Officers (ATCOs) have to handle more and potentially
less predictable traffic, their workload is amplified (Chua et al., 2017). Consequently, the taxiing
operations may become less efficient.
To deal with these issues, previous and ongoing SESAR projects considered how to increase the
efficiency and predictability of taxiing operations through automation in general (MOTO, 2016;
TaCo, 2017). Other projects examined more specifically how to reduce emissions by integrating
engine-off taxiing techniques (AEON, 2021), or how to enable human-automation teamwork in
the operations through higher levels of automation (ASTAIR, 2023). Moreover, different aspects
of airport surface movement operations (ASM Ops) were studied in previous work: for example,
2

(Morris & Pa, 2016; Roling & Visser, 2008) studied automation within ASM Ops, (Atkin et al.,
2010) reviewed optimization approaches, (Weiszer et al., 2020; Zhang et al., 2019) investigated
how to create optimal trajectories under multiple objectives, and (Liu et al., 2014) assessed metrics
to evaluate predictability when automating airport surface movements.
In practice, the Advanced-Surface Movement Guidance and Control System (A-SMGCS) pro-
vides specifications for four services to increase efficiency through the use of automation: the
surveillance service to track vehicles, the airport safety support service to alert controllers of
potential conflicts, the routing service to determine conflict-free trajectories, and the guidance
service to guide the vehicles during taxiing (ICAO, 2004; Lane et al., 2020). Moreover, towards
2050 and beyond, both EASA and SESAR JU define a vision and roadmap to eventually reach
autonomous air traffic management, i.e. level 3 in EASA’s AI roadmap (AI Roadmap 2.0, 2023),
or level 4-5 in the roadmap of SESAR JU (Automation in ATM, 2020). However, the challenges
to achieve this long-term vision are manifold. For instance, the role of the human during and
beyond the transition as well as the implications of such fully-automated operations remain
largely unknown.
In this paper, we explore which operational consequences may result from autonomous surface
movement operations at large airports: in terms of efficiency, predictability, delays on taxiways,
and runway capacity, while sustaining safety levels. We represent such operations as multi-agent
system (MAS) model that plans conflict-free routes for all aircraft on the ground and controls
their execution. Multi-agent systems modelling and simulation allow for inherent modularity,
flexibility, and expressiveness of a system’s structure and dynamics. Both heterogeneous agent
properties and different types of interactions between actors as well as randomness can be
integrated into an agent-based model (Helbing & Balietti, 2015).
However, to evaluate the operational impact of such a new operational concept, the model
must represent the airport surface movement operations realistically. As pushback and engine-
start operations considerably effect delays in planning systems (Stergianos et al., 2016), we ex-
plicitly model these processes in the MAS control model. Furthermore, we include inbound and
outbound holding as these can considerably affect the taxi times of the respective flights: some
arriving aircraft must wait at a remote holding location until their stand is free, while some
departing aircraft must be delayed to comply with the Calculated Take-Off Time (CTOT) issued
by Eurocontrol. Moreover, the taxi times are strongly influenced by the chosen agent kinematics
as pilots speed up on straight segments, and slow down in front of curved taxiways. To this end,
we define a maximal and minimal velocity, and use speed limits along curves dependent on the
radius of curvature that we calibrate using historic ADS-B data. To the best of our knowledge,
this is the first study to use such detailed kinematics to compute trajectories of taxiing aircraft. We
thus provide an overview of values found in the literature when describing the model calibration
in Section II-D.
Airport surface movement operations must be safe at all times. To plan conflict-free routes
for a set of agents, many different multi-agent path finding algorithms have been developed
that model agents as a point (Stern & Sturtevant, 2019). However, to include agent shapes and
kinematics in path planning, the domain of multi-agent motion planning (MAMP) offers more
suitable algorithmic concepts (Cohen et al., 2019). We combine and extend such state-of-the-art
MAMP algorithms to form a routing algorithm that addresses the requirements of airport surface
movement operations. Section II-C summarises how it computes conflict-free trajectories for all
aircraft.
With an implementation of the MAS model in Python, we simulate the autonomous taxiing
operations using the flight schedule of two of the busiest days at Amsterdam Airport Schiphol to
date. Section III outlines the experimental setup. We then analyse and discuss key performance
indicators in relation to their historic counterparts in Section IV, and end with concluding remarks
in Section VI.

II. M ULTI -A GENT S YSTEM M ODEL

The developed multi-agent system (MAS) model for autonomous aircraft taxiing operations has
a distributed-hierarchical structure of both centralized and distributed agents, which is illustrated
in Fig. 1. The centralized Airport Operations Agent defines and updates the flight schedule and
runway configuration, the centralized Routing Agent plans conflict-free trajectories for all Aircraft
 3

Agents which are instructed and monitored by distributed Guidance Agents while executing their
planned routes. For the autonomous operations considered in this paper, it is assumed that the
full control and decision making is done by the agents. Furthermore, we assume that digital
means of communication via a datalink such as AeroMACS as well as the surveillance service
of the A-SMGCS specification are fully operational.

FS + RMO
Airport
Routing
Operations
Agent
Agent
trajectories

Aircraft Guidance
Agents Agents

instructs + monitors
Figure 1. Overview of multi-agent system for autonomous airport surface movement operations

A. Model specification
The airport taxiing infrastructure is represented by a graph G = (V, E) comprising vertices V
and directional edges E. As example, the layout of Amsterdam Airport Schiphol, which is also
used in the simulations presented in this paper, is shown in Fig. 2. Vertices denote aircraft stands
(green), taxiway intersections (black), holding points (orange), or stopbars in front of runway
entries (red). Each bidirectional taxiway segment between two vertices is constructed from two
unidirectional edges that connect the vertices. Taxiway edges (black) are obtained from the actual
locations of these taxiways at Schiphol.
The Airport Operations Agent schedules all flights, and updates them whenever new predic-
tions of the underlying A-CDM milestones are available. When the allocated stand of an arriving
flight is still occupied by a departing aircraft, or Eurocontrol issued a Calculated Take-Off Time
(CTOT) for a departing aircraft, the agent marks the corresponding flight. Such flights are subject
to special routes assigned by the Routing Agent to account for the necessary holding, detour,
or prioritization during taxiing, as outlined in Section II-B. Furthermore, the Airport Operations
Agent defines the runways in use, i.e. the runway mode of operation (RMO). Active runways
and the resulting flight path of arriving or departing flights must not be crossed. Thus, the
Airport Operations Agent blocks such taxiway segments by setting layout constraints on them.
This mechanism is also applicable for taxiway segments that are temporarily unavailable.
Both the flight schedule and constraints are shared with the Routing Agent that computes
conflict-free routes for all taxiing aircraft within the upcoming planning window wplng . It re-
computes the routing plans when it receives updates from the Airport Operations Agent, or
latest after the re-planning period hplng has passed. We use motion planning to account for
vehicle kinematics and shapes in planning. To ensure conflict-free paths, we deploy a two-level
search based on Priority-Based Search (PBS) (Ma et al., 2019) with an augmented version of the
Safe Interval Path Planning (SIPP) algorithm (Phillips & Likhachev, 2011). This routing algorithm
is presented in Section II-C, and we describe how we calibrate the model in Section II-D.
The resulting trajectories are sent to the Guidance Agents which are positioned at every
intersection in the taxiway system. Each Guidance Agent controls those Aircraft Agents that are
moving towards its location. It instructs them to execute the next part of the planned trajectories,
and monitors that the instructions are carried out accordingly. To do so, the Guidance Agents use
the airport radar, which reports the position, speed, and heading of all Aircraft Agents while they
move over the airport surface. In case the executed movements deviate from the planned routes,
the Guidance Agents locally adjust the trajectories to minimize these deviations. However, when
the impact becomes too extensive, they request central replanning from the Routing Agent. Once
one of the Aircraft Agents has passed the location of a Guidance Agent, it passes the guidance
responsibility for that aircraft to the next Guidance Agent along the aircraft’s route.
4

18C
TWY-Y

36L

18L
TW
Y-
V

W5
TWY-D

TWY-C

TWY-A

24
TWY-Z

TW B
Y-
36C Y-
C TW

36R
06

Figure 2. Graph of layout of Amsterdam Airport Schiphol with edges for runways (grey) or taxiways (black),
and vertices for taxiway intersections (black), stopbars (red), holding points (orange), or gates (green)

Aircraft Agents represent the aircraft (auto-)pilots and are modelled to be fully cooperative:
they thus carry out the instructions as accurately as possible. To account for the different sizes
of aircraft, all flights are categorized as one of the 6 aircraft types from the ICAO aerodrome
reference codes (ICAO, 2016). They are assumed to have a circular shape with a pre-defined
radius according to the type. Table IIb in Section III lists these shape-radii.

When planning the trajectories, a safety zone is added around all agents. To this end, we
define a general safety distance, as well as a safety distance that an agent has to keep when
it is trailing another aircraft. Both safety measures are defined in relation to the shape radii of
the corresponding pair of agents. Moreover, two aircraft that consecutively take off from the
same runway must have a minimal separation to mitigate the wake turbulence of the preceding
aircraft. We use the time-based separation minima from RECAT-EU for that (Rooseleer & Treve,
2018).
 5

B. Activity sequence of Aircraft Agents

To take the various surface movement operations into account during path planning, the route
of an Aircraft Agent is expressed as a combination of the following three activities:
• Go-to activities have one start vertex and a set of goal vertices. Thus, the routing algorithm
gets two degrees of freedom: the path between the vertices, and the time to traverse this
path. The regular taxiing between one point to another point at the airport is an exemplary
go-to activity.
• Follow activities comprise a predefined ordered list of edges that must be part of the route.
Therefore, during routing, time is the only remaining variable as the path cannot be changed.
Pushback and push-pull manoeuvres of departing aircraft are examples of such.
• Wait activities define a vertex at which an agent has to wait for a fixed duration. For instance,
a wait activity is used to specify the place at which the pushback-truck is decoupled from the
aircraft, or the necessary direction-switch of the push-pull manoeuvre within the pushback
operations occurs.
Using a combination of these activities, the Routing Agent defines an activity sequence for both
departing and arriving aircraft, as depicted in Fig. 3.

ACTIVITIES follow go-to wait

RUNWAY

regular pushback via

decouple
runway exit (regular or rapid-exit) to stand

STAND
taxiing
AIRCRAFT

engine cool-down
ARRIVING

RUNWAY

inbound via runway exit to holding point (HP) hold to stand

STAND
holding engine cool-down

RUNWAY
regular pushback decouple hold to runway entry (RWY)
STAND

taxiing
DEPARTING

engine-start
AIRCRAFT

RUNWAY
outbound hold pushback decouple hold to HP hold to RWY
STAND

holding engine-start

Figure 3. Activity sequence for regular taxiing of arriving and departing aircraft as well as inbound and outbound
holding. While engine-start (orange box) of departing aircraft is accounted for, engine cool-down (dotted orange
box) is neglected.

1) Engine warmup and cooldown: In the sequence, the warmup and cooldown of the engines
represent special cases. The routing algorithm takes the warmup-phase as part of the engine-start
manoeuvre and on basis of the aircraft-specific engine-start duration as input value into account.
Therefore, if this duration exceeds the time needed till decoupling from the pushback-truck,
additional waiting in form of holding is added to the route. We do not model engine cooldown,
as it does not have an influence on the routing regarding the kinematics, since the engines are
switched off after standstill at the gate.
2) Inbound holding: When an aircraft arrives at the airport, but its stand is still occupied by
a departing flight, the Routing Agent has three options to resolve the anticipated stand-conflict:
for long conflict durations (case 1), it sends the arriving flight to the remote holding platform
(cp. Fig. 2). Otherwise, it defines a detour along the taxiways (case 2), or reduces the agent’s
taxi speed for short conflicts (case 3). To this end, the Routing Agent first calculates the single-
agent route directly to the stand, i.e. the trajectory without accounting for other aircraft agents,
to estimate the severity of the stand-conflict. Then, it computes a single-agent trajectory via the
remote holding points. When this detour is insufficient to resolve the stand-conflict, the Routing
Agent assigns the remaining time as remote holding duration (case 1), and updates the agent’s
activity sequence accordingly. In contrast, when the taxi duration now exceeds the time at which
6

the departing aircraft has cleared the stand (case 3), it keeps the original activity sequence of the
agent.
3) Outbound holding to comply with CTOT-slots: Similar to inbound holding, the Routing Agent
deals with departing flights for which Eurocontrol issued Computed Take-Off Times (CTOT-slots).
However, as long as no arriving flight requires the stand, it assigns a holding duration at the
agent’s stand so that the agent arrives at the runway at the beginning of the CTOT-slot. In case
an arriving flight is scheduled for the stand, the Routing Agent sends the departing flight to a
remote holding location close to the scheduled runway. It updates the activity sequence of the
departing flight accordingly.

C. Routing algorithm
The Routing Agent carries out multi-agent motion planning for all Aircraft Agents that taxi
within the planning window. This two-level routing algorithm uses a low-level search to calculate
individual trajectories per aircraft, and coordinates all agents in its high-level search to yield
conflict-free trajectories. For the low-level, we extended the Safe Interval Path Planning (SIPP)
algorithm (Phillips & Likhachev, 2011), and adapted the Priority-Based Search (PBS) algorithm
(Ma et al., 2019) to serve as high-level solver.
PBS constructs a priority order between agents to deconflict their space-time trajectories. In
its priority tree, each parent-node has up to two child-nodes. Thus, a priority-relation between a
conflicting pair of agents is established. In each child-node, one additional priority-pair is added
with which one of the two agents that were previously in conflict must give way to the other
agent along its entire route. Then, PBS checks the child node that has the lowest sum-of-cost of
all agent trajectories for conflicts between those agents that do not yet form a priority-relation
with each other. We define the cost of a trajectory as sum of the taxiing duration and travelled
distance. Once a child-node is expanded without any collisions, PBS returns the resulting conflict-
free trajectories.
In the low-level search, the route of a deprioritized agent has to be adapted, either by changing
its path or altering the speed profile along the path. To this end, we translate all paths into a
set of graph reservations: an aircraft temporarily blocks a set of edges during each movement
between one vertex and another. The blockage times and set of blocked edges are dependent on
the agent’s shape, velocity profile, the shapes of other agents, and the safety zone between the
shapes.
The SIPP algorithm represents moving obstacles as collision intervals and subsequently defines
a set of Safe Intervals (SIs) per graph location, representing time intervals during which an agent
can occupy that location. Furthermore, states are defined on vertices and motion profiles with
piecewise constant acceleration map the trajectory between states. We augmented SIPP to facilitate
the activity sequence of an aircraft as defined by the Routing Agent, and to take the travelling
direction as well as the kinematic agent properties into account. Additionally, we use SIs also on
edges to deal with the reservations of agents higher in priority.
In the motion generation, we are bound to the agent’s kinematic properties for the current
activity and the velocity in the current state. A motion that is part of the follow-activity for
pushback is for example constrained by a lower maximum speed than regular taxiing in a go-to
activity. In addition, vehicles that have maximum velocity in the current state, might not be able
to decelerate enough to satisfy a reservation on the next edge or vertex. In this case, it might be
required to start decelerating on the edge before the current state. To efficiently account for this,
we anticipate based on the agent’s current velocity, braking distance, and reservations or velocity
restrictions within the braking distance.

D. Model calibration
In the MAS model, the agents’ motions during route planning are modelled based on constant
longitudinal acceleration/deceleration and do not account for slip, i.e. are steady-state motions.
Nonetheless, to the best of our knowledge, this is the first study to use such detailed kinematics
to compute trajectories of taxiing aircraft. In the following, we thus include an overview of related
values found in the literature.
We define a general speed limit of 15 m/s in line with the design taxi speed given in the A-
SMGCS manual from ICAO (2004). Except for the dedicated wait-locations, agents must taxi at
 7

least with the minimal velocity of 1.5 m/s to avoid stop-and-go during taxiing. For curved taxiway
segments, the ICAO manual mentions that speeds up to 10 m/s may occur. Most previous studies
on airport surface movement operations define curved segments as turns with a maximal velocity
of 5 m/s (Bakowski et al., 2015; Chen et al., 2015; Udluft, 2017). Since we model taxiway curves
explicitly, we define a speed limit vcurve per edge by using
√
vcurve = alat ∗ rcurve (1)
with the lateral acceleration alat and the radius of curvature rcurve of the respective edge.
For passenger comfort in public transport, Bae et al. (2022) provide a range for both longitudinal
and lateral accelerations of ±0.9 m/s2 . Furthermore, they claim that a car driver with a normal
driving style experiences a lateral acceleration of up to ±4 m/s2 and a longitudinal acceleration
of −2 m/s2 to 1.47 m/s2 . In contrast, De Winkel et al. (2023) found in empirical studies that the
acceptable limits for passenger comfort are 1.23 m/s2 for longitudinal and 0.98 m/s2 for lateral
acceleration. As noted above, previous studies did not consider lateral accelerations to define turn
speeds. For longitudinal acceleration/deceleration, different values are reported: for example,
±0.98 m/s2 (Bakowski et al., 2015; Chen et al., 2015), or 0.26 m/s2 as acceleration and −0.8 m/s2
as deceleration (Udluft, 2017).
To find realistic values for the longitudinal and lateral accelerations of taxiing aircraft, we use
historic track data from Schiphol captured by ADS-B receivers that record the aircraft positions
during taxiing with a rate of 1 Hz. To this end, we map the positions onto the graph representing
the taxiway centerlines, and smooth the resulting trajectories with a Savitzky-Golay filter (window
length of 11 s, linear polynomial). This yields the travelled taxi distance along the graph edges
as well as the speed and acceleration at each time point of the trajectory. However, we only use
the data on the edges that correspond to the main taxiways: while the tracks become too noisy
in the bay areas and at aircraft stands, the accelerations on runways for takeoff and landing are
not representative of those experienced during taxiing.
In Fig. 4, the acceleration over velocity of each data point is visualized as 2d-histogram.
The 1% and 99% percentile lines of the acceleration values per 0.5 m/s step show that the
longitudinal acceleration/deceleration remain similar across different taxi speeds. Therefore, we
set the acceleration to 0.4 m/s2 and deceleration to −0.5 m/s2 independent of an agent’s speed.
While these values seem low compared to those mentioned in the literature, we argue that using
these in planning increases the flexibility during execution: the Guidance Agents have more
options to locally adjust the trajectories if necessary (cp. Section II-A).

Figure 4. Calibration of longitudinal acceleration and deceleration values with historic track data. Comfort limit
obtained from (Bae et al., 2022)

Fig. 5 visualizes the historic curve speeds of different aircraft types as average speed along a
curved edge with radius r. The average speed is calculated as v̄ = ∆d ∆t
with the time difference
∆t and travelled distance ∆d of the data points per edge along each trajectory. Although higher
curve speeds exist, we define the speed limit in curves based on a lateral acceleration of 1.5 m/s2
for small aircraft (left plot in Fig. 5) and 1.125 m/s2 for large aircraft (right plot). Using Eq. (1),
we assign a speed limit per edge within the velocity bounds of 1.5 m/s to 15 m/s.
 8

,&$2$_,&$2%_,&$2& ,&$2'_,&$2(_,&$2)




VSHHG>PV@


 KLVWRULF,&$2'
KLVWRULF,&$2% KLVWRULF,&$2(
 KLVWRULF,&$2& KLVWRULF,&$2)
 OLPLWWKLVVWXG\ OLPLWWKLVVWXG\
YHORFLW\ERXQGV YHORFLW\ERXQGV

         
UDGLXVRIFXUYDWXUH>P@ UDGLXVRIFXUYDWXUH>P@
Figure 5. Calibration of curve speed with historic track data

Table I summarizes the kinematic values and lists the main algorithmic parameters used by
the routing algorithm. In general, two aircraft agents have to keep a minimal safety distance
between them equal to the average of their shape radii. However, when an aircraft is trailing
another agent, it has to keep a safety distance of at least 3-times the shape radius of the preceding
aircraft, which is in accordance with experts. The planning window wplng and replanning period
hplng are provided as ranges with the requirement that hplng < wplng .

TABLE I. K INEMATIC AND ALGORITHM PARAMETERS THAT ARE USED IN THE ROUTING ALGORITHM

parameter value unit

maximal speed vmax 15 m/s
minimal speed vmin 1.5 m/s
curve speed vcurve per edge m/s
acceleration acc 0.4 m/s2
deceleration dec -0.5 m/s2
safety distance in general 1 averaged shape radius
safety distance trailing 3 shape radii of preceding aircraft
planning window wplng 15 to 60 min
replanning period hplng 5 to 50 min

E. Verification and validation

Verification and validation of the simulation model were performed in accordance with vali-
dation techniques and tests as described by Sargent (Sargent, 2011). To validate the conceptual
model, operational experts from Amsterdam Airport Schiphol were consulted. During implemen-
tation, continuous verification was performed. The model was developed in different modules,
allowing for the independent testing of the building blocks. In addition, assertion conditions
were added to ensure correctness of the internal processes in the code and compiler errors
were resolved. With visual animations, we verified that the routes were executed as planned.
Furthermore, small test scenarios were created to verify the model’s behaviour in the bay areas
for pushback, push-pull, and engine start manoeuvres. The activity-based path planning was
verified with small test scenarios ensuring correct timings and kinematics. With these scenarios,
face validation was performed to ensure that the model performance was as expected. Finally,
individual agent behaviour was carefully followed throughout the system to ensure correctness.
As safety must not be compromised, we confirmed that indeed no collisions between agents
occurred during the execution of the planned conflict-free routes.
 9

Figure 6. Exemplary historic, un-calibrated, and calibrated speed profiles of a departing flight travelling from
stand C14 to runway 36L

As validation of the calibrated agent kinematics, Fig. 6 shows an exemplary historic speed
profile over travelled distance (black line) as well as simulated single-agent trajectories along the
same path for both the un-calibrated agent kinematics used in (von der Burg & Sharpanskykh,
2023) (green line) and the calibrated ones of this work (orange line). Especially on the long straight
taxiway segments towards runway 36L, the historic trajectories indicate that aircraft often taxi
faster than the velocity limit of 15 m/s.

III. E XPERIMENTAL S ETUP

In this section, we present the experimental setup to simulate the flight schedules on 17th and
18th July 2019 as two of the most busiest days to date at Amsterdam Airport Schiphol. To this end,
we outline additional assumptions for our study and give an overview of the traffic situation
on these two days. Furthermore, we describe the analyses and key performance indicators to
evaluate the study.

A. Study assumptions
We draw the following additional assumptions with respect to the concept of the control
architecture outlined above:
• the final flight schedule of the two operational days is used and remains static throughout
the simulation
• arriving aircraft are spawned at their historic Actual Landing Time (ALDT); their velocity
at rapid-exit taxiways vRET = vmax , or at regular exits vexit = vturn
• departing aircraft are spawned at their historic Actual Off-Block Time (AOBT), but are
allowed to hold at the stand; they use the standard pushback path of the stand according
to the airport manuals (see (Schiphol - Standaard Pushback per Positie, n.d.))
• the routing algorithm can freely determine the takeoff sequence, but has to comply with the
CTOT-slots of the historic data
• all vehicles execute the instructions from the Guidance Agents perfectly, i.e. no deviations
to planned routes
• the simulation is executed sequentially, i.e. paused when routes are planned
• the standard taxiway directions at Schiphol are ignored

These assumptions decrease the complexity within the simulation, and give the routing algorithm
more freedom to optimize, limiting the risk of an incomplete solution, i.e. situations in which the
routing cannot be done.

B. Traffic situation
At Schiphol, two main runway mode of operations (RMOs) exist: RMO North (active on
17th July 2019), and RMO South (active on 18th July 2019). During each day, different runway
combinations are set. These RMO phases are visualized in Fig. 7.
10

RMO North
off-peak arrival-peak transition departure-peak

06│36L 06+36R│36L 06+36R│36L+36C 06│36L+36C

RMO South
off-peak arrival-peak transition departure-peak

18R│24 18R+18C│24 18R+18C│24+18L 18R│24+18L

Figure 7. Runway mode of operations (RMO) at Schiphol: different phases of RMO North (top) and RMO South
(bottom) with active runways for arrivals (orange) and departures (blue). Background map adapted from (NielsB,
2007).

In the simulation, the same runways are activated according to those that were active in the
historic operations. The runway 09/27 was not active during the two days. Aircraft landing on or
departing from runway 04/22 are not modelled as these general aviation flights remain foremost
within Schiphol East. Table IIa lists the total number of flights, arrivals, departures, the main
RMO, and the number of RMO phases for the two simulated days. As mentioned above, each
aircraft is categorized as one of the six ICAO-types with an associated shape radius and wake
turbulence category (WTC). Table IIb lists these parameters along with the count over the two
days.

TABLE II. O VERVIEW OF (A) TRAFFIC DATA AND (B) PARAMETERS AND DAILY COUNTS OF ICAO- TYPES
( B ) ICAO- TYPES . WTC = WAKE TURBULENCE CATEGORY
( A ) TRAFFIC DATA
parameters count per day
date 17-07-2019 18-07-2019 shape [m] WTC 17-07-2019 18-07-2019
flights 1489 1492 ICAO-A 12 CAT-F 0 0
arrivals 745 744 ICAO-B 25 CAT-E 22 20
departures 744 748 ICAO-C 40 CAT-D 1195 1198
RMO RMO North RMO South ICAO-D 54 CAT-C 37 43
RMO phases 19 19 ICAO-E 72 CAT-B 213 206
ICAO-F 80 CAT-A 22 25

C. Overview of analysis
We carry out two types of analyses to explore the operational consequences of fully automated
airport surface movement operations (ASM Ops), as outlined in the following. While we refer
to key performance indicators, they are further outlined in Section III-D. All simulations use the
historic flight schedules and RMOs defined above.
 11

Analysis of algorithmic parameters to investigate the influence of wplng and hplng on the
trade-off between:
1) efficiency of operations: How do the parameters influence the distribution of taxi times?
2) predictability of taxi times: How predictable is the remaining taxi time when the planning
window changes?
3) computational efficiency of routing algorithm: How long does it take to coordinate the
agents per planning round with respect to different sets of wplng and hplng ?
To this end, we simulate the operations with different values of wplng and hplng within the ranges
specified in Table I. We analyse the first part with the taxi time distributions, the second with the
deviations between predicted and actual taxi times, and the third with the computational time
in relation to the number of agents per planning round. As outcome of the trade-off, we choose
a set of parameters to conduct the following analysis.
Analysis of operational consequences of autonomous airport surface movement operations
with respect to:
1) delay hotspots in the simulated operations: In the routing algorithm, three mechanisms
exist to let an agent avoid the reservations of other agents: an alternative path is chosen,
its speed is reduced down to the minimal velocity of 1.5 m/s, or the agent has to hold at
the stand or engine-start location. The latter two create additional taxi time that we sum
per 5 m-segment of the graph underlying the airport layout. This yields the locations where
most delays across all flights occur. Per simulated day of operations, we visually examine
these congestion areas in the airport layout.
2) efficiency gain on each of the two days: We compare the distribution of taxi times between
historic and simulated operations, and summarize the efficiency gain as change in the key
performance indicators related to the taxi time.
3) predictability of simulated operations compared to the historic operations: We use the
predictability metrics outlined in Section III-D.
4) inbound and outbound holding: We assess how many arriving and departing aircraft must
hold, and the effect on the taxi times. To this end, we compare the taxi time distributions
between historic and simulated operations of all flights that must hold during taxiing.
5) compliance with CTOT-slots: We analyse the compliance with CTOT-slots and the distribu-
tion of takeoff times within the CTOT window between historic and simulated operations.
6) runway sequence and usable capacity: We investigate how the simulated operations affect
the runway sequence and usable capacity compared to the historic operations. We show
the takeoff order and count the remaining slots per departure runway.

D. Key performance indicators

The taxi time distributions form the basis of our analysis as they are directly linked to the
efficiency, predictability, and resulting emissions of airport surface movement operations. Since
the taxi times are not necessarily normally distributed, we report the averages, medians, and
interquartile ranges. Furthermore, we exclude flights with a hold-type assigned by the Routing
Agent due to the potential impact on their taxi times.
Besides, we assess the predictability of the taxi times by defining metrics similar to those used
by Liu et al. (2014):
• The change in taxi time variability as actual vs. predicted value with respect to the remaining
time till the end of taxiing of the simulated operations: recall that per planning round, the
MAS plans the routes for all aircraft that are scheduled to start within the planning window
and those that are already taxiing. Conflicts with other agents that occur beyond wplng are
ignored. This yields a predicted taxi time that is updated in the subsequent planning round,
which takes place latest after the replanning period has passed.
• The predictability of takeoff times as difference between the Estimated taXi-Out Time (EXOT)
and Actual taXi-Out Time (AXOT) of both historic and simulated operations: to this end, we
calculate the root-mean-square prediction error RM SE and the inequality coefficient U as
outlined by (Liu et al., 2014) and defined mathematically below.
• Likewise, the predictability of in-block times as difference between the Estimated taXi-In
Time (EXIT) and the Actual taXi-In Time (AXIT) of both historic and simulated operations.
 12

The prediction error RM SE is defined as

v
u n
u1 X
RM SE = t (AXTi − EXTi )2 (2)
n i=1

with the corresponding actual taxi time (AXT) and estimated taxi time (EXT) per flight i as well
as the total number of flights n. The inequality coefficient U quantifies the relative prediction
error and is calculated as pPn
i=1 (AXTi − EXTi )
2
U= p Pn . (3)
2
i=1 (AXTi )

When this ratio is zero, all predictions are perfect, while unity denotes that the RMSE equals the
error of always predicting a taxi time of zero.
As indication of the runway use, we provide the maximal hourly throughput and occupancy
rate of any departure or arrival runway. The occupancy rate is calculated as the relative time that
the runway is blocked due to the minimal wake turbulence separation between two consecutive
aircraft. Furthermore, we calculate the remaining takeoff slots per day that comply with the wake
turbulence separation of the flights before and after. We compare the historic to the simulated
slot count to estimate the surplus runway capacity.

IV. E XPERIMENTAL R ESULTS

To give an overview of the traffic situation, Fig. 8 shows the hourly count of all flights for
both the historic (black dotted line) as well as simulated operations (grey line) over the two
days of 17th and 18th July 2019. The two curves almost match each other, with the simulated
operations showing a slightly lower total count due to the lower taxi times as discussed below.
Furthermore, the chart visualizes the count of arriving vs. departing flights: the alternating trend
between landings and takeoffs that is characteristic for a hub-and-spoke airport such as Schiphol
is clearly visible. This is also reflected in the frequently changing RMO phases over the course
of the two days, illustrated by the colored shades in the figure.

 KLVWRULF
 WRWDO
DUULYLQJ
KRXUO\FRXQW>@

 GHSDUWLQJ




        
WLPHRIGD\
Figure 8. Hourly count of flights over the two days. Shades denote the RMO phase: off-peak (white), arrival-peak
(orange), transition (grey), and departure-peak (blue)

Table III lists the mean, median, and interquartile range (IQR) of the taxi times for inbound
and outbound flights. The underlying distributions exclude flights that hold explicitly during
taxiing, i.e. arriving flights without a free gate and departing flights with an issued CTOT-slot
exceeding the required taxi time. The distributions are further discussed in Section IV-B, while
the excluded flights are analysed separately in Section IV-F. As predictability metrics, we list the
prediction error RM SE and inequality coefficient U , and refer back to these in Section IV-C.
Moreover, for any runway in use, we report the maximal throughput and occupancy rate per
hour. For arrivals, the indicators are identical between the historic and simulated operations as
we used the actual landing time (ALDT) as spawn-time in the simulation (cp. the assumptions
in Section III). For departures, the maximal hourly throughput is similar between historic and
 13

simulated operations, while the maximal hourly occupancy rate has increased for the simulated
operations. Section IV-H provides more details on the runway sequence and capacity. The listed
CTOT-slot violations are further discussed in Section IV-G.

TABLE III. C OMPARISON OF HISTORIC AND SIMULATED OPERATIONS WITH wplng = 20 min, hplng = 10 min

date 17-07-2019 18-07-2019

operations historic simulated historic simulated
mean taxi time 04:38 03:03 10:15 07:40
median taxi time 04:00 02:53 10:44 08:37
IQR taxi time 02:51 01:40 05:59 05:13
ARR

RMSE taxi time prediction 01:47 00:02 02:16 00:02

U taxi time prediction 33.9% 1.1% 20.6% 0.5%
RWY throughput* 42 38
RWY occupancy* 68.7% 69.3%
mean taxi time 14:29 11:18 10:20 07:05
median taxi time 14:32 11:20 10:06 06:34
IQR taxi time 06:32 04:56 04:20 02:29
RMSE taxi time prediction 02:45 00:40 02:44 00:35
DEP

U taxi time prediction 18.1% 5.5% 25.2% 7.9%

RWY throughput* 45 45 43 42
RWY occupancy* 74.7% 76.1% 73.5% 77.0%
CTOT-slot violations 5 3 0** 0
*: maximal hourly value for any runway
**: corrected after checking historic A-CDM milestones

Figure 9. Aircraft positions at historic A-CDM milestones for inbound (left) and outbound flights (right)
 14

A. Influencing factors on taxi time comparison

As visualized in Fig. 9, most A-CDM milestones of landings and takeoffs, i.e. Actual Landing
Time (ALDT) and Actual Take-Off Time (ATOT), are located on the runway and are thus viewed
as sufficiently precise. Likewise, most block-times, i.e. Actual In-Block Time (AIBT) of arriving
and Actual Off-Block Time (AOBT) of departing flights, are fairly accurate. However, the tracks
of some flights end (1.6 %) or start (4.9 %) further away from the respective stand. Especially
for departing flights, an AOBT situated outside the bay area may lead to significantly lower
taxi times: the pushback and engine-start procedures require multiple minutes, but are likely not
captured in the resulting historic taxi time.
Moreover, the historic tracks of departing flights often do not follow the standard pushback
paths that the simulated flights must follow. Additionally, the actual engine-start duration cannot
be extracted from the historic tracks. In the simulation, aircraft are only allowed to enter the
runway when they can immediately take off, i.e. use a rolling takeoff. However, we do not
model the takeoff, and instead use the time point of entering the runway as ATOT. While all of
these factors introduce potential offsets between the historic and simulated taxi times, we believe
that those leading to shorter historic times have a greater impact. Therefore, we did not attempt
to correct the historic A-CDM milestones, but use them as such when comparing the taxi times
in the following.

B. Comparison of taxi time distributions per runway

Fig. 10 displays the distributions of taxi times for each arrival and departure runway for the
historic and different simulated operations as box-and-whisker plots. These represent the median,
first and third quartiles as box, while outliers are marked by whiskers and points. Like Table III,
all distributions exclude flights that hold explicitly during taxiing, which are analysed separately
in Section IV-F.

DUULYDOV GHSDUWXUHV
 RSHUDWLRQV RSHUDWLRQV
 KLVWRULF KLVWRULF
Z PLQK PLQ Z PLQK PLQ
 Z PLQK PLQ Z PLQK PLQ
Z PLQK PLQ Z PLQK PLQ
WD[LWLPH>PLQ@

 Z PLQK PLQ Z PLQK PLQ

Z PLQK PLQ Z PLQK PLQ




 /5 && 5/  /5 && 5/
5:<VWULS 5:<VWULS
Figure 10. Box-and-whisker plot of historic and simulated taxi times for arrivals and departures per runway-strip

In general, the taxi times from the simulated operations are shorter and vary less. Since the
runway 18R/36L is far away from the central part of Schiphol, taxiing to/from this runway takes
more time than to any of the other runways. As the departing aircraft have to start their engines
to taxi after pushback, their taxi time to any runway is in general longer than for aircraft that
land on the same runway. Furthermore, since we used an engine-start time of 6 min for large
aircraft (ICAO-D to ICAO-F) in comparison to 3 min for small aircraft, the taxi times of departing
aircraft vary more than those of arriving aircraft in the simulated operations.
The five simulations carried out for the algorithmic analysis do not show significant differences,
both for arriving and departing flights from all runways. Therefore, we conclude that both the
 15

planning window wplng as well as the replanning period hplng do not impact the efficiency of
operations within the ranges that we tested. We also tested values for wplng below 15 min, but
the routing algorithm did not succeed to find a solution throughout each of the two days.

Z PLQK PLQ Z PLQK PLQ


SUHGLFWHGWD[LWLPH>PLQ@

TXDQWLOHV
WRRVKRUWYVWRRORQJ





Z PLQK PLQ Z PLQK PLQ

SUHGLFWHGWD[LWLPH>PLQ@
WRRVKRUWYVWRRORQJ





         
WLPHWRHQGRIWD[LLQJ>PLQ@ WLPHWRHQGRIWD[LLQJ>PLQ@
Figure 11. Predictability of taxi time

C. Predictability of taxi times

Fig. 11 shows the variability of the taxi time predictions with respect to the remaining actual taxi
time for four sets of wplng and hplng . The red lines mark the 1 % and 99 % quantiles as indication of
the accuracy over the remaining time. For all four simulations, the first predictions underestimate
the actual taxi time: the aircraft start taxiing almost at the end of the planning window and most
conflicts are thus not yet resolved. The deviation to the actual taxi time decreases strongly in
the following planning rounds. When the remaining taxi time is less than wplng , the difference
between predicted and actual taxi time is negligible for more than 50 % of all flights. The accuracy
further increases towards the end of taxiing, with almost all predictions being accurate within
the replanning period. As listed in Table III, the RM SE and U values decrease significantly in
comparison to the historic operations. Note that this may change when deviations to the planning
arise during execution, which we did not model in this work.

D. Computational efficiency of routing algorithm

The multi-agent system including the routing algorithm are implemented in Python. We ran
the simulations on a Windows 10 laptop equipped with a 1.80 GHz Intel Core i7-10610U CPU
and 16 GB RAM. In Fig. 12, the runtime of each planning round is plotted over the number
of agents that had to be routed for different sets of wplng and hplng . The planning rounds on
day 1 with RMO North are marked by circles, and those of day 2 with RMO South are marked
by crosses. The exponential nature of the runtime in relation to the number of agents is clearly
visible, while the RMO has a subordinate effect. From the parameter settings we tested, we deem
that wplng = 20 min with hplng = 10 min yield a good trade-off between a fast algorithmic runtime
and accurately predicting the remaining time within the last 10 min of taxiing. Thus, we chose
these parameters to analyse the operational consequences of the autonomous operations.
 16

 Z PLQK PLQ

 Z PLQK PLQ
Z PLQK PLQ
UXQWLPH>V@

 Z PLQK PLQ

Z PLQK PLQ
 5021RUWK
5026RXWK


      
DJHQWFRXQW
Figure 12. Runtimes of planning rounds over the number of agents

E. Delay hotspots in the taxiway network

Fig. 13 visualizes congested areas in the taxiway network on each of the two days. On both
days, hotspots form in front of the runway stopbars, the bay areas and stands, as well as on some
taxiway segments. On 17th July, one of the latter hotspots is located in front of the crossing of
the second departure-runway 36C, which is active only in departure-peaks and transition-phases
(cp. Fig. 7). Aircraft slow down in front of the crossing to await its opening, which the Routing
Agent determined to be faster than letting the aircraft go around the Southern end of 18C/36C.
In front of the runways, aircraft queue with reduced velocity until the minimal separation time
due to wake turbulence of the preceding aircraft has passed. In the bay areas, aircraft mostly
hold at their stand, with some more holding after engine-start occurring on the second day.

17-07-2019 (day 1) 18-07-2019 (day 2)

Figure 13. Additional taxi time per layout location on each of the two days, simulated with wplng = 20 min,
hplng = 10 min

F. Holding of inbound and outbound flights

In the taxi time analysis in Section IV-B, we excluded flights with a hold-type assigned by
the Routing Agent. Fig. 14 compares the taxi times of these flights between historic operations
and the different hold-types of the simulated operations as box-and-whisker plot. In general,
not many flights are holding. In comparison to Fig. 10, the taxi times of flights with inbound
holding are significantly higher, and are similar between historic and simulated operations, also
considering that some of the historic A-CDM milestones end at the holding locations. For most
outbound aircraft that must be delayed due to their CTOT-slots, the Routing Agent lets them
hold at their stand. The taxi times of the historic operations are slightly longer, which has an
influence on the moment that the aircraft take off within the CTOT-slot, as we analyse further
below.
 17

DUULYDOV GHSDUWXUHV
DLUFUDIWFRXQW DLUFUDIWFRXQW







































 KROGW\SH
 KLVWRULF
UHPRWH
 VWDQG
WD[LWLPH>PLQ@



 KROGW\SH
KLVWRULF
 GHWRXU
UHPRWH

 /5 && 5/  /5 && 5/
5:<VWULS 5:<VWULS
Figure 14. Box-and-whisker plot of historic and simulated taxi times dependent on the hold-type of arrivals and
departures per runway-strip; simulation with wplng = 20 min and hplng = 10 min

G. CTOT-slots of outbound flights

Over the two days, a total number of 442 CTOT-slots are assigned by Eurocontrol. Fig. 15
visualizes the compliance between the takeoff times to the CTOT-slots of both historic and
simulated operations. While the historic times almost follow a normal distribution centered
around the time issued by Eurocontrol, those of the simulated operations are skewed towards the
beginning of the CTOT-window. In its current implementation, the routing algorithm optimizes
for lowest taxi times and does not attempt to let aircraft take off closer to their calculated
takeoff time. From the 442 flights, 8 historic flights (1.8 %) do not comply with their CTOT-
slots. In comparison, only a single simulated flight takes off outside its CTOT-slot. However, we
noticed that for this flight, the historic A-CDM milestone occurs multiple minutes after the actual
pushback, rendering it a faulty outlier.

KLVWRULFPHGLDQ ,45  VLPXODWHGPHGLDQ ,45 


FRXQW>@




       
RIIVHWWR&727>PLQ@ RIIVHWWR&727>PLQ@
Figure 15. Compliance to CTOT-slot (white area) of historic (left) and simulated operations (right)

H. Runway sequence and capacity of outbound flights

As defined in Section III, we do not use a pre-defined takeoff sequence. Thus, the order of
aircraft departing from a runway is different between the historic and simulated operations,
as exemplary shown in Fig. 16. The minimal separation between flights according to RECAT-
EU is illustrated by the red shades. While the historic order mostly adheres to the RECAT-EU
18

separation, in some cases, two aircraft are separated less than the minimum. We confirmed with
the track data that indeed some flights take off around 52 s after the previous one, despite having
a larger separation than indicated in the figure due to inaccuracies of the A-CDM milestones.

Figure 16. Comparison of exemplary takeoff order at runway 36L between historic (top) and simulated operations
(bottom), with actual takeoff time (blue lines) and minimal WTC separation (red shades)

DFFXPXODWHGUHPDLQLQJ,&$2&VORWVSHUGD\


SULPDU\5:<VHFRQGDU\5:<
FRXQW>@

 VLPXODWHG VLPXODWHG

KLVWRULF KLVWRULF


DFFXPXODWHGGLIIHUHQFHVLPXODWHGKLVWRULFSHUGD\
 SULPDU\5:<
VHFRQGDU\5:<

FRXQW>@



        
WLPHRIGD\
Figure 17. Remaining takeoff slots for ICAO-C aircraft per day for both historic and simulated operations. Shades
denote the RMO phase: off-peak (white), arrival-peak (orange), transition (grey), and departure-peak (blue)

As emergent property of the MAS, whenever possible, flights are grouped together with
minimal separation time between each takeoff. Based on this observation, we count the remaining
slots per departure runway for each of the two days, as visualized in Fig. 17. For both historic and
simulated operations, takeoff slots remain even during busy departure peaks marked by the blue
shades. However, throughout the two days, more slots remain for the simulated operations: in
relation to the total number of departing flights, 10.9 % and 12.6 % additional slots are available
on each of the two days, respectively. Thus, the MAS better utilizes the potential runway capacity.
 19

V. F UTURE W ORK
In this study, we considered a futuristic scenario of autonomous operations. However, before
real-world airport surface movement operations are fully automated, a long transition period will
be necessary with many challenges to investigate and resolve. In general, it needs to be explored
which tasks could be automated and how controllers could then interact with the automated
support tools. Moreover, a suitable architecture of the human-automation teamwork that keeps
controllers in the loop is needed.
As continuation of the research presented in this paper, we first plan to analyse how the engine-
off taxiing techniques that were explored in AEON (2021) affect the key performance indicators
of autonomous taxiing operations. Besides this, aircraft were assumed to execute all commands
exactly as instructed. However, many uncertainties arise in real-world operations. Future research
should determine how severe different sources of uncertainty are, and how to include these in
the model.
Furthermore, as part of surface movement operations, various types of ground vehicles may
come in close contact to parked and moving aircraft, foremost in the aprons. Their movements
must be coordinated with each other and the aircraft. Future versions of the model should include
such operations to explore the operational consequences that may result from their automation.

VI. C ONCLUSION
In this paper, we presented a hierarchical multi-agent system (MAS) model for autonomous
taxiing operations at large airports in which different agents coordinate and control all movements
on the airport surface. As centralized agents, the Airport Operations Agent handles the flight
schedule and runway configuration, while the Routing Agent computes conflict-free trajectories
for all Aircraft Agents. Their execution is then instructed and monitored by the Guidance Agents.
We accounted for aircraft shapes and kinematics during path planning on a high-resolution airport
layout. To this end, we calibrated the model using historic ADS-B data. Furthermore, important
airport surface movement elements and processes were explicitly included in the model such as
pushback, engine-start, inbound holding, complying with CTOT-slots, and adhering to a minimal
safety distance during taxiing as well as minimal wake turbulence separation during takeoff. The
routing algorithm was evaluated to be well suited for planning conflict-free trajectories of all
aircraft.
We analysed the proposed model using the real-world schedules of two of the busiest days
at Amsterdam Airport Schiphol, including 19 different runway configurations per day. For the
considered simulation conditions, the autonomous operations controlled by the MAS model
reduced the average taxi time per flight by around 3 min, or 30 %. Furthermore, the simulated
operations decreased the taxi time variability and yielded a high accuracy in predicting the
remaining time till the end of taxiing. Counting the remaining takeoff slots at departure runways
yielded an increase of approximately 11 % in comparison to the historic operations, meaning that
the MAS model better utilizes the available runway capacity. All obtained results demonstrate
the potential of the MAS as control model for autonomous airport surface movement operations
that are more efficient, predictable, and hence produce less emissions.

A CKNOWLEDGMENT
Not included in this draft.

R EFERENCES
AEON - Advanced Engine-Off Taxiing Operations. (2021). AEON. Retrieved September 25, 2023, from
https://www.aeon-project.eu/
AI roadmap 2.0: Human-centric approach to AI in aviation. (2023, May). EASA.
Atkin, J. A. D., Burke, E. K., & Ravizza, S. (2010). The Airport Ground Movement Problem: Past
and Current Research and Future Directions. In 4th International Conference on Research in Air
Transportation.
Automation in air traffic management: Long-term vision and initial research roadmap. (2020). SESAR
JU.
20

Bae, I., Moon, J., Jhung, J., Suk, H., Kim, T., Park, H., Cha, J., Kim, J., Kim, D., & Kim, S. (2022,
November 18). Self-Driving like a Human driver instead of a Robocar: Personalized comfortable
driving experience for autonomous vehicles. arXiv: 2001.03908 [cs, eess]. Retrieved March 8,
2024, from http://arxiv.org/abs/2001.03908
Bakowski, D. L., Hooey, B. L., Foyle, D. C., & Wolter, C. A. (2015). NextGen Surface Trajectory-Based
Operations (STBO): Evaluating Conformance to a Four-dimensional Trajectory (4DT). Procedia
Manufacturing, 3.
Chen, J., Weiszer, M., Stewart, P., & Shabani, M. (2015). Toward a More Realistic, Cost-Effective,
and Greener Ground Movement Through Active Routing—Part I: Optimal Speed Profile
Generation. IEEE Transactions on Intelligent Transportation Systems, 17.
Chua, Z., Cousy, M., Causse, M., & Lancelot, F. (2017). Initial assessment of the impact of modern
taxiing techniques on airport ground control. Proceedings of HCI-Aero 2016.
Cohen, L., Uras, T., Kumar, T. K. S., & Koenig, S. (2019). Optimal and bounded-suboptimal multi-agent
motion planning. Proceedings of SoCS-19.
De Winkel, K. N., Irmak, T., Happee, R., & Shyrokau, B. (2023). Standards for passenger comfort in
automated vehicles: Acceleration and jerk. Applied Ergonomics, 106.
Eurocontrol. (2018). European aviation in 2040 - challenges of growth. Eurocontrol StatisticsForecast
Service.
Eurocontrol. (2021, November 30). Data snapshot #22 on lower summer taxi-out times.
Helbing, D., & Balietti, S. (2015). How to do agent-based simulations in the future: From modeling
social mechanisms to emergent phenomena and interactive systems design.
IATA. (2021, October 4). Net-zero carbon emissions by 2050. Press Release No. 66. Retrieved April 4, 2023,
from https://www.iata.org/en/pressroom/pressroom-archive/2021-releases/2021-10-04-03/
ICAO. (2004). Advanced Surface Guidance and Control Systems (A-SMGCS) Manual.
ICAO. (2016). Aerodrome design and operations. In Annex 14 to the convention on international civil
aviation (7th ed.).
Lane, R., Dubuisson, S., Ellejmi, M., & Cerasi, E. (2020, April 22). EUROCONTROL specification for
A-SMGCS services.
Liu, Y., Hansen, M., Gupta, G., Malik, W., & Jung, Y. (2014). Predictability impacts of airport surface
automation. Transportation Research Part C: Emerging Technologies, 44.
Ma, H., Harabor, D., Stuckey, P. J., Li, J., & Koenig, S. (2019). Searching with consistent prioritization
for multi-agent path finding. Proceedings of AAAI-19.
Morris, R., & Pa, C. S. (2016). Planning, Scheduling and Monitoring for Airport Surface Operations.
In AAAI-16 Workshop on Planning for Hybrid Systems.
NielsB. (2007, August 6). Schiphol (Amsterdam Airport) overview map. Wikimedia Commons. Retrieved
September 28, 2023, from https://commons.wikimedia.org/wiki/File:Schiphol-overview.png
Phillips, M., & Likhachev, M. (2011). SIPP: Safe Interval Path Planning for dynamic environments.
Proceedings of ICRA-11.
Roling, P. C., & Visser, H. G. (2008). Optimal Airport Surface Traffic Planning Using Mixed-Integer
Linear Programming. International Journal of Aerospace Engineering, 2008.
Rooseleer, F., & Treve, V. (2018, February 14). RECAT-EU: European wake turbulence categorisation and
separation minima on approach and departure.
Sargent, R. G. (2011). Verification and validation of simulation models. Proceedings of WSC-10.
Schiphol - standaard pushback per positie. (n.d.). Schiphol. Retrieved August 30, 2022, from https://www.
schiphol.nl/en/operations/page/sleep-en-pushbackbewegingen/
SESAR Joint Undertaking | ASTAIR - Auto-Steer Taxi at Airport. (2023). Retrieved September 25, 2023,
from https://www.sesarju.eu/projects/astair
SESAR Joint Undertaking | MOTO - the Embodied Remote Tower. (2016). Retrieved September 25, 2023,
from https://www.sesarju.eu/projects/moto
SESAR Joint Undertaking | TaCo - Take Control. (2017). Retrieved September 25, 2023, from https :
//www.sesarju.eu/projects/taco
Stergianos, C., Atkin, J., Schittekat, P., Nordlander, T., Gerada, C., & Morvan, H. (2016). The importance
of considering pushback time and arrivals when routing departures on the ground at airports.
Proceedings of ICAOR-16.
Stern, R., & Sturtevant, N. R. (2019). Multi-agent pathfinding: Definitions, variants, and benchmarks.
Proceedings of SoCS-19.
Udluft, H. (2017). Decentralization in Air Transportation [Doctoral dissertation, Department of Control
and Operations, Aerospace Engineering, TU Delft].
 21

von der Burg, M., & Sharpanskykh, A. (2023). Multi-Agent Planning for Autonomous Airport Surface
Movement Operations. SESAR Innovation Days 2023.
Weiszer, M., Burke, E. K., & Chen, J. (2020). Multi-objective routing and scheduling for airport ground
movement. Transportation Research Part C: Emerging Technologies, 119.
Zhang, M., Huang, Liu, & Li. (2019). Multi-Objective Optimization of Aircraft Taxiing on the Airport
Surface with Consideration to Taxiing Conflicts and the Airport Environment. Sustainability,
11.

View publication stats