0% found this document useful (0 votes)

16 views58 pages

SSRN 4916874

Uploaded by

lwh978122452

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

16 views58 pages

SSRN 4916874

Uploaded by

lwh978122452

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 58

Towards Autonomous ASM Ops using Hierarchical Multi-Agent Planning

ed
Towards Autonomous Airport Surface Movement Operations
using Hierarchical Multi-Agent Planning

iew
Malte von der Burg [email protected]
Jorick Kamphof
Joost Soomers
Alexei Sharpanskykh [email protected]
TU Delft, Faculty of Aerospace Engineering
Kluyverweg 1, 2629 HS Delft, The Netherlands

ev
Abstract
Coordinating the movements of aircraft along the surface of busy airports is a complex

r
task involving both humans and machines. To reduce the aircraft emissions, new engine-
off taxiing techniques are available. The most promising one uses tugs to tow aircraft
to and from the runways. As this adds additional vehicles to the traﬀic, the operational

er
complexity for Air Traffic Controllers increases. Especially in congested traffic situations,
this may overwhelm their cognitive capacity. Artificial intelligence systems could provide
decision-support, and take over repetitive tasks of human operators, such as air traffic
controllers. This paper explores how a hierarchical multi-agent system model can facilitate
pe
both safe and efficient airport surface movement operations in a fully-automated setting. To
this end, we combine and extend state-of-the-art multi-agent motion planning algorithms,
yielding the Multi-Agent Motion Planning on Airport Surfaces (AS-MAMP) algorithm: it
builds upon Priority-Based Search (PBS) and its greedy variant GPBS with the new Safe
Interval Motion Planning (SIMP) algorithm as its low-level solver. We include the different
processes during aircraft taxiing such as pushback, tug coupling, and decoupling into path
planning using a novel activity-based search. Motions of aircraft and ground vehicles are
ot

based on finite acceleration rates, and are calculated in continuous space and time. SIMP
detects and resolves conflicts taking the two-dimensional, circular shapes of vehicles into
account. This ensures collision-free routing also for the confined spaces of the taxiway
tn

systems of large airports. By evaluating four cases based on the potential bottlenecks in
taxiway networks, we confirm that our system design yields safe and eﬀicient operations.
Using sensitivity analyses, we show that both the chosen search-mode in PBS as well as
changes to the acceleration of agents have the strongest influence on the taxi time and
taxi distance as key performance indicators. Moreover, given our simulation conditions, we
show that the multi-agent system model can handle runway throughput levels that match
rin

or exceed those of large European airports.

Keywords: multi-agent system; multi-agent motion planning; airport surface movement
operations; automation; engine-off taxiing; air traﬀic control
ep

1. Introduction
The air transport sector faces three connected challenges. First, the demand is predicted to
increase to around 10 billion annual passengers by 2050 (IATA, 2021), more than doubling
the amount of 2019. Second, the sector strives to achieve net-zero emissions by the same
Pr

time (IATA, 2021) to meet the Paris Agreement (UNFCCC, 2015). As third challenge,
the level of safety must be sustained or rather increased. Zooming in on airports as the

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
key points in the air transport network, recent incidents during taxiing and at runways
(SKYbrary, 2024a, 2024b) underscore the importance of safety of the operations.
The safe and eﬀicient planning and execution of all ground movements is referred to
as airport surface movement operations (ASM Ops), and is a challenging task especially

iew
at large airports involving both humans and machines (Morris et al., 2016). Air Traﬀic
Control Oﬀicers (ATCOs) provide clearances and guidance for all vehicles moving along the
taxiway network. Once an arriving flight has touched down and vacated the runway, its
pilot follows the route instructed by ATCOs through the airport’s taxiway system towards
the bay area and its assigned stand. In some cases, the stand is still occupied by another
aircraft, so that the arriving aircraft must wait e.g. at a remote holding platform. After

ev
reaching the stand, the turn-around processes commence, some of which are to unload the
baggage, refuel, clean and let new passengers board the aircraft. Once ready for departure,
the aircraft is pushed back from the stand, the pilots start the engines, and taxi along the
taxiways towards a designated runway. They enter the runway and take off once cleared
to do so by a tower controller. Besides aircraft, ground vehicles such as aircraft towing

r
vehicles that use or cross active taxiways must be integrated into the flow as well. While
the aircraft pilots and ground vehicle drivers must ensure the safety of their vehicle at all

er
times, they rely on the overall coordination by and communication through ATCOs.
For ASM Ops, the three challenges of meeting the predicted demand, reducing emissions,
and improving safety are highly interrelated. To address the demand, either bigger or
more aircraft are needed, and airports must facilitate this growth with additional capacity.
pe
However, it is expected that infrastructural expansions of airports are insuﬀicient to achieve
higher airport capacity (Eurocontrol, 2018). Thus, the congestion on the airport surface
will increase. This, in turn, will amplify the workload of ATCOs (Chua et al., 2017)
which increases the risk of human errors and may induce additional delays and congestion.
Furthermore, the taxi times of aircraft, i.e. the time they spent taxiing over the airport
ot

surface, become harder to predict. In addition to having a direct influence on the respective
flights, this may trigger knock-on effects on the entire network, resulting in even higher
ineﬀiciencies (Eurocontrol, 2021).
Regarding emissions, the taxi times can reach a significant share of the total flight time
tn

especially for short-haul flights, with combined taxi-in and taxi-out times reaching up to
30 min for European flights (Eurocontrol, 2023). Furthermore, an analysis using a fuel-
burn model showed that between 2% (long-haul flight, 9630km) and 17% (short-haul flight,
350km) of total fuel is burnt during taxiing (Rowland, 2022). The corresponding emissions
rin

can be effectively reduced when the aircraft do not have to use their engines for taxiing.
In recent years, various new engine-off techniques have been explored and potentially yield
significant decreases in carbon and noxious emissions (Lukic et al., 2019; Zoutendijk et al.,
2023). On the other hand, solutions such as the TaxiBot concept from IAI (2013), which
can tow aircraft to runways and back, add additional vehicles to the traﬀic, again amplifying
ep

the issues of airport capacity, ATCO workload, and increased probability for human errors.
To address these interrelated challenges, both technological and operational advances are
urgently needed that make the operations safer, more eﬀicient, predictable, and minimize
the associated emissions. The Airport Collaborative Decision Making (A-CDM) concept
Pr

from Eurocontrol (2016) has been implemented at various airports yielding more informed,
predictable, and eﬀicient decision making. Furthermore, multiple airports are working on

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
Towards Autonomous ASM Ops using Hierarchical Multi-Agent Planning

ed
further implementing the services of the Advanced Surface Movement Guidance and Control
System (A-SMGCS), offering improved safety nets and situation awareness for ATCOs
(ICAO, 2004; Lane et al., 2020). One possible direction is to further increase the level
of automation of the operations. Artificial intelligence systems could provide extensive

iew
decision-support for ATCOs, and even take over certain repetitive tasks. To this end,
both NASA and DLR have developed future concepts of operations for air transport in
the US and in Europe, respectively (Hooey et al., 2014; Okuniek et al., 2016). In both
concepts, the surface trajectory-based operations (STBO) use four-dimensional trajectories
(4DTs) along the taxiing route of aircraft. Each concept was evaluated using human-in-
the-loop experiments. In a joint effort, NASA and DLR have also developed a shared

ev
concept of operations and evaluated it using fast-time simulations (Okuniek et al., 2018).
Moreover, previous research assessed the achievable positioning tolerances of pilots and
auto-pilot systems when following the 4DTs, and revealed that trajectory errors below 22 m
are feasible to achieve, with mean errors even below 7 m for (partially) automated aircraft
taxiing (Bernatzky, 2019).

r
However, high levels of automation are deemed to be reachable only in the far-term
(Hooey et al., 2014). Indeed, such human-automation teaming does not yet exist in real-

er
world applications since today’s technology is not sophisticated enough (Rieth & Hagemann,
2022). To gain the trust of ATCOs, the artificial side must be proven to be capable of han-
dling tasks autonomously in a safe and eﬀicient way. This necessitates a control system
model that realistically incorporates the interrelated processes of ASM Ops as well as the
pe
new engine-off taxiing techniques to compute conflict-free yet environmentally friendly tra-
jectories also in congested traﬀic situations. To this end, both the various processes as
well as the heterogeneous fleet of differently sized aircraft and ground vehicles must be
represented in detail. Nonetheless, in previous work, such models used multiple simplifying
assumptions to handle the inherent complexity. For instance, the fast-time simulations to
ot

test the shared concept of NASA and DLR relied on fixed taxiing routes within only one bay
area of the studied airport layout (Okuniek et al., 2018). As part of an exploratory control
system model to test new operational concepts such as engine-off taxiing or the effect of
existing ATC procedures, a powerful and flexible path planning algorithm is missing that
tn

is capable of dealing with the intricacies of ASM Ops in detail.

In this paper, we present a novel concept of a multi-agent system (MAS) model with
both centralized and distributed agents that is tailored to autonomously control the airport
surface movement operations. Both heterogeneous agent properties and behaviour as well
rin

as randomness can be explicitly modelled with an agent-based model (Helbing & Balietti,
2015). Furthermore, multi-agent simulations allow for inherent modularity, flexibility, and
expressiveness that are beneficial when considering a detailed model. We embedded the
newly developed algorithm for multi-agent motion planning on airport surfaces (AS-MAMP)
into the MAS model. To yield fast solutions, we deploy a two-level, sub-optimal solver. As
ep

its high-level search, multi-agent coordination is based on the Priority-Based Search (PBS)
algorithm (Ma, Harabor, et al., 2019) and its greedy variant GPBS (Chan et al., 2023).
Trajectories that satisfy the constraints from the high-level are calculated with the novel
Safe Interval Motion Planning (SIMP) algorithm, which is one of the main contributions
Pr

of this work. It is inspired by the Safe Interval Path Planning (SIPP) algorithm (Phillips
& Likhachev, 2011) and its variants. SIMP further extends the capabilities of SIPP-based

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
low-level planners to deal with all important characteristics of ASM Ops such as the agent
kinodynamics and sizes, safety zones around agents, and the different operations like push-
back, engine-start, tug coupling, and decoupling.
Based on the MAS concept, we developed a simulator in Python that implements the

iew
AS-MAMP algorithm. We already used it in two previous papers (von der Burg & Sharpan-
skykh, 2023, 2024). Their focus was on the application and evaluation of operational con-
sequences of AS-MAMP-based models for realistic ASM ops in existing airports, and not
on the algorithmic basis underlying these models, which is the main focus of this paper.
In the following, we first provide a description of ASM Ops and define a list of re-
quirements that we deem indispensable to model the operations realistically (Section 2).

ev
Based on these, we outline previous research related to modelling ASM Ops and multi-
agent path planning algorithms in Section 3. We then describe the multi-agent system
model for autonomous ASM Ops (Section 4) including a detailed description of the AS-
MAMP algorithm integrated into the MAS (Section 5). We evaluate the path planning in

r
Section 6 by discussing the influence of important parameters on the AS-MAMP algorithm
using four case studies based on potential bottleneck in airport layouts. In Section 7, we
provide an overview of future explorations of the path planning algorithm and the MAS

er
model. Finally, a summary of the main findings and contributions of this work are outlined
in Section 8.
pe
2. Airport Surface Movement Operations (ASM Ops) and Modelling
Requirements
Currently, aircraft use their own engines to taxi on the airport surface. We refer to this
taxiing mode as multi-engine taxiing (MET) in this paper. For outbound aircraft, the
existing procedures can be split into three parts: 1) the pushback and (possibly) push-pull
ot

manoeuvre, 2) pushback-truck decoupling and engine-start, and 3) taxiing to the runway.

Before step 3) can be carried out, the pilot has to start the aircraft engines. The engine-start
procedure usually begins once the pushback commences. Only in the case of a push-pull
manoeuvre, engine-start must be performed entirely after finishing the pull manoeuvre as
tn

it is carried out to avoid engine blast. After landing, arrival flights taxi to the stand, during
which their engines cool down.
In future ASM Ops, engine-off taxiing techniques may be deployed such as the TaxiBot
system (2013), which we refer to as tug-enabled taxiing (TET) in this work. Based on the
rin

TaxiBot specification and interviews with operational experts, outbound TET operations
can be divided into four stages: 1) pushback and direction switch, 2) tug-enabled taxiing to
a decoupling location close to the runway, 3) decoupling and all-clear signal between tug and
aircraft, and 4) taxiing onto the runway. Dependent on the airport regulations, engine-start
can potentially be performed during part 2). Nonetheless, holding at the decoupling location
ep

may be necessary when the duration of pushback, taxiing, and decoupling is shorter than the
time needed to start the engines. For arriving aircraft, TET comprises three relevant stages:
1) exiting the runway and taxiing to a coupling location close to the runway, 2) coupling to
the waiting tug, and 3) tug-enabled taxiing to the ramp. For both flight directions, the tugs
Pr

have to drive to the coupling location, and travel to the next assignment or the tug-base
once decoupled. Airports may provide extra service-roads for the tugs, or require them to

ed
use the taxiway network managed by ground control. Eventually, the tugs are expected to
be equipped with electric engines (Zoutendijk et al., 2023). In turn, this requires recharging
over the day of operations, either at dedicated charging stations or at the stand, affecting
the task assignment and route planning of tugs after decoupling.

iew
To enable autonomous operations, a control system must be capable to coordinate the
movements of the mixed fleet of aircraft deploying either taxiing technique as well as the
return movements of tugs. Based on the current ASM Ops and with the focus on path
planning, we identified the following characteristics of ASM Ops. They pose modelling
requirements when developing a comprehensive model to study new operational concepts
of ASM Ops in detail:

ev
(1) Multiple actors: pilots, ground vehicle drivers, and other ground personnel have to
share limited resources, i.e. runways, taxiways, and gate areas. Furthermore, these
actors have to fulfil the general objectives of safe and eﬀicient operations as well
as their individual tasks. Since these can be conflicting with one another and between

r
actors, control and coordination is necessary.

(2) Layout: most airports have a marked network of taxiways creating a two-dimensional

er
layout with confined space to be traversed. For large airports, such as Amsterdam
Airport Schiphol (AAS), the taxiway system is rather complex and contains not only
straight and curved segments, but also compounded intersections, runway crossings
pe
and go-arounds, as well as long single-lane taxiways. While aircraft and other vehicles
can generally stop and wait at arbitrary locations, airports usually have designated
holding points. In most cases, aircraft cannot turn around on a taxiway segment,
meaning that deadlocks must be anticipated as these are time-consuming to resolve.

(3) Online setting: aircraft appear and disappear continuously at the airport boundaries
ot

(i.e. the runways).

(4) Interrelated operations: at most airports, multiple routes between a runway and
bay area exist, and various processes take place in parallel (Stergianos et al., 2016).
tn

In current operations, this comprises:

a. different routing tasks such as regular taxiing, pushback-manoeuvres, engine-

start, or tug operations;
b. ATC procedures like temporarily blocked taxiway segments due to chosen run-
rin

way mode of operation (RMO) determining the active runways, safety distances
between vehicles, traﬀic rules such as standard taxiway directions, or prioritiza-
tion of certain vehicles;
c. time constraints that must either be surpassed or reached in time (i.e. dead-
ep

lines), e.g. an arriving flight whose assigned stand is still occupied, a departing
flight that must take off within a departure slot due to airspace restrictions (i.e.
CTOT-slots assigned by Eurocontrol), or aircraft in need of a tug or towing
vehicle;
Pr

d. interactions for instance between aircraft in queues, landing aircraft vacating

the runway, or in the bay areas that should be coordinated safely and eﬀiciently.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
(5) Heterogeneous vehicle fleet: besides different aircraft types, the use of different
engine-off taxiing techniques will lead to a heterogeneous fleet of aircraft and addi-
tional vehicles like tugs. Factors include:

iew
a. differing vehicle dimensions between different aircraft types as well as in
relation to tugs, since the geometric shapes may overlap with multiple other
taxiway segments or entire intersections;
b. kinodynamics or non-holonomic properties of the vehicles, accounting for fi-
nite acceleration and deceleration, minimal speed to avoid stop-and-go, speed
restrictions on both straight as well as curved segments, and suﬀicient safety

ev
distances.

(6) Real-time computational eﬀiciency: for application in a real-world setting, the

underlying algorithms need to have low run-times.

r
(7) Uncertainty in operations: in actual operations, changes to planned actions or
anticipated effects arise that have to be handled accordingly, also during the routing

er
of vehicles (Lee, 2014). For that, examples include:

a. rate of environmental change, such as sudden weather changes and other

system dynamics affecting the operations, are within the order of magnitude of
pe
the taxiing times of aircraft;
b. changes in schedules, especially the flight schedule, task assignments of tugs,
or RMO, with often short lead times (e.g. RMO changes may be known less than
30 min in advance) and are potentially influenced by the rate of environmental
change;
ot

c. positioning tolerances, due to sensor noise or technological limits in precisely

executing the assigned vehicle trajectories;
d. changes to vehicle routes, as pilots and other vehicle drivers may inadver-
tn

tently, or in some cases also intentionally, divert from cleared routes or com-
mands from ATCOs;
e. delays, in procedures / actions, communication, and various other forms (Morris
et al., 2016);
rin

f. disruptive events, at least those that are confined, e.g. vehicle breakdowns or
blocked taxiways.

(8) Interaction with human operators: to conduct future research in the direction of
human-automation teaming, the model needs to be flexible enough to eventually deal
ep

with human input and/or interactions throughout the simulated operations.

In summary, the long-term goal of a model of an autonomous control system for ASM
Ops must be to ensure safe and eﬀicient operations that are as predictable and environmen-
Pr

tally friendly as possible, while handling operational uncertainties as well as inputs from
human operators.

ed
3. Related Work

In its current form, ASM Ops are controlled and coordinated within a centralized system by
air traﬀic control oﬀicers (ATCOs) who coordinate the flow of aircraft from a central loca-

iew
tion, i.e. the air traﬀic control tower. Most previous work modelled ASM Ops as centralized
system, and studied related sub-problems. Other research has focused on developing new
concepts of operations of ASM Ops as a centralized, but also as a distributed system. We
will describe these in the next subsection.
Most previous research used a pre-defined priority order to plan the routes of aircraft
sequentially, usually based on the first-come-first-served principle. However, paths can also

ev
be planned without such stringent order with potentially better solutions. To this end, the
domain of multi-agent path planning (MAPP) provides a multitude of algorithms to solve
both simplified and complex problems. Interestingly, many MAPP-related papers motivate
their work with ASM Ops (e.g. Cohen et al., 2019; Ma et al., 2016; Walker, 2022), while
such algorithms are often neglected in operations-focused research (e.g. Jiang et al., 2021;

r
Yin et al., 2024; Zaninotto et al., 2021; X. Zou et al., 2018). To close this gap, we provide
a comprehensive review of MAPP algorithms and their underlying concepts in Section 3.2.

er
Finally, in Section 3.3, we outline how we address the modelling requirements of ASM
Ops defined in Section 2. Our approach uses a multi-agent system model with integrated
multi-agent motion planning algorithm. We conclude this section by listing the scientific
pe
contributions presented in this paper.

3.1 Modelling of Airport Surface Movement Operations

In previous research, specific sub-problems of ASM Ops were studied. The Ground Routing
Problem (GRP) addresses the task of scheduling the aircraft movements on the airport
ot

surface between runways and stands in an eﬀicient and safe manner while respecting the
operational requirements. Atkin et al. (2010) reviewed and compared various mixed-integer
linear programming (MILP) and genetic algorithm (GA) approaches. While these were
used in various models for (multi-)objective optimization to study variations of this routing
tn

and scheduling problem, the authors conclude that more research is needed that takes the
interdependencies of the processes affecting ASM Ops into account. The Runway Schedul-
ing Problem (RSP) addresses the task of maximizing the utilization of the available runway
capacity through scheduling the landing or takeoff sequence at runways. Both Bennell et
rin

al. (2011) and Lieder et al. (2015) reviewed the common solution approaches that used
either mixed-integer programming (MIP), dynamic programming (DP), branch-and-bound
(B&B), or heuristic algorithms. Furthermore, Guépet et al. (2017), provide a review of
studies that fully or partially integrate ground and runway operations at airports, i.e. com-
bining both GRP and RSP. Using a heuristic sequential approach, they show that a better
ep

integration of these two sub-problems reduces the taxi times significantly. Other studies
argued that most ineﬀiciencies of taxiing operations originate from runway congestion, and
that these are effectively reduced through stand holding, i.e. by delaying the time the
aircraft is pushed back from the stand (Balakrishnan & Jung, 2007; Ravizza et al., 2014).
Pr

Guépet et al. (2016) provide a review of deterministic and queuing models that address this
sub-problem of ASM Ops. Furthermore, Ashok et al. (2017) assessed that a pushback con-

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
trol strategy, but also derated takeoffs, i.e. taking off with lower thrust levels, are effective
to reduce the air quality impacts of aircraft operations.
Li, Gong, et al. (2019) argued that the congestion and uncertainties in the air transport
system are the major constraints on the available capacity. They used a fast-time simulation

iew
for anticipating delays through a probabilistic approach in planning and scheduling of ASM
Ops. Yin et al. (2024) reviewed approaches for taxi time prediction and analysis. Different
models are used to evaluate and optimize taxi time predictions under uncertainty, e.g. using
a stochastic model (Lee, 2014), chance-constraint programming (X. Wang et al., 2021),
or a fuzzy rule-based approach (Brownlee et al., 2018). Other studies used data-driven
approaches to predict the landing times of aircraft (Dhief et al., 2020; Z. Wang et al.,

ev
2018), or to visually show the runway exit prediction (Woo et al., 2022).
As another research direction, new concepts of operations (CONOPS) are studied. Mor-
ris et al. (2016) proposed a centralized automation system to handle tower operations. They
used discretized time and a grid-based layout with obstacles to form taxiways of Dallas Fort

r
Worth airport (DFW). Uncertainty during taxiing of aircraft is handled through continuous
scheduling and monitoring. Using a discrete-time Markov chain (DTMC) they were able
to partly transfer uncertainties during plan execution into probabilistic behaviour models.

er
Both NASA and DLR have developed new CONOPS to increase eﬀiciency and predictabil-
ity as well as to reduce the environmental impact of ASM Ops. NASA’s CONOPS for
far-term Surface Trajectory-Based Operations (STBO) envisions that pilots actively partic-
ipate in precisely meeting time-based goals (Hooey et al., 2014). It comprises two phases: in
pe
phase 1, minimal flight deck changes are required for pilots to conform to time-based clear-
ances at specific trajectory points. In phase 2, high precision and conformance is enabled
by providing full 4D trajectories comprising (x,y)-locations at all times t along the taxi
route, which requires advanced flight deck equipment. However, the authors note that this
phase demands further research. They developed decision-support tools for ATCOs to eval-
ot

uate their CONOPS using human-in-the-loop simulations, which showed promising results
(Hayashi et al., 2015; Jung et al., 2011). Furthermore, NASA conducted studies to test how
well pilots are able to conform to 4D taxi clearances when supported with different flight
deck displays providing trajectory information (Bakowski et al., 2015, 2017). DLR imple-
tn

mented both a departure management system (DMAN) and a surface management system
(SMAN) to test the Airport Collaborative Decision Making (A-CDM; Eurocontrol, 2016)
and the Advanced Surface Movement Guidance and Control System (A-SMGCS; ICAO,
2004) concepts with human-in-the-loop simulations (Gerdes & Schaper, 2015). The simula-
rin

tor has a built-in conflict detection and resolution functionality to monitor the conformance
to the 4DT clearances. In a joint effort, NASA and DLR also developed a shared concept of
operations and evaluated it using fast-time simulations (Okuniek et al., 2018). They used
pre-defined routes between each origin-destination pair combined with the first-come-first-
served principle to assign priorities between aircraft. However, both aspects likely lead to
ep

sub-optimal solutions and may not be applicable to complex traﬀic situations.

In general, centralized systems may lack knowledge of local conditions affecting oper-
ations, and communication might be impeded by capacity and bandwidth issues (Udluft,
2017). In comparison to that, ASM Ops can also be modelled with decentralized control,
Pr

as studied by Udluft (2017). In such a system, communication and control mechanisms are
shifted to multiple local entities. While they result in higher robustness, resilience, and bet-

ed
ter solutions for local problems, a global optimum cannot be guaranteed and the available
global resources may not be used eﬀiciently.

3.2 Multi-Agent Path Planning Algorithms for Conflict-free Routing

iew
Much research has concentrated on the classical multi-agent path finding (MAPF) problem
in which several assumptions are utilized to simplify the search: the environment is defined
as static, both time and space are discretized, agents are depicted as points that only
occupy a single location per timestep, and move between these within unit time (Stern,
2019). Moreover, it is typically assumed that the generated plans are perfectly executed

ev
(Ma et al., 2017). Furthermore, in most MAPF-related research, an offline setting (i.e. pre-
known start and goal locations of all agents) is assumed. Comprehensive overviews of the
MAPF problem and its algorithmic landscape are provided in (Stern & Sturtevant, 2019,
with MAPF definitions, variants, benchmarks), (Felner et al., 2017, with MAPF variants

r
focused on search-based optimal solvers), and (Gao et al., 2024, with classic but also beyond
classic MAPF solvers).

er
Two-level solvers have proven to be the most suitable to solve such problems: they
combine the strengths of different algorithms and yield better solution quality and lower
computational times than other approaches such as reduction-based, rule-based, or other
one-level search-based solvers. Prominent examples of optimal two-level MAPF solvers are
pe
M* (Wagner & Choset, 2011), the Increasing Cost Tree Search (ICTS) (Sharon et al., 2013),
and Conflict-Based Search (CBS) (Sharon et al., 2012). Especially the latter one has found
broad use for classical MAPF instances. Furthermore, to reduce the computational time
needed to run this optimal solver, sub-optimal variants were developed that still return
near-optimal solutions in most cases (Barer et al., 2014; Cohen et al., 2016).
ot

CBS and its variants are based on the ideas of building a binary conflict-tree during the
search, and separating conflict detection and conflict resolution into two different search
levels. In the high-level, conflicts are detected sequentially. For every occurring conflict,
two new child nodes, one per conflicting agent, are added to the search tree including a
tn

new constraint that prohibits one of them of occupying the conflicting state. Then, the
low-level search resolves the conflict by finding a new path for the constrained agent, for
which any single-agent path finding (SAPF) algorithm such as the well-known A* algorithm
may be used. This process is repeated until the solution is conflict-free, usually employing
rin

a best-first search strategy. Barer et al. (2014) proposed the bounded-suboptimal variant
Enhanced CBS (ECBS). It uses a focal search on the low-level of CBS to find paths which
satisfy the high-level constraints, but also minimize the unresolved conflicts with the paths
of other agents. By reducing the number of conflicts that the high-level still has to resolve,
this procedure speeds up the search. Li, Ruml, and Koenig (2021) introduced Explicit
ep

Estimation CBS (EECBS), which is currently deemed to be the state-of-the-art bounded-

suboptimal algorithm. EECBS uses an explicit estimation search (J. T. Thayer & Ruml,
2011) instead of the focal search in the high-level of ECBS. Furthermore, they used online
learning (J. Thayer et al., 2011) as inadmissible heuristics to guide the high-level search, and
Pr

include other recent improvements of CBS. While these bounded-suboptimal CBS-variants

showed promising results in trading off solution quality for runtime, they may still be too

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
slow for applications requiring close to real-time path planning computations, especially
when relaxing the assumptions of classical MAPF.
In contrast to CBS and its variants, Ma et al. introduced the Priority-Based Search
(PBS) algorithm. PBS generates a priority-tree to explore the state space by applying

iew
different priority orderings for the agents. During conflict resolution, the lower prioritized
agent has to avoid all paths of those higher in priority. To speed up the algorithm, a
depth-first search is applied, and an optionally provided initial priority order can further
improve its runtime. While PBS is unbounded suboptimal and incomplete, the authors
claim that it returns near-optimal solutions. For that, they compared PBS to CBS, also on
an airport-like map (DAO-map brc202d), yielding similar solution qualities but a greater

ev
performance of PBS regarding success rate and runtime. PBS was used in several recent
studies (Li et al., 2023; Zheng et al., 2023), and Chan et al. (2023) deem PBS to be the
leading suboptimal yet fast MAPF algorithm. They further improved its runtime especially
for MAPF instances with many agents and obstacles by introducing Greedy PBS (GPBS)
that greedily minimizes the number of conflicts between agents to speed up its execution.

r
In the depth-first search within the priority-tree, the child node with the lowest number of
remaining conflicting agent pairs is expanded next. Furthermore, they introduced several

er
GPBS extensions to further improve its execution speed. They adopt the partial expansion
(PE) technique from EPEA* (Goldenberg et al., 2014) to randomly create only one of
the two child nodes (the second child node is only created when backtracking), and target
reasoning (TR; Li et al., 2020) to first resolve conflicts with an agent waiting at its goal.
pe
Furthermore, besides the new priority order added to a child node of N , they propose to
also include the set of induced constraints (IC) between the existing priorities: when agent
ai is given higher priority than agent aj , ICi≺j (N ) mark the implicit priority relations in N
between all agents higher in priority than ai that must have higher priority than any agents
lower in priority than aj . In GPBS with IC, the conflicting pair with the largest |ICi≺j (N )|
ot

is resolved by creating child nodes of N . Moreover, instead of backtracking upon reaching a

dead end node N , they used soft restarts (SR) in which N is set as new root node removing
all priorities but keeping the paths.
As general approach to decrease execution times of algorithms, Silver (2005) introduced
tn

the idea of windowing the search, which is also referred to as bounded-horizon planning in
other articles (Li, Tinka, et al., 2021). Using a windowed approach, conflict resolution is
done only up to a pre-defined time horizon w, while conflicts beyond that limit are ignored.
This decreases the search space and therefore results in a lower runtime of the path finding
rin

algorithm.
In previous research, some of the assumptions of classical MAPF mentioned above were
relaxed. Švancara et al. (2019) explored the online MAPF variant in which new agents may
dynamically appear and disappear. In such an online setting, highest solution quality can be
achieved by re-planning the paths of all existing and newly appearing agents. Despite being
ep

computationally expensive, this may change existing agent plans, which can be undesirable
in certain cases as it requires additional communication and reaction overhead. Walker et al.
(2018) defined the MAPFR problem to solve MAPF instances with continuous, non-uniform
edge weights. Furthermore, they generalized ICTS for non-unit costs called Extended-ICTS
Pr

(E-ICTS) by increasing the target cost in the high-level search with a cost-interval instead
of a unit-cost increment: the low-level search has to find a solution whose cost is within

ed
the updated cost-range. Besides the optimal version, they also introduced two bounded-
suboptimal variants of E-ICTS.
Li, Surynek, et al. (2019) explored MAPF instances with Large Agents (LA-MAPF), in
which agents occupy not one but a set of points in a graph at any timestep. While this

iew
poses the possibility to introduce safety distances between agents, it also necessitates new
conflict types and increases the amount of possible collisions. Especially for conflict-based
approaches like CBS, this may lead to many additionally created nodes in the constraint
tree increasing its runtime. To mitigate this effect, the authors introduced Multi-Constraint
CBS (MC-CBS) in which closely related conflicts are added as multiple constraints to a
newly generated child node in the high-level search of CBS. Walker et al. (2020) generalized

ev
MC-CBS through the use of bipartite reduction (BR), forming CBS with time-annotated
bicliques (CBS+TAB). They showed that CBS+TAB is complete and optimal, and claim
that it can be used on both discretized and continuous environments. Furthermore, they
introduced two sub-optimal yet complete variants called CBS+TAB with conditional con-
straints (CBS+TCC) and CBS+TAB with conflicting paths (CBS+TCP). Both are based

r
on the idea of additionally using conditional constraints in the low-level solver. When the
low-level cannot find a solution, it is invoked again with deactivated conditional constraints

er
to guarantee completeness while otherwise allowing for improved performance.
Ma et al. (2018) formulated the problem of MAPF with Deadlines (MAPF-DL) in which
the number of agents reaching their goal within a deadline shall be maximized. To this end,
they adapted CBS to deal with deadlines (CBS-DL), introduced Death-Based Search (DBS),
pe
and combined both to form meta-agent DBS (MA-DBS). Like CBS, DBS is a complete and
optimal two-level solver. However, rather than imposing agent constraints on nodes or
edges, DBS labels individual agents as unsuccessful. Using a best-first search, it finds a
conflict-free solution with the smallest set of unsuccessful agents.
Ma et al. (2017) generalized the MAPF problem by taking delay probabilities into
ot

account (MAPF-DP) and provide robust plan execution policies. In comparison to this,
Atzmon et al. (2020) directly considered robustness during planning. In their approach, k-
robust MAPF plans are generated that allow up to k unexpected delays to occur per agent
before being invalid. Semiz and Polat (2021) studied Incremental-MAPF (I-MAPF) to re-
tn

plan when environmental changes occur that temporarily make locations unavailable for
agents. MAPF plans can usually not be directly executed by real-world robots (Yakovlev
et al., 2019). To address this issue, Hoenig et al. (2016) proposed MAPF-POST which
uses a temporal network to post-process MAPF solutions so that non-holonomic robots can
rin

execute the plans with guaranteed safety distances between them.

Nevertheless, the strong assumptions to solve such MAPF instances are often not appli-
cable to real-world applications (Ma et al., 2016). Dependent on the application, continuous
time and space dimensions as well as non-holonomic, heterogeneous properties of the agents
may need to be accounted for directly in path planning. The combination of these is often
ep

referred to as multi-agent or multi-robot motion planning (MAMP or MRMP). According

to González et al. (2016), the latter addresses open spaces with less restrictions to obey,
while MAMP focuses on road networks where restrictions such as traﬀic rules apply. In
their review, they classify various techniques for motion planning of vehicle trajectories.
Pr

However, from the literature that they reviewed, only the search-based and sampling-based
techniques have been applied for multi-agent cases. Sampling-based approaches are often

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
based on the RRT* algorithm and are utilized to find paths and generate roadmaps in open
environments. For quadrotor swarms in three-dimensional open spaces, Hönig et al. (2018)
presented an approach that generates a common roadmap between agents. Their MAPF
solver with Generalized Conflicts (MAPF/C) then annotates the roadmap with conflicts

iew
and yields coordinated, discretized plans for all agents that are then iteratively refined to
yield arbitrarily smooth trajectories. Kottinger et al. (2022) introduced Kinodynamic CBS
(K-CBS) which is a decentralized MRMP algorithm. In the high-level search, K-CBS de-
tects collision time intervals in an existing motion graph generated by any sampling-based
solver. To resolve conflicts, this graph is subsequently refined by sampling an alternative
motion and checking its set of discretized points for validity. They claim that their approach

ev
is general, scalable, and yields probabilistically complete solutions.
However, the taxiway layouts of airports do not resemble open spaces, but rather consist
of sparsely located intersections connected by roads in a continuous, two-dimensional Eu-
clidean environment. Sampling-based techniques are thus not applicable to coordinate the
agents in ASM Ops. Instead, the taxiway network can be transformed into a graph using

r
the physical coordinates of intersections as nodes and the length of the interconnecting taxi-
ways as edge weights, yielding a representation in continuous space. Note that alternative

er
transformations are also feasible as done by Morris et al. (2016), which used a grid-based
layout with obstacles to form the taxiway network of Dallas Fort Worth airport (DFW).
For both, the above mentioned two-level search-based solvers can be applied when cou-
pled with a low-level algorithm that can deal with the continuous dimensions and non-
pe
holonomic agent properties. Phillips and Likhachev (2011) introduced the Safe Interval
Path Planning (SIPP) algorithm, which is in first instance a SAPF solver. SIPP is based on
the idea of summarizing timesteps into time intervals: the duration that a moving obstacle
occupies a graph location is defined as unsafe interval (USI), while the time between two
USIs as safe interval (SI). This representation of the time dimension reduces the amount of
ot

potential states at every graph location and therefore the search space. Besides decreasing
the runtime significantly, this also enables planning with continuous time. In contrast to
A*, SIPP generates successive states based on the set of possible motions from the current
one. The earliest arrival time at the new location is then compared against the safe intervals
tn

of that location. A successor is only generated when the motion is feasible, i.e. fulfils the
underlying kinematic conditions. This makes SIPP well suited to cope with continuous time
and space, arbitrary motion durations, and different agent speeds.
For this reason, SIPP was subsequently applied as low-level solver to solve multi-agent
rin

instances. Andreychuk et al. (2019) introduced and later improved (2021) CCBS as optimal
solver based on CBS and SIPP. Their algorithm is capable of dealing with non-uniform
action durations, continuous time, as well as accounting for the geometric shape of agents
and obstacles. Although the environment is discretized into cells, i.e., expressed as a 2k -
connected grid, they claim that CCBS can work on any graph. Recently, Walker et al. (2024)
ep

further improved the performance of CCBS through symmetry-breaking enhancements,

resulting in CCBS with disjoint k-partite cliques (CCBS+DK).
For any-angle path planning generating shorter and smoother paths in grid environ-
ments, Any-Angle SIPP (AA-SIPP) was proposed as greedy algorithm (Yakovlev & Andrey-
Pr

chuk, 2017), and later improved to yield optimal solutions called Time-Optimal Any-Angle
SIPP (TO-AA-SIPP; Yakovlev & Andreychuk, 2021). However, the authors note that TO-

ed
AA-SIPP might be too slow to be applicable to multi-agent planning. To address this issue,
Y. Zou and Borst (2024) recently introduced Zeta*-SIPP that reduces its runtime by up
to 90 % on average. However, these algorithms do not yet account for both kinodynamic
constraints and the size of agents. To this end, Yakovlev et al. (2019) proposed modifica-

iew
tions to AA-SIPP in combination with a prioritized multi-agent planner to generalize the
algorithm to handle different sizes and speeds of agents as well as movements in continuous
time.
Ma, Hönig, et al. (2019) used SIPP with a reservation table (SIPPwRT) to generalize
SIPP in continuous time and large agents moving on 4-neighbour grids. The shape of agents
is expressed by a safety radius and can be increased to add a safety margin around the agent.

ev
Together with the velocities of two agents, but without accounting for acceleration and
deceleration rates, they then compute time offsets to set safe intervals for lower prioritized
agents. The authors claim that SIPPwRT is therefore guaranteed to be collision-free, and
tested their algorithm for multi-agent pickup and delivery tasks in warehouses.

r
Li et al. (2022) generalized SIPP to handle soft obstacles, called SIPP with Soft Con-
straints (SIPPS). Each state S in SIPPS also contains a lower bound estimate of the number
of soft collisions along the path to S. Soft collisions mark the pending conflicts with the

er
plans of other agents that are not yet set as hard constraints. SIPPS thus minimizes the
number of unresolved conflicts when creating a new plan: in the open-list, the state with
the lowest number of soft collisions is expanded next, breaking ties in favour of states with
pe
lowest f-score. The authors used SIPPS in their proposed algorithm MAPF-LNS2 that
carries out a large neighbourhood search (LNS) to gradually minimize the total number of
collisions by repeatedly replanning for a subset of agents. Within a defined runtime limit,
MAPF-LNS2 is thus able to either return a conflict-free solution or one with only a few
unresolved conflicts. Moreover, SIPPS could also be adapted to pick the state with the
lowest f-score and breaking ties based on the lowest number of soft collisions so that it
ot

preserves the properties of e.g. CBS and its variants like EECBS when used as its low-level
solver. In GPBS, Chan et al. (2023) deployed both alternatives of SIPPS.
Ali and Yakovlev (2023) argued that all previous work on SIPP does not account for
tn

the kinodynamic constraints of agents prohibiting them to stop instantaneously. They show
that these agent properties make vanilla SIPP and its variants incomplete. To address this
issue, they proposed SIPP with (Wait) Interval Projection (SIPP-IP) and prove that it is
both complete and optimal. In SIPP-IP, each state S has an associated non-overlapping,
projected safe interval representing the time interval that can be safely reached from the safe
rin

interval at a previous wait-location in the path to S. We extend their idea to be applicable

also to states that did not originate from a wait-location, as summarized below and further
outlined in Section 5.7.
ep

3.3 Our Modelling Approach and Scientific Contributions

Taking the advantages and limitations of both centralized and distributed systems into con-
sideration, we created a novel architecture of a multi-agent system (MAS) model with both
centralized and distributed agents to coordinate and control ASM Ops. Both heterogeneous
Pr

agent properties and behaviour as well as randomness can be integrated into an agent-based
model (Helbing & Balietti, 2015). Furthermore, multi-agent simulations allow for inherent

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
modularity, flexibility, and expressiveness that are beneficial when considering a detailed
model. These allow us, in combination with the hierarchical structure of the MAS model, to
address Requirements 1, 7, and 8 from Section 2. Further details are provided in Section 4
that outlines the architecture of the MAS model.

iew
Based on the path planning concepts reviewed in Section 3.2, we embedded a tailored
algorithm for multi-agent motion planning on airport surfaces (AS-MAMP) into the multi-
agent system to satisfy the remaining requirements from Section 2 for modelling ASM Ops
in detail. As outlined in Section 5, we combine PBS with our newly developed Safe-Interval
Motion Planning (SIMP) algorithm. The latter is an augmented version of SIPP that
takes the heterogeneous, non-holonomic agent properties as well as their dimensions into

ev
account (satisfying Requirement 5). It uses motions in continuous time and space along
a graph that represents the taxiway network (Requirement 2). Furthermore, SIMP can
cope with the different interrelated processes (Requirement 4) and has a low computational
time (Requirement 6) given the complexity of the search with kinodynamics and shapes of
agents. To further cut down on runtime and to satisfy Requirement 3, we use a windowed

r
search. This also allows us to deal with uncertainty in the operations (Requirement 7): the
routes can be re-planned if, for example, the schedules change or taxiways are blocked or

er
become unavailable. Moreover, we can guarantee conflict-free plans for all concurrent routes
(Requirement 1) since we explicitly detect and resolve all conflicts between the shapes of two
agents that account for both safety distances and positioning tolerances. We can verify this
property directly through the simulator, in which we assume instantaneous communication
pe
and perfectly executed commands at this point. Time constraints of agents (Requirement
4c) are handled by the Routing Agent as part of the MAS: it assesses whether e.g. holding is
necessary and adjusts the activity sequence representing the taxiing processes of the agents
accordingly.
In summary, the main contributions of this paper are:
ot

• the architecture of the hybrid multi-agent system model to autonomously control

airport surface movement operations;

• the AS-MAMP algorithm for conflict-free multi-agent motion planning that accounts
tn

for the heterogeneous fleet of agents, their shapes and kinodynamic constraints;

• the SIMP algorithm as low-level solver in PBS that extends the ideas of Ali and
Yakovlev (2023) to deal with kinodynamic agent properties, and that calculates the
rin

motions between states on-the-fly;

• the underlying motion principles based on constant acceleration as well as the treat-
ment of space-time reservations.

Although we focused on creating a multi-agent motion planning algorithm that is tai-

lored to ASM Ops, we believe that the presented ideas could be useful for other applications
requiring a detailed representation of agent kinodynamics and shapes. Due to the inherent
complexity of embedding the operational details of such real-world systems in the algo-
rithms, we did not yet implement high-level alternatives like EECBS or MAPF-LNS2, or
Pr

techniques to further improve SIMP such as using the soft constraints from SIPPS. We
leave this for future work, as further outlined in Section 7.

ed
4. Multi-Agent System Model for Autonomous Airport Surface
Movement Operations
In this section, the conceptual architecture of our hybrid multi-agent system (MAS) for
autonomous airport surface movement operations (ASM Ops) is described. Fig. 1 shows

iew
its control structure, comprising both centralized and distributed agents. The centralized
Airport Ops Agent schedules the flights and tasks of ground vehicles, and determines con-
straints for the taxiway network (i.e., the Airport Layout) originating from the runways in
use or temporarily unavailable taxiway segments. Both schedules and layout constraints
are shared with the centralized Routing Agent. Its task is to compute conflict-free routes
for all vehicles that are moving or will move over the airport surface within the predefined

ev
planning window wplng , and to re-compute these at least once every re-planning period
hplng . The generated routing plans are communicated to the Localized Agents that are
positioned at every junction in the taxiway network. Each Localized Agent controls those
Moving Agents that are travelling towards its location, sends them instructions according

r
to the routing plans, and monitors the execution of those instructions. To do so, they use
the Surface Movement Radar (SMR) to obtain the position and speed of a Moving Agent.
All necessary information such as the past and planned route, activity sequence, and cur-

er
rent SMR data is stored in the Datastrip (DS) associated to each Moving Agent. Once a
Moving Agent has passed the position of a Localized Agent, it hands over the control to the
next Localized Agent by transferring the respective DS. The Moving Agents represent the
pe
aircraft pilots and tug drivers, and are modelled to be fully cooperative. We thus assume
that they execute the received instructions to the best of their possibilities when moving
over the airport surface.

MULTI - AGENT SYSTEM

E CENTRAL AGENTS
N Global Clock
V Airport Ops Agent
I
tn

Airport Layout
R
Routing Agent
O
N Surface Movement Radar
M
LOCALIZED AGENTS
E
rin

N Datastrips
T MOVING AGENTS

Figure 1: Architecture of hybrid multi-agent system (MAS) for autonomous airport surface
ep

movement operations (ASM Ops)

4.1 Specification of Environment

The Environment comprises a Global Clock to synchronize the time between all agents, the
Airport Layout representing the taxiway network, the Surface Movement Radar to track

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Moving Agents on the airport surface, and the Datastrips associated to every Moving Agent.
The latter is used to pass control between Localized Agents and to store the necessary
operational as well as analysis data.

iew
The taxiway network is represented by a graph G = (V, E) with nodes V and directional
edges E. An exemplary Airport Layout is shown in Fig. 2. Tug decoupling locations (orange
nodes) are specific to each runway (grey edges) and are used for tugs to couple to or decouple
from aircraft that taxi with tug-enabled taxiing. Each bidirectional road segment between
two nodes is constructed from two unidirectional edges that connect the nodes. Taxiway
edges (black) as well as service road edges (blue) can be obtained from the actual locations

ev
of these taxiways at an airport, such as Amsterdam Airport Schiphol. Ramp nodes (green)
represent the aircraft parking positions. The airport layout graph is fully accessible to
the Routing Agent and remains static throughout the simulation. However, edges can be
constrained to prohibit vehicles to be routed over these. Such layout constraints are set by
the Airport Ops Agent and described below.

r
er
pe
ot
tn
rin

Figure 2: Exemplary layout of taxiway network. Taxiway centerlines as well as pushback-

paths are shown as black, runways as grey, and service roads as blue edges. Small nodes
represent taxiway intersections (black) or service road intersections (blue). Large nodes
indicate aircraft ramps (green), stopbars (red), decoupling locations (orange), or their as-
sociated all-clear points (grey).
Pr

ed
4.2 Specification of Airport Ops Agent
The Airport Ops Agent manages all information that influences the airport surface move-
ment operations as a whole: it schedules the flights and tasks of ground vehicles, handles
necessary changes of these, and determines the active runways, i.e., the runway mode of

iew
operation (RMO) for the next 15 minutes. Furthermore, it notifies the Routing Agent of
any temporarily unavailable taxiway segments. The Airport Ops Agent has the following
properties:
P1. Receive Request Property: both Routing Agent and Localized Agents can send
information that affect the operations to the Airport Ops Agent, which then acts

ev
accordingly. For instance, a Localized Agent could report a blocked taxiway segment,
or the Routing Agent may request a change in the schedules to allow for conflict-free
routing.
P2. Update Schedule Entry Property: the Airport Ops Agent schedules all flights

r
and tasks of tugs and updates the entry in the respective schedule accordingly, in-
cluding any necessary changes that arise during the operations (cp. P1). The updated
schedules as well as the time point of the earliest change are shared with the Routing
Agent.
er
P3. Update RMO Property: at each time point t, the Airport Operations Agent checks
whether the active runways have to change. If so, it adjusts the blockage intervals
pe
from t + 15 min onwards for all edges affected by the RMO change. It then shares the
new set of layout constraints with the Routing Agent.
P4. Request Replanning Property: if the time point of earliest change in either sched-
ules or RMO falls within the planning window wplng , the Airport Ops Agent sends a
request for immediate re-planning of the routes to the Routing Agent.
ot

4.3 Specification of Routing Agent

The role of the Routing Agent is to perform the central route planning. To do so, it
tn

gathers the updated schedule data, layout constraints, along with local information from
the Localized Agents, and converts this into route parameters per agent. Then, it computes
a plan for all vehicles within the planning window wplng using the AS-MAMP algorithm
specified in Section 5, and shares the updated routing plans with the Localized Agents. The
rin

properties of the Routing Agent are as follows:

P1. Receive Request Property: the Routing Agent can receive a re-planning request
from both Airport Ops Agent and Localized Agents. The former triggers re-planning if
changes to the schedules or the layout constraints become effective within the planning
ep

window, while the Localized Agents do so when the deviation between planned and
executed route of a Moving Agent exceeds time thresholds. The request is stored
internally and further handled in P2.
P2. Check Replanning Property: at each time point t the Routing Agent checks
Pr

whether a re-planning request was made or the internal re-planning timer elapsed. If
so, the timer is reset, and the routes are re-planned by executing P3, P6, and P7.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
P3. Update Constraints for Route Planning Property: the layout constraints from
the Airport Ops Agent are gathered and stored internally. These change foremost
when the RMO switches, and allow the Routing Agent to anticipate on upcoming
taxiway closures or re-openings corresponding to the altered RMO.

iew
P4. Handle Time Constraints of Moving Agent Property: to address potential
time constraints of a Moving Agent (Requirement 4c), the Routing Agent assesses
whether these are achievable and whether remote and/or stand holding is necessary,
which influences the activity sequence of the Moving Agent that it defines through
P5. Furthermore, it sets operational constraints pertaining to the time constraints,

ev
to account for these during path planning done by its P7.

P5. Define Activity Sequence of Moving Agent Property: to include the various
parts of the route of a Moving Agent such as push-pull manoeuvres, coupling / de-
coupling of tugs, or holding in path planning, the Routing Agent converts these into

r
one of the following three activities:

er
• go-to activities comprise a start location and a set of goal locations, so that
in path planning, two degrees of freedom exist: time and route. This activity
is used as input to taxi from one point at the airport to another point at the
airport, for example for regular taxiing.
pe
• follow activities consist of a predefined list of path segments that must se-
quentially be part of the route. Therefore, time is the only remaining degree of
freedom in planning, and the path cannot be changed. This activity is used for
instance for pushback and push-pull manoeuvres.
• wait activities prescribe a waiting location and waiting duration that have to
ot

be accounted for in path planning. Operations such as coupling to or decoupling

from a tug are examples of such.
tn

Using a combination of these, the Routing Agent defines an activity sequence accord-
ing to the schedule of a Moving Agent, which has to be executed in the respective
order.

P6. Update Route Parameters of Moving Agents Property: per Moving Agent
rin

that is scheduled to taxi within the planning window, the Routing Agent collects all
necessary parameters to plan its route. This encompasses the kinodynamic represen-
tation according to the vehicle type of the Moving Agent, the activity sequence derived
from its schedule (shared by the Airport Ops Agent), and in case the Moving Agent
is already taxiing also the latest execution data (shared by the controlling Localized
ep

Agent).

P7. Do Route Planning Property: the Routing Agent invokes the AS-MAMP algo-
rithm (cp. Section 5) to re-plan the routes of all agents that are part of the current
Pr

planning round. The resulting new conflict-free routes are shared with the Localized
Agents.

ed
4.4 Specification of Localized Agents
Each Localized Agent controls the Moving Agents that are travelling towards its location
and hands over the control to the next Localized Agent once a Moving Agent has passed its
location. Based on the generated routing plans, it gives instructions to the Moving Agents,

iew
and monitors their execution. It acquires the necessary position and speed data from the
radar. When the executed route deviates from the plan, a Localized Agent can adjust the
route locally to minimize the deviations. If the deviation to the planned route exceeds
time thresholds, it can request re-planning of the routes. The Localized Agents thus act
as link between the centralized planning and the local operations. They have the following
properties:

ev
P1. Receive Updated Routing Plan Property: once a new routing plan is issued by
the Routing Agent, the Localized Agent updates the planned routes of the Moving
Agent under its control and generates new corresponding instructions by executing

r
P2.

P2. Send Instructions to Moving Agent Property: as one-way means of commu-

er
nication, the Localized Agents send instructions for executing the next part of the
planned routes to the Moving Agents. This property is thus executed whenever a
Moving Agent has to change its speed or heading within the pre-defined execution
pe
time window wexec . Furthermore, coupling/decoupling as well as engine-start instruc-
tions are sent.

P3. Monitor Execution of Planned Routes Property: at each time point t, each
Localized Agent obtains the positions, headings, and velocities of all Moving Agents
under its control from the radar. The agent then estimates the time and speed at
ot

which each Moving Agent will pass its location, and shares these with the Routing
Agent. The Localized Agents also use the radar data to monitor that their instructions
are executed according to the planned routes. If deviations arise, they can primarily
tn

re-instruct the Moving Agents to close the gap. As last resort, they can execute P4
to trigger centralized re-planning of the routes.

P4. Request Replanning Property: with this property, the Localized Agents request
re-planning from the Routing Agent iff the deviation between a planned and executed
rin

route of a Moving Agent exceeds pre-defined thresholds.

P5. Communicate with Moving Agent Property: the Localized Agents communi-
cate with the Moving Agents under their control whenever required. Thus, they can
pass on information from or about a Moving Agent to the centralized agents.
ep

P6. Handover Control Property: whenever a Moving Agent has passed the location
of a Localized Agent, it transfers the respective Datastrip and thus the control re-
sponsibility to the closest next Localized Agent on the planned route of the Moving
Pr

Agent. Once a Moving Agent has reached its goal, the DS is placed in storage, and
can be utilized for post-operation analyses.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
4.5 Specification of Moving Agents
Moving Agents represent either an aircraft or a tug. They are modelled to have a circular size
with a shape diameter corresponding to their respective type. All aircraft are categorized
according to the six size-types of the ICAO aerodrome reference codes (ICAO, 2016), as

iew
listed in Table 1. Tugs are modelled with a diameter of 12 m based on the reference TaxiBot
system (2013).

ICAO-type shape diameter [m] exemplary aircraft type

ICAO-A 12 Cirrus Vision SF50

ev
ICAO-B 25 Cessna, Learjet
ICAO-C 40 A320, B737, EMB 170/190
ICAO-D 54 A300F, B767
ICAO-E 72 A350, B747, B787
ICAO-F 80 A380, B747-800F

r
Table 1: Categorization of aircraft sizes into ICAO-types

er
When consecutive aircraft take off from a runway, the Moving Agents have to keep a
time-based separation minima to ensure suﬀicient wake turbulence separation. The times
are dependent on the type-combination of leading and following aircraft, as listed in Table 2.
pe
The values were mapped to the ICAO-types from Table 1 on the basis of the RECAT-EU
time-based separation minima (Rooseleer & Treve, 2018).

Following
ICAO-A ICAO-B ICAO-C ICAO-D ICAO-E ICAO-F
Leading
ot

ICAO-A 80 60 60 60 60 60
ICAO-B 100 60 60 60 60 60
ICAO-C 120 60 60 60 60 60
tn

ICAO-D 120 100 80 60 60 60

ICAO-E 140 120 100 60 60 60
ICAO-F 180 160 140 120 100 60

Table 2: Minimal separation times in [s] for departing aircraft to mitigate wake turbulence.
rin

The aircraft types correspond to those defined in Table 1

The role of the Moving Agents is to execute the instructions that they received from the
Localized Agents. They thus have the following properties:
ep

P1. Execute Instructions Property: the Moving Agents execute the received instruc-
tions to the best of their possibilities to limit the amount of necessary re-instructions.
The trajectories provided to the Moving Agents include an allowable positioning tol-
erance around the speed profile in analogy to the Ownship-display used in (Bakowski
Pr

et al., 2017).

ed
P2. Communicate with Localized Agent Property: in any case, the Moving Agents
can communicate with the controlling Localized Agent to exchange status updates
and relevant operational information. They thus notify the Localized Agent whenever
a process such as tug decoupling or engine-start is estimated to finish. As further

iew
example, they could notify the Localized Agent that their preferred maximal speed is
lower than expected, so that the routing plans can be adapted accordingly.

5. Multi-Agent Motion Planning on Airport Surfaces (AS-MAMP)

To address the problem statement defined above, we have developed the multi-agent motion

ev
planning algorithm for ASM Ops (AS-MAMP): it is a two-level solver with a low-level
search to find individual agent trajectories, and a high-level search to coordinate all agents
so that their trajectories are conflict-free. AS-MAMP combines the newly developed Safe
Interval Motion Planning (SIMP) algorithm as low-level with an adapted version of PBS (cp.
Section 3.2) as high-level. In the following, we first provide a general overview of the AS-

r
MAMP algorithm, and elaborate on the underlying concepts in the successive subsections
of this section.

er
PBS de-conflicts agents whose space-time trajectories intersect by constructing a priority
ordering: it maintains a priority tree with each parent-node having up to two child-nodes
to establish a priority-relation between a conflicting pair of agents. Per child-node, a new
priority-pair is added that constrains one of the two agents to avoid the path of the other
pe
agent. Based on the sum-of-cost of the paths of all agents in each child node, the algorithm
then chooses the priority order that results in the lowest overall cost, and expands this node
next. In contrast to the original PBS algorithm, we define the cost of an agent trajectory as
a linear combination of its taxi time and taxi distance. The high-level search continues until
a child-node is expanded without any collisions. The changes to PBS are further outlined
ot

in Section 5.1, and the corresponding pseudo-code is provided in Algorithm 3.

A collision between two agents is detected by comparing their trajectories with each
other, accounting for their respective shapes that are defined in Section 5.3. To do so
eﬀiciently, we convert the plan of each agent into graph reservations. In short, all agents
tn

temporarily block the nodes and edges along their path as well as those swept by their
shape. The corresponding time durations per node, edge, or part of an edge form unsafe
intervals (USIs) during which other agents must not traverse the respective parts of the
graph underlying the airport layout. Further details on the graph reservations are discussed
rin

in Section 5.4, and Section 5.5 describes the conflict detection procedure.
To resolve a conflict, the plan of each of the two agents involved in the collision is
updated in the newly created PBS child-nodes by invoking the low-level solver. Its task
is to compute a new trajectory for a single agent that respects all reservations of agents
higher in priority. As such, we have developed the Safe Interval Motion Planning (SIMP)
ep

algorithm. It is based on the SIPP algorithm (cp. Section 3.2) as it reasons with safe time
intervals, defining states on nodes, and using motion principles for the mapping between
states. However, we modified the search by defining safe intervals not only on nodes but
also on edges, and embedded new mechanisms to deal with the kinodynamic restrictions,
Pr

the constraints set by the Routing Agent, and the reservations of agents higher in priority.
These are motivated and briefly described in the following.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Due to the kinodynamic constraints of Moving Agents, speeding up or slowing down
may take place over multiple edges. Additionally, it may not be possible to reach the
required lower speed to traverse curved segments. Therefore, we do not assume a constant
velocity per edge, but compute motions with piecewise constant acceleration. Per vehicle

iew
type, a maximal speed limit is set for both straight and curved edges. Furthermore, to avoid
stop-and-go movements, the agents need to have a minimal velocity with the exception of
special locations in the taxiway network to fulfil wait-activities. Instead of using a set of
predefined motions between states, we calculate feasible motions on-the-fly that respect the
kinodynamic agent constraints. To this end, we define motion principles as outlined in
Section 5.2.

ev
Due to the kinodynamically confined motions, and dependent on the reservations set by
other agents, reaching a node as early as possible does not necessarily lead to the fastest
motion over multiple edges. Therefore, in contrast to classical MAPF solvers and also the
original SIPP algorithm, past motions influence both the feasibility and cost of upcoming
states. As noted by (Ali & Yakovlev, 2023), this violates a core assumption of SIPP and

r
makes it incomplete. To handle the minimal speed requirements, we take their approach of
projecting wait intervals further: states are defined with Feasible Motion Intervals (FMI)

er
marking both time interval and speed range that can be reached from the previous state.
This approach is further outlined in the state definition and representation of the state
space (cp. Section 5.7).
We minimize the amount of generated states by defining a heuristic function that uses
pe
domain-specific circumstances to estimate the remaining cost to reach the goal location: it
takes the vehicle category, travel direction, and activity sequence including the start and
goal velocities into account (cp. Section 5.6). Furthermore, when generating successor states
(cp. Section 5.9), we anticipate on upcoming speed restrictions and potential overlaps with
the reserved space of other vehicles. Nonetheless, due to the nature of the motion principles,
ot

overlaps may remain along edges. If the contemplated motion over an edge overlaps with
the motion in the same direction of another vehicle, we analytically detect an overlap and
resolve it accordingly. If necessary, we use a backtracking scheme to do so (cp. Section 5.8).
The SIMP algorithm terminates if either a new plan is found, or no unexplored states are
tn

left, and the high-level search continues as described above.

5.1 Adaptations to PBS as High-level Search of AS-MAMP

The pseudo-code of the adapted version of PBS as high-level solver of AS-MAMP is provided
rin

in Algorithm 3 in Appendix A for reference. Changes to the original algorithm from (Ma,
Harabor, et al., 2019) are marked by red line-numbers.
PBS optimizes the sum-of-cost of the agent trajectories. Thus, we define the cost of a
PBS search-node N as
∑A
ep

cN = (cprio )i ∗ (cplan )i (1)

i=0

with the number of agents A, the cost cplan of the trajectory of agent i, and (cprio )i as
dimensionless cost to weigh the importance of the route of agent i for the overall optimization
Pr

goal. For instance, aircraft can be assigned a higher value for cprio than tugs to increase
their prioritization: an increase in plan cost of the aircraft is penalized more than that of the

ed
tug, potentially influencing the resulting priority order. The cost of each agent’s trajectory
is defined as linear combination of taxi time ttaxi and taxi distance dtaxi :

cplan = (ttaxi + cd ∗ dtaxi ) (2)

iew
with cd [s/m] translating the travelled distance into a cost with unit time. The low-level
solver of AS-MAMP uses the same cost-function to determine an agent trajectory, which is
further outlined in Section 5.6.
In the root-node of PBS, the trajectories of all agents are calculated taking merely the
constraints set by the Routing Agent (cp. its P3 and P3) into account. This represents the
single-agent motion planning (SAMP) solution of the AS-MAMP search for each Moving

ev
Agent. Note that a partial priority-ordering can be set in the root-node of the adapted
version of PBS as well.
The next search-node is picked based on the chosen search-mode: in this work, we dis-
tinguish between either a depth-first, best-first, or greedy search. The depth-first search

r
uses the top node of the stack as new parent-node. Newly created child-nodes are thus
placed on top of the stack in order of decreasing cost. The greedy search follows the
approach of Chan et al. (2023) to greedily place the child-node with the lowest number of

er
remaining conflict-pairs on top of the stack irrespective of its sum-of-cost. However, in
contrast to the original Greedy PBS (GPBS) implementation, we do not alter the low-level
solver. With these two search-modes, PBS only backtracks when it cannot find a solution
pe
in the explored branch. While these quickly lead to a solution, the resulting priority or-
derings may be sub-optimal and thus lead to higher taxi times of the agents. In contrast,
the best-first search returns the optimal solution within the priority-tree as it continues
to pick the node with the lowest cost in the priority-tree until the solution is conflict-free.
However, this search-mode is both time-wise and memory-wise only feasible for small agent
populations. In Section 6, we use this search-mode to assess the suboptimality of the depth-
ot

first and greedy searches for the analysed case studies. Note that other search-modes can
be used as well, such as a focal search, or other means to pick child nodes that lead to a
faster and/or higher quality solution. We will explore such in future research.
tn

The authors of GPBS also introduced the concept of using induced constraints (IC)
to further speed up the high-level search. While their use does not necessarily sacrifice
solution quality, it does affect the chosen conflict to be resolved (cp. Section 5.5) and thus
the resulting priority order. We adopt this technique as variation of all three search-modes,
and study its influence in Section 6. In the following, we will outline the remaining building
rin

blocks of AS-MAMP, starting with the motion principles.

5.2 Motion Principles

To model the motions of the vehicles on the airport surface along the taxiway network, we
ep

make the following assumptions:

• the motions are approximated with piecewise-constant acceleration

• since the radius of curvature r of all taxiway segments is large in comparison to the
Pr

travelled distance and the change in velocity within that distance, we neglect the
normal component of the acceleration

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
• despite that, velocity limits have to be met to traverse curved taxiway segments that
have a radius of curvature below a pre-defined threshold value.

These allow us to use the equations for rectilinear motions, i.e. with constant accelera-
tion along a straight line:

iew
acceleration a(t) = const. in [m/s²] (3)
velocity v(t) = vi + a t in [m/s] (4)
1
distance d(t) = di + vi t + a t2 in [m] (5)
2

ev
with elapsed time t, initial velocity vi , and initial distance di . Since we consider relative
distances when traversing taxiway segments, di = 0. Rewriting these equations yields:
√
final velocity vf = vi + a t = vi2 + 2 a d (6)

r
2d vf − vi
traversal time t= = (7)
vi + vf a
travelled distance
er 1
d = vi t + a t 2 =
2
vi + vf
2
t

We apply these equations to calculate on-the-fly motions along the graph edges during
(8)
pe
single-agent path planning. Since the above equations are only valid for constant accel-
eration, we split motions into piecewise segments of constant acceleration if needed, and
allow for kinks both at nodes and anywhere on edges. In general, the velocity profile is
optimized for smallest traversal time and highest final velocity. The velocity is bounded
by an agent-dependent maximal velocity vmax as well as a minimal velocity vmin when in
motion.
ot

During planning, reaching a full stop is only allowed at the end-nodes of edges. The
agents can wait and hold arbitrarily long at these, as long as no reservations from other
agents exist. While waiting is bound to nodes, holding may also occur at the beginning
(but not at the end) of a motion along an edge. By convention, a motion always connects
tn

the state at the beginning of an edge with the state at the beginning of a subsequent edge.
A trajectory over multiple edges is thus split into one motion per edge before generating
the resulting states on the respective nodes (cp. Section 5.7). Note that the duration of
motions can have arbitrary values, allowing for execution in continuous time.
rin

Fig. 3 illustrates different motions over one or multiple edges with a path length of
d = 250 m. The exemplary motions are plotted in both a (d,t)-graph (top) and a (d,v)-
graph (bottom). Speed limits and minimal traversal times as restrictions are depicted by
red shades. In the first column, a velocity limit of vf,lim = 5 m/s must be reached at
the end of the path. Thus, to minimize the traversal time, acceleration from vi < vf,lim
ep

to vf = vf,lim takes place immediately (blue line), while deceleration from vi > vf,lim to
vf = vf,lim happens towards the end of the path (orange line).
In contrast, when a minimal traversal time due to time restrictions has to be exceeded
(second column), the motion is optimized for maximal final velocity: acceleration is carried
Pr

out towards its end (blue line), while deceleration occurs immediately (orange line). Notice
that while both motions take the same time, the blue motion will have a significantly shorter

ed
YUHVWULFWLRQ WUHVWULFWLRQ YWUHVWULFWLRQ DFFGHFPRWLRQ

WLPHW>V@

iew

YHORFLW\Y>PV@

r

GLVWDQFHG>P@
Figure 3: Exemplary motions visualized per column both in (d,t)-graph (top) and (d,v)-
graph (bottom). The velocity and time restrictions for the different examples are marked
by red shades.
er
pe
traversal time on the subsequent edges due to its higher final velocity. This highlights that
the fastest overall motion depends not only on past but also upcoming restrictions. Arriving
as fast as possible in each node is thus not always ideal.
The two remaining columns depict combinations of both velocity and time restrictions.
ot

As convention for regular motions shown in the third column, acceleration (blue line) and
deceleration (orange line) take place in between the two segments with constant velocities
vi and vf = vf,lim so that both constraints are met. In contrast, when vf,lim = 0 m/s
(green line), a constant velocity vconst with vf > vconst ≥ vmin is used in between the two
tn

deceleration segments to meet the minimal traversal time. Note that a set of constraints
requiring vconst < vmin lead to an infeasible motion by definition.
The regular motions with their monotonic profiles described up to this point fall short
of special cases such as vi = vf,lim = 0 m/s. Furthermore, such profiles are too constraining
rin

when motions are generated over multiple edges as necessary for backtracking and an-
ticipation (cp. Section 5.8). Therefore, we define acc-dec-motions that comprise both
acceleration and deceleration segments to/from the intermediate velocity vmid . vmid results
from the restrictions at the beginning and end of the travelled distance, and is bound by
vmax and vmin . Thus, to cover the entire distance, the acc-dec-motion can also form an
ep

in-between segment with constant velocity vconst = vmid . Note that by convention, except
for holding at the start of an acc-dec-motion, a constant velocity is neither allowed at its
start nor its end.
The last column in Fig. 3 shows examples of such motions. In comparison to the third
Pr

column, the acc-dec-motion allows the blue and orange motions to reach the end of the
respective other time restriction depicted by the red shades. The green and dotted red

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
lines depict the special case mentioned above: while the dotted red motion ends in the red
shaded area and is therefore infeasible, the green motion includes holding at its start and
thus exceeds the minimal traversal time. For these cases, a maximal holding time thold,max
can be defined as additional constraint if needed. Potentially, this results in a constant

iew
intermediate segment with a smaller vmid .

5.3 Definition of Motion Profiles, Shapes, and Safety Distances

In general, agents have a circular shape with a radius according to their type as listed in
Table 1. Each shape has a center, front, and back, as illustrated in Fig. 4. The center
represents an agent’s reference point as defined in (Li, Surynek, et al., 2019). A motion

ev
thus specifies the movement of the center. We refer to the center’s collective motion over
multiple edges as a motion profile, i.e. the agent’s four-dimensional trajectory (4DT).
The motion profile specifies the time points and velocities at the absolute distances at
which the acceleration changes. Therefore, for each segment in the motion profile with

r
constant acceleration, we can use Eqs. (6) to (8) to determine any time point t, travelled
distance d as absolute or relative value, and velocity v on the continuous scale. As the front
and the back of a shape move in unity with the center, their respective motion profile is

er
obtained by shifting the absolute distances of the central motion profile by the radius of
the shape. Then, the corresponding motion over an edge in the path of the agent can be
determined in analogy to obtaining a single point in the profile as described above.
pe
In real-world operations, taxiing vehicles have to keep a minimal safety distance. For
aircraft, this is mainly due to the engine blast that poses a safety risk to any object or
person in their wake. We know from expert interviews that other vehicles should keep a
safety distance of around twice the length of the aircraft in front of them to mitigate the
risk of engine blast. The experts highlight that the actual safety distance maintained by
pilots and ground vehicle drivers often falls short of the recommended length. Since pilots
ot

and auto-pilots can only follow planned trajectories with limited precision (cp. Requirement
7c), we include an allowable deviation as positioning tolerance within the safety zone.
In AS-MAMP, we hence define that all agents have to keep a minimal safety distance dS
tn

between each other in dependence of their shape-radii as well as the positioning tolerance
δpos for each agent. Since the taxiway network is represented as a graph, we distinguish
between the minimal safety distance between agents on the same path (cp. Fig. 4.1), and
that between agents on different paths (cp. Fig. 4.2). When travelling on the same path,
we define the minimal safety distance between two agents in relation to the shape-radius of
rin

the leading agent. As illustrated in Fig. 4.1a, for an agent ai following an agent aj along
the same path, their minimal safety distance dS,j is defined as

dS,j = σpath · rj + δpos,i + δpos,j (9)

with the safety factor σpath multiplying the shape-radius rj of agent aj , and δpos for each of
the two agents. We make use of the safety factor to provide an operational parameter that
can be adjusted when studying airport surface movement operations with the multi-agent
system. Nonetheless, we recommend to use values for σpath between 2 − 6 to match the
Pr

practical considerations of the experts. We will evaluate the effect of different positioning
tolerances with a sensitivity analysis of δpos in Section 6.

ed
(1) on the same path (1b) dS,j dS,j
center (1a) ri 2 2 rj
ri dS,j
ai
rj
δpos,i σpath∙rj δpos,j
aj

iew
back front back front

(1c)
safety zone path
ri dS,j rj
travelling direction
back front
following leading center

ev
agent agent
back front

(2) on different paths along neighbouring edges

ai (2a) (2b) (2c)

r
ri

safety zone
er
dS,ij

ri
dS,ij

rj
pe
dS,ij rj
aj

rij
ot

Figure 4: Definition of shapes and safety distances when two agents ai and aj travel (1) on
the same path and (2) on different paths
tn

When travelling on different paths, we define the minimal safety distance between two
r +r
agents ai and aj in relation to their average shape-radius i 2 j as
r i + rj
dS,ij = σN E · + δpos,i + δpos,j (10)
rin

2
with the safety factor σN E . Since the risk of engine blast is lower when agents do not
directly trail each other, we recommend to use values for σN E between 1 − 3.
When determining the remaining gap between two moving shapes, different representa-
tions of their shapes are possible, which is illustrated in the sub-figures (b) and (c) in Fig. 4.
ep

For two agents on the same path, half of the minimal safety distance can be attributed to
each agent, respectively (1b). This representation is useful in conflict detection, as outlined
in Section 5.5. Since SIMP plans a new trajectory for the center of a shape, (1c) depicts a
representation in which one of the agents is represented as point while both shape-radii as
Pr

well as the minimal safety distance are associated to the other agent. The same approach is
feasible for shapes on neighbouring edges (cp. 2c). Additionally, when setting graph reser-

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
vations as outlined in Section 5.4, the representation in (2b) is needed: it attributes both
shape-radii as combined shape to a single agent while the other agent keeps the minimal
safety distance as safety zone with radius dS,ij . To this end, we define the radius of the
combined shape as rij = ri + rj .

iew
5.4 Graph Reservations
To account for dynamic obstacles during single-agent path planning (i.e. the low-level
search), the motions of other agents must be known to the planner. While the original SIPP
implementation defines collision intervals to do so, these do not account for the shape and
kinematics of other agents. As noted by Walker and Sturtevant (2019), an algebraic equation

ev
to detect collisions between accelerated motions must still be derived. The alternative
approaches reviewed in Section 3.2 draw on assumptions that are invalid for graphs with
arbitrarily long, curved, and oriented edges. Moreover, we must also account for layout
constraints such as temporarily blocked edges due to changing RMOs during planning.

r
Therefore, we outline our approach to handle dynamic obstacles and time constraints as
graph reservations in the following.
Once the SIMP algorithm has terminated for an agent, the resulting plan is converted

er
into graph reservations. These allow us to account for the agents’ shapes and are needed
both for conflict detection and when updating the plans of agents lower in priority. We
distinguish between three types of graph reservations:
pe
• path-reservations on edges along the path of an agent;
• neighbourhood-reservations on all edges not on the path of an agent but swept
by its shape, which includes edges on the path in reverse direction;
• WTC-reservations on runway entries to ensure suﬀicient temporal separation ac-
cording to the wake turbulence category (WTC) of an agent with respect to other
ot

agents.
Path-reservations are necessary to let an agent follow – but not overtake – a higher
prioritized agent that is travelling in front of it, while keeping the minimal safety distance
tn

as defined in Section 5.3. Such reservations are illustrated in the top part of Fig. 5: the
orange agent AC-2 (ICAO-C) traverses the edges (1,2), (2,3) and (3,4) towards the runway
entry (4,10). It is followed by the green agent AC-1 that has to use the runway entry (8,11)
due to its larger size (ICAO-E). They thus share the edges (1,2) and (2,3). To reserve the
rin

traversed path, the motion of an agent’s center as well as the radius of its shape are stored
separately per edge as path reservation.
In contrast, neighbourhood-reservations are set from an agent traversing an edge
onto neighbouring edges. In the illustrated example in Fig. 5, the orange agent AC-2 cannot
traverse edge (3,4) as long as the blue agent AC-3 still occupies edge (6,5), since their shapes
ep

would collide. A neighbourhood-reservation comprises two parts to represent the space-time

collision interval: the reservation is set for the edge-interval as reserved segment of the other
edge which must not be traversed by other agents during the time-interval as its second
part.
Pr

An edge-interval marks the start and end positions normalized by the length of the edge
for which the reservation is set. These positions result from the geometrical overlap based

ed
10 11
runway
entry
following leading
agent agent stopbar
rE 4 8

iew
rC
dS,C

1 2 3 5 7
AC-2
AC-1 (ICAO-C)
safety zone
(ICAO-E)
AC-3

ev
(ICAO-E) 6 9

path-reservations and exemplary conflict

safety zone dS,C rC rE
t

r
back center
AC-1 front

v
AC-2 overlap = conflict
er vturn
vmax
d
pe
d
1 2 3 4 5

neighbourhood-reservations and exemplary conflict

a) spatial overlap of shapes l(3,4) l(3,4)
δ (3,4)
ot

2
δ2(3,4)
δ (3,4)
rC rE 1
l(6,5) spatial
δ2(6,5) overlap
tn

δ1(3,4)
δ1(6,5)
l(6,5)
rCE δ1(6,5) δ2(6,5)
b) temporal overlap of shapes
center
rin

t dS,CE t dS,CE temporal overlap = conflict

τ2(3,4) τ2(6,5)
τ1(6,5) tδ1AC-2 τ2(6,5) tδ2AC-2
tδ2 AC-2 safety tδ2 AC-3
reserves zone reserves edge (3,4)
tδ1 edge (6,5) tδ1 edge (3,4)
τ1(6,5) edge (6,5)
ep

τ1(3,4) d d t
δ1(3,4) δ2(3,4) δ1(6,5) δ2(6,5) tδ1AC-3 τ1(3,4) tδ2AC-3 τ2(3,4)
3 4 6 5

Figure 5: Graph reservations and exemplary conflicts (illustrated in top part), with path-
reservations on edges along an agent’s path (middle part) and neighbourhood-reservations
Pr

on all other edges that are swept by an agent’s shape (bottom part)

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
on the combined shape of two agents as defined in Fig. 4.2b. For the illustrated example, the
bottom row of Fig. 5 depicts the radius rCE as combined shape of AC-2 with type ICAO-
C and AC-3 with type ICAO-E. With the combined shape, we calculate the geometrical
overlap between a pair of edges: when the combined shape traverses edge (3,4) it overlaps

iew
(3,4) (3,4)
with edge (6,5) between δ1 and δ2 . The overlap-interval of edge (6,5) with edge (3,4)
(6,5) (6,5)
is likewise calculated as δ1 and δ2 . These normalized positions on the edge-pair form
the edge-intervals of each edge when setting a neighbourhood-reservation from the other
edge. Since the spatial overlap between an edge-pair depends on the combined shape of
two agents, each combination of two agent types yields one edge-interval-pair per edge-pair.
The intervals are pre-calculated for all edge-pairs and shape-combinations.

ev
The time-interval as second part of a neighbourhood-reservation is determined with
the motion profile over an edge: as outlined in Section 5.3, we determine the time points
at which the safety-zone around the point-shape of an agent (green agent ai in shape-
representation 2b in Fig. 4) enters and leaves the normalized positions δ1 and δ2 on the

r
edge from which the reservation is set. In the example, agent AC-2 reserves the edge
(3,4) (3,4)
(6,5) between τ1 and τ2 , given the safety distance dS,ij around it. This results in
(6,5) (6,5)
the neighbourhood-reservation on edge (6,5) between the normalized positions (δ1 , δ2 )

(3,4) (3,4)
(3,4) (3,4)
er
as edge-interval during (τ1 , τ2 ) as time-interval. Likewise, agent AC-3 reserves the
(6,5) (6,5)
edge-interval (δ1 , δ2 ) of edge (3,4) for the time-interval (τ1 , τ2 ).
We account for the required takeoff-separation according to the wake turbulence cat-
pe
egory (WTC) of a departing aircraft by defining WTC-reservations. These are set on
the directional edges of all entries to the respective runway as time intervals per type-
combination of the agent with any other aircraft type. A time interval consists of the start
time point ts and end time point te :

ts = tStB − tW TC
(i departs after j) (11)
ot

i j→i
te = tStB
i + tW TC
i→j (i departs prior to j) (12)

with tStB
i denoting the time at which the center of agent i is at the stopbar before entering
tn

the runway, and tW T C being the durations required for the WTC-separations between agents
ai and aj . In the example in Fig. 5, both agents AC-3 and AC-1 use edge (8,11) as runway
entry. Presuming that AC-3 will be in front, it reserves the runway entries (8,11) and
(4,10) between ts and te with tStB = t8 as the time of passing the stopbar at node (8) and
tW TC WTC
j→i = ti→j = 60 s from Table 2.
rin

Before updating the plan of a de-prioritized agent, all reservations from agents higher in
priority are gathered. The illustrated collisions in Fig. 5 are thus resolved when a priority-
order between the three agents is set. The next subsection outlines how the necessary
conflict detection is carried out.
ep

5.5 Conflict Detection

Based on the graph reservations per agent, we execute the conflict detection scheme to
find all conflicts between each pair of agents. The corresponding pseudo-code is outlined
Pr

in Algorithm 4 in Appendix A for reference. We use an approach based on an exhaustive

search: the reservations of each agent ai are compared against those of every other agent

ed
aj . Conflict detection is skipped if a priority order already exists between the two agents,
or their plans are disconnected in time. Otherwise, we compare the path-, neighbourhood-,
and WTC-reservations of agent ai with those of agent aj .

iew
The agents are in a path-conflict if their path-reservations overlap on an edge that
both of them traverse. To detect such a conflict, we first determine the sequence of the
agents passing the edge. In the example in Fig. 5, agent AC-2 is in front of agent AC-1 on
the edges (1,2) and (2,3). Using the shape-representation 1b in Fig. 4, we then compute
the minimal gap between the front-line of the shape of the trailing agent (AC-1) with the
back-line of the shape of the leading agent (AC-2). Note that half of the minimal safety
distance is included in each shape as depicted in Fig. 4.1b. The minimal gap between the

ev
two shifted motion-profiles (cp. Section 5.3) is analytically calculated by determining the
minimal distance between the underlying motion-functions. If the distance is negative, the
motions overlap, and thus form a path-conflict. As marked in the example, the motions
of AC-1 and AC-2 overlap on edge (2,3), since AC-2 slows down for the curved edge (3,4)

r
while AC-1 can keep its maximal velocity to go straight ahead along edge (3,5).

er
In contrast, agents are in a neighbourhood-conflict if the neighbourhood-reservations
of one agent overlap with the motion of the other agent’s center on the same edge. Recall
that a neighbourhood-reservation consists of a time-interval during which another agent
must not traverse the corresponding edge-interval of the edge for which the reservation is
pe
set. To detect such conflicts, we iterate over each edge in the plan of agent ai . For each
neighbourhood-reservation of agent aj , we determine the time-interval at which the center
of ai enters and leaves the normalized edge-positions according to the edge-interval of the
reservation. The agents are in a neighbourhood-conflict if the resulting time-interval of ai
overlaps with the time-interval of the neighbourhood-reservation set by aj . In Fig. 5, such
a conflict occurs between agent AC-2 and agent AC-3 between the edges (3,4) and (6,5):
ot

the motion of the safety zone around the center of AC-2 traversing edge (3,4) overlaps with
the neighbourhood-reservation of AC-3 that are set for that edge. Likewise, the reservation
of AC-2 on edge (6,5) overlaps with the motion of AC-3 over that edge.
tn

In analogy, agents are in a WTC-conflict if the WTC-reservation of one agent overlap

with the motion of another agent on any edge that is an entry to the same runway. Using
Eq. (12), the time interval of agent ai occupying the runway is determined. The agents are
in a WTC-conflict if it overlaps with the time interval that agent aj reserved on all runway
rin

entries.

The conflict detection scheme is applied to every combination of two agents and yields
all existing conflicts. In general, the first one in time is selected. However, when using
a PBS search-mode with induced constraints (IC), we follow the approach of GPBS with
ep

IC (Chan et al., 2023): the agent pair with the largest IC-set is selected, breaking ties by
choosing the pair with the smallest set of agents lower in priority. This is complemented by
breaking ties in favour of the earliest conflict in time. To resolve the selected conflict, recall
that two new PBS search-nodes are created with each conflicting agent being added to the
Pr

others priority list. Per search-node, a new list of conditions for the de-prioritized agent is
formed and SIMP is invoked.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
5.6 SIMP: Objective Function and Heuristic Search
Like A* and SIPP (cp. Section 3.2), the Safe Interval Motion Planning (SIMP) algorithm
searches for a path from a start to a goal location by minimizing the associated cost of the
resulting trajectory. The f-score f (S) of a state S estimates the path cost:

iew
f (S) = g(S) + h(S) (13)

with the g-score g(S) as accrued cost between the start location and S, and the h-score h(S)
as estimation of the remaining cost to reach the goal. Apart from using the shortest distance
to the goal, h(S) is also based on the current velocity vS at S, the velocity restriction at the

ev
goal, and the agent’s maximal velocity vmax . Furthermore, we take the current travelling
direction, the agent’s activity sequence, and the agent’s vehicle type into account. The
latter determines which edges of the graph can be traversed in general, as some edges are
not accessible to all agents: for example, service roads cannot be used by aircraft, and some
taxiways are wingspan-restricted.

r
To force SIMP to follow the agent’s activity sequence, we calculate h(S) as sum of the
heuristic per activity h(acti ): ∑
h(S) =
er(h(acti )).
For go-to activities, multiple end nodes may be defined (cp. P5 of Routing Agent). In that
case, the minimum of each h(S) per combination j of end nodes within the sequence is
(14)
pe
chosen so that h(S) remains consistent:

h(S) = min(hj (S)) (15)

Each h(acti ) is calculated by using Eq. (2) with dtaxi = dh (acti ) and ttaxi = th (acti ) as
well as the cd defined for the respective agent. For dh (acti ), the node-to-node shortest
ot

distance along the graph edges between the start and end node of the respective activity
is used in general. However, since aircraft cannot turn on the spot at a state S in their
current activity actS , we use an edge-to-node shortest distance for dh (actS ) based on the
directional edge in the travelling direction of the agent. These distances are pre-calculated
tn

for all edge-node-combinations in the layout. Per dh (acti ), and dependent on the velocities
at the start and end nodes of acti , we make use of the motion principles to calculate the
corresponding minimal traversal time th (acti ). Note that for waiting-activities, dh = 0 and
th equals the waiting duration.
rin

Throughout the search, to ensure that the heuristic is admissible and consistent, we
verify that
h(S) ≤ cm (S → S ′ ) + h(S ′ ) (16)
with cm (S → S ′ ) representing the motion cost between state S and its successor S ′ and
ep

h(S ′ ) denoting the h-score from S ′ to the goal. cm is calculated by using Eq. (1) with
ttaxi = tm as the duration of motion m and dtaxi = dm as the traversed distance.

5.7 SIMP: State Space

SIMP calculates the agent trajectories that are, once the AS-MAMP algorithm has found a
conflict-free solution, used as routing plans by the Localized Agents to instruct the Moving

ed
Algorithm 1: Safe Interval Motion Planning (SIMP) Algorithm
Input: route parameters of agent, reservations of higher prioritized agents
1: closed ← ∅

iew
2: open ← getInitStates() # creates start states Sinit
3: while open ̸= ∅ do
4: S ← state in open with lowest f-score
5: if S fulfils goal-conditions then
6: plan, cost ← getPlan(S) # backtracks plan from S to Sinit
7: return plan, cost # solution found: SIMP can terminate
8: if S not in closed then

ev
9: add S to closed
10: else
11: chooseStateToKeep(S) # same state-ID in closed: keeps the state with
lower f-score and updates child-states
continue # successors of this state were already created before

r
12:

13: successors ← getSuccessors(S)

14: foreach S* in successors do
15:
16:
17:
cfg, FMI ← extract from S*

open ← open ∪ S*
er
S* ← make FMI of S* temporally disjunct
# indexed by (cfg, FMI)
pe
18: return ∅, ∞ # no solution found
Output: plan from start to goal node and its cost, if existent
ot

(1): 300m tSmax

tSmin
S3

M2
P

M1 SP M1

(2): 500m
SP
vSmax
rin

M2 P

(3): 750m
vSmin
P
S4
ep

Figure 6: Exemplary trajectories through layout (left) and state space (right) to a stand
Pr

that is occupied till t = 100 s resulting in a long goal constraint (red shade at 0 m)

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Agents during plan execution. To this end, SIMP shall output plans that consist of an
agent’s configuration at the beginning of each edge along the path with a feasible motion
following the motion principles from Section 5.2 connecting two entries in the plan. As
SIMP plans with states, we require that each state is defined on a node N , with a next

iew
node N ′ provided as travel direction. This indicates which edge is about to be entered by
the following motion. Furthermore, as the agents must follow the activity sequence defined
by the Routing Agent, each entry in the plan shall contain the activity-ID actID that is
valid from the state onwards.
In SIMP, we keep track of the state space with an open-list and a closed-list, in
analogy to other search algorithms such as A* and SIPP. However, the purpose of these

ev
two lists differs, influencing how states are identified. Furthermore, all states also store
additional information, such as a reference to their parent-state, the g-, h-, and f-scores, the
arrival distance, and velocity vmid to indicate an acc-dec-motion (cp. Section 5.2). In the
following, we outline these state definitions along with the core part of the SIMP algorithm,
which is provided in Algorithm 1.

r
The purpose of the open-list is to mark unexplored states. To follow the logic behind
SIPP, open shall contain states that cluster the motions along the following edge towards

er
the subsequent node into intervals to reduce the number of similar states in the search
space. To this end, we define the agent configuration of state S in open as

Scf g = (N , N ′ , vlim , actID ) (17)

pe
with the speed limit vS,lim dependent on the edge leading to N as well as the entered edge
(N, N ′ ).
However, since agents cannot stop instantaneously, the safe time intervals that SIPP
uses are not applicable as such, as noted by (Ali & Yakovlev, 2023). The minimal speed
requirement at most edges and nodes further complicates this issue, illustrated in Fig. 6:
ot

the designated stand of a landing aircraft is still occupied for a while, i.e. a long goal-
constraint is placed on its goal-node, indicated by the red shade at 0m in the upper right
plot. Since the aircraft is neither allowed to stop along the direct route to its stand, nor
travel slower than the minimal velocity, it must take a detour. In the low-level search, states
tn

along path (1) in blue, then along path (2) in orange, and finally along path (3) in green
would be created. However, in vanilla SIPP with a single safe time interval, the state of the
orange path at the merge-location M1 would be discarded as its f-score is higher than the
already existent state of the blue path at M1. Likewise, the state at M2 of the green path
rin

would not be added to open. Thus, the low-level search would fail to find a solution.
To address this issue, we take the approach of (Ali & Yakovlev, 2023) further, and define
Feasible Motion Intervals (FMI) marking both arrival time interval and speed range that
can be reached from the previous state satisfying the kinodynamic constraints. We define
the FMI of a state S in open as
ep

SF M I = (tS,min , tS,max , vS,max , vS,min ) (18)

with the quickest feasible motion arriving at tS,min with speed vS,max , and the slowest
feasible motion arriving at tS,max with vS,min . For the velocities,
Pr

vmin ≤ vS,min ≤ vS,max ≤ vS,lim , (19)

ed
with the minimally allowed velocity vmin . For illustration, in the upper right plot in Fig. 6,
the arrival time interval of state SP is visualized by the shaded green area between the
solid and dotted lines representing the earliest arrival time tSP ,min and latest arrival time
tSP ,max , respectively. Likewise, the lower right plot shows the speed range of SP between

iew
vSP ,max = 15 m/s (solid green line) and vSP ,min = 5 m/s (dotted green line).
In open, each state-ID comprises a configuration and an FMI. For an existing configu-
ration in open, we ensure that the states have temporally disjunct FMIs: for two states S
and S ∗ with the f-scores fS < fS ∗ , we adjust the FMI of S ∗ so that tS ∗ ,min ≥ tS,max and
vS ∗ ,max ≤ vS,min . In case tS ∗ ,max < tS,max , we do not add this state to open. If fS = fS ∗ ,
we keep the FMI with the earlier arrival time, and adjust the other accordingly. When the

ev
FMI is adjusted, the g-, h-, and f-scores are recalculated as well.
At the merge location M1 in the example in Fig. 6, the FMI of the orange state does
not overlap with that of the blue state, and is directly added to open. At M2, the FMIs of
the orange state So,M 2 and green state Sg,M 2 are respectively defined as

r
So,F M I = (tS,min , tS,max , vS,max , vS,min )o = (23 s , 45 s , 13 m/s , 5 m/s)
Sg,F M I = (tS,min , tS,max , vS,max , vS,min )g = (38 s , 95 s , 13 m/s , 5 m/s)

er
which overlap on the interval between (38 s, 45 s). The g-score of the green state must be
higher since its path is longer. Therefore, fSo ,t=38s < fSg ,t=38s . As a result, the motion of
the green state must be adjusted to create a temporally disjunct FMI: its adjusted earliest
pe
arrival time tSg ,min ∼
= tSo ,max = 38 s with a speed of vSg ,max = vSo ,min = 5 m/s. Subsequent
states are created in analogy to keep the FMIs temporally disjunct. In Fig. 6, this is
visualized by the solid green line matching the dotted orange line.
To create a trajectory with minimal cost (cp. Section 5.6), the state S in open with
the lowest f-score is picked and removed. Successors of S are generated by executing the
ot

function getSuccessors, which is further outlined in the next subsection. S is then moved
to the closed-list (Lines 8 to 12). Once a state in open is picked that meets the goal
conditions, SIMP terminates and returns the resulting plan and cost (Lines 5 to 7).
In contrast to clustering states, the purpose of the closed-list is to quickly obtain
tn

a parent-state during the search as well as generate the final plan once the search has
terminated without recalculating the motions between two states. We thus define the unique
state-ID SID for states in closed as

SID = (N , N ′ , tS , vS , actID ).
rin

(20)

with the arrival time tS and arrival speed vS at node N as well as the activity-ID actID
that is valid from the state onwards. If a state with identical state-ID is added to closed,
SIMP keeps the one with the lower f-score, and updates its state-values as well as those of
ep

all its successors accordingly. Correspondingly, all child-states are recursively updated.

5.8 SIMP: Backtracking and Anticipation

Continuing with the example in Fig. 6, the aircraft still arrives too early at the goal when
Pr

following the resulting green trajectory (solid line). Thus, the motions along the previous
edges must be adjusted so that the goal is reached just after the end of the goal constraint,

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
as illustrated by the purple trajectory along path (3). We use this case to introduce two
schemes that form an integral part of SIMP: the backtracking scheme and the anticipation
scheme.
The purpose of the backtracking scheme is to find a motion from a previous state that

iew
allows to create a state S with a predefined time tS and speed vS , which are tS = 100 s
and vS = 0 m/s in the example to meet the goal conditions. To this end, the motion from
a parent state along multiple edges to state S is checked for feasibility. This is recursively
done until a feasible motion along the combined path is found. Nonetheless, this scheme
may fail if the acc-dec-motion violates any constraints along its path. In the example, the
backtracking scheme has to iteratively go back to the parent state SP at approximately

ev
400m remaining distance to the goal to create this combined motion, as shown by the
purple line. To satisfy the requirement that each state must be defined on a single edge,
the motion along the combined path must be split into a motion per edge of that path. We
refer to the new set of states created along the path to S as intermediate states. Since they
were created specifically to form a motion to S, the intermediate states are added directly

r
to closed, while S is added to open.
In contrast, the purpose of the anticipation scheme is to anticipate on time and velocity

er
restrictions at upcoming edges that require a motion from SP to allow for feasible successor
states. This allows to create the purple trajectory directly from SP , and avoids the recursive
calculation of motion feasibility that are necessary in the backtracking scheme. Although we
pe
primarily use the anticipation scheme in SIMP, we can only anticipate on those constraints
that are valid for a node or an entire edge. We thus need the backtracking scheme to resolve
for instance infeasible motions that overlap with path-reservations of other agents that vary
over an edge.

5.9 SIMP: State Generation

In this subsection, we outline the generation of successors based on the pseudocode of the
function getSuccessors and an illustrating example provided in Fig. 7. Due to our state
definition, multiple successor states S ′ may be defined for traversing the edge on which state
tn

S is defined. To limit the amount of repetitive computations for the motion along an edge,
we first gather the velocity restrictions and safe time intervals that must be satisfied to
enter the next edges. These form multiple sets of boundary conditions for the motion along
an edge. We cluster them along with the next edges that require an identical set of these
rin

motion boundaries (Line 5 in Algorithm 2). Furthermore, we sort them based on earliest
arrival time within the safe interval, breaking ties in favour of highest speed limit vlim .
Then, following the sorted order, a motion is computed that satisfies the set of conditions,
and the respective successors are created (Lines 8 to 22).
In the example in Fig. 7, three operations are needed to create five regular successors
ep

(ID1 to ID5): first, the motion to enter edge (3,5) is created taking 12 s with vi = vf =
15 m/s (ID1). Then, meeting the speed limit of vlim = 5 m/s for edges (3,4) and (3,6), the
fastest motion takes 16.4 s, arriving within the first SI of each edge (ID2 and ID3). Finally,
the successors ID4 and ID5 are created to arrive at the beginning of the second SIs on edges
Pr

(3,5) and (3,6). To form a motion that spans 40 s, it is necessary to immediately start to
decelerate from vi to vf = 1.4 m/s, and keep a constant velocity for the remaining length

ed
Algorithm 2: SIMP function getSuccessors(S)
Input: state S
1: successors ← ∅
## special case: wait-activity, switching to next activity

iew
2: if S.activity = "wait-activity" then
3: S’ ← getSuccessorWaitActivity(S)
4: return S'

## get motion restrictions for next edges

5: motionBounds ← getMotionBounds(S) # based on v-restrictions
6: if not motionBounds then

ev
7: return ∅ # no way forward

## calc motions for next edges. Start with quickest possible motion
8: motionBounds.sorted() # sort for quickest motion
9: motionPrev ← ∅

r
10: foreach motionBound in motionBounds do
11: if motionPrev slower than motionBound allows then
12: continue # previous motion slower than this bound, skip recalculation
13:
14:
15:
16:
motion ← calcMotion(S, motionBound)
foreach edgeID in nextEdgeIDs do
S’ ← createSuccessor(motion)
successors ← successors ∪ {S’}
er # calculate motion based on bounds

# create successor based on motion

pe
## use backtracking iff vFinal < vLimit
17: if vFinal < vLimit then
18: S” ← createSuccsBacktrack()
19: successors ← replace S’ with S”
## anticipate on upcoming motion restrictions
20: S’ ← createSuccsAnticip()
ot

21: successors ← successors ∪ {S’}

22: motionPrev ← motion
23: return successors
tn

Output: successor-states S’ of state S

of the edge (2,3). In contrast, a successor of S that enters edge (3,4) at the beginning of
rin

its second SI is infeasible: its motion would require a duration of 110 s, but the SI on edge
(2,3) already ends at t = 100 s, thus requiring the motion to last less than 90 s.

The two successors created for the second SIs on edges (3,5) and (3,6) enter these edges
with a lower velocity than the speed limit of 5 m/s: to reach the required traversal time,
ep

the agent has to slow down significantly. In this case, arriving later and/or with a smaller
velocity at node (2) would yield a higher final velocity at node (3). Therefore, for the sake
of this example, we use the backtracking scheme (Lines 17 to 19) that was introduced in
Section 5.8 to optimize the respective final velocity for entering the two edges: the function
Pr

uses an acc-dec-motion explained in Section 5.2 from the parent state of S to satisfy the
restrictions. In the example, backtracking yields the two successors ID6 and ID7 that replace

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
iew
r ev
er
pe
Figure 7: Exemplary successor generation. From state S on edge (2,3), different successors
are created on edges (3,4), (3,5), (3,6), and (5,7) due to speed limits at straight segments
(blue lines, vlim = vmax ), curved segments (orange lines, vlim = vturn ), as well as safe
ot

intervals (SIs) on edge (2,3) (shaded in yellow) and for entering the upcoming edges (shaded
in green). All resulting successors S ′ are displayed.
tn

the successors ID4 and ID5, respectively. Note that the velocity limit may not be reachable
in all cases due to other constraints on the previous edges.
On the other hand, the final velocity of a regular successor may be too high to decelerate
rin

suﬀiciently for the velocity restrictions on upcoming edges. We use the anticipation scheme
introduced in Section 5.8 to check for such cases and to compute a motion from state S
directly to the start node of that future edge. In the example, this is illustrated with edge
(5,7): the length of edge (3,5) does not suﬀice for an agent to decelerate from 15 m/s to
the turn speed of 5 m/s. Therefore, the function implementing the anticipation scheme
ep

(Lines 20 to 21) creates a successor of state S (ID8) directly for entering edge (5,7).
Motion bounds do not have to be calculated when the agent has to wait at the current
node due to a wait-activity: the successor can directly be created based on the waiting
duration (Lines 1 to 4). If multiple SIs exist in which the agent can enter a subsequent
Pr

edge, additional successors for holding until each start of these SIs are created as well. Note
that holding is only possible until the end of the SI at the waiting-node.

ed
5.10 Replanning
When the AS-MAMP algorithm is invoked by the Routing Agent, some agents may already
be taxiing. Thus, these agents are likely in motion and in arbitrary places in the taxiway
network at the planning time tplng . Since states are always defined on nodes, the Localized

iew
Agents predict when and with which velocity these agents will arrive at the next node on
their respective paths (cp. their P3). The (node, time, velocity)-information is then used
as initial conditions when updating their plans in SIMP. To set the reservations correctly,
the graph reservations explained in Section 5.4 include the motion and edges along the
already executed part of the route. However, reservations that already ended before tplng
are dropped, since these lie in the past and agents cannot infringe them anymore. Thus,

ev
only those previous reservations that are still relevant during re-planning are accounted for.

6. Model Evaluation

r
In this section, we demonstrate the applicability of the AS-MAMP algorithm for airport
surface movement operations and evaluate its performance. To do so, we define four test
cases on the basis of potential bottlenecks in airport taxiway networks using a layout that

er
includes all important features of airport surfaces. Then, we explore the influence of the
search-mode in PBS (Section 6.2), of key operational parameters (Section 6.3), and the
planning window (Section 6.4). Finally, to stress-test the algorithm, we analyse the maximal
attainable runway throughput for both multi-engine taxiing (MET, cp. Section 2) and tug-
pe
enabled taxiing (TET), and put it into perspective with that of large European airports
(Section 6.5).

6.1 Definition of Test Cases

As visualized in Fig. 8, potential bottlenecks in taxiway networks of airports may appear
ot

in the following locations:

• runways due to the wake turbulence separation between two consecutive aircraft
tn

• bay areas due to pushback-operations and engine-start in case of multi-engine taxiing

• narrow sections in the taxiway network

• coupling / decoupling locations of tugs when aircraft use tug-enabled taxiing

rin

To analyse the bottlenecks separately, we define two schedules each with 7 flights in
total as listed in Table 3. In the scenario Queuing, the 7 aircraft depart from all available
gates in the layout (cp. Fig. 8). To isolate the runway as bottleneck, test case A considers
MET without pushback-operations and with engines being already started. To optimize the
ep

trajectory costs, the AS-MAMP algorithm must determine a suitable runway sequence so
that the WTC-separation (cp. Table 2) leads to minimal queuing delay before entering the
runway (cp. Fig. 9). For test case B, pushback-operations and engine-start are included,
congesting the bay area: decoupling from the pushback-truck takes 60 s, and engine-start
Pr

either 180 s (ICAO-C) or 360 s (ICAO-E). Thus, AS-MAMP must find a suitable pushback-
sequence. Note that agents can either hold at their gate, or on the bay area after pushback.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
R18/R36

runways

iew
r ev
decoupling
locations
er
narrow sections
G2
G4
G6
bay areas
G1
G3
G5
G7
pe
Figure 8: Taxiway layout from Fig. 2 with potential bottlenecks marked by red circles that
form the basis of the test cases.

case A case B case C case D

ot
tn
rin

Figure 9: Simulation snapshots illustrating the bottleneck of each test case

To investigate the bottleneck of tug-decoupling locations, test case C uses the same setup
ep

as case A but with tug-enabled taxiing. AS-MAMP has to choose one of the four decoupling
locations and must account for the decoupling duration of 120 s. In test case D, we use
the Crossflow scenario with three large arriving and four medium-sized departing aircraft,
without considering pushback-operations, to isolate the narrow section in the layout as
Pr

bottleneck. Therefore, the AS-MAMP algorithm must optimize the sequence in which the
departing and arriving aircraft pass through it.

ed
ID Type At From To ID Type At From To
1 ICAO-C T + 0s G1 R36 1 ICAO-E T + 0s R18 G2
2 ICAO-E T + 20 s G2 R36 2 ICAO-E T + 60 s R18 G4
3 ICAO-C T + 40 s G3 R36 3 ICAO-E T + 120 s R18 G6

iew
4 ICAO-E T + 60 s G4 R36 4 ICAO-C T + 0s G1 R36
5 ICAO-C T + 80 s G5 R36 5 ICAO-C T + 40 s G3 R36
6 ICAO-E T + 100 s G6 R36 6 ICAO-C T + 80 s G5 R36
7 ICAO-C T + 120 s G7 R36 7 ICAO-C T + 120 s G7 R36
(a) Queuing scenario (b) Crossflow scenario

ev
Table 3: Flight schedules used for the sensitivity analyses with T as fictive reference time

The default values for the main parameters such as the kinodynamic properties, cost
factors, safety factors, and positioning tolerance are provided in Table 4. The maximal

r
velocity for straight and curved segments resulted from expert interviews as well as current
technological limitations of tugs1 . To evaluate the capabilities of the AS-MAMP algorithm,

er
we chose rather low and differing absolute values for the acceleration and deceleration of the
agents, which have a large influence on the KPIs (see Section 6.3). Recall from Section 5.3
that the safety radii along an agent’s path and the neighbouring edges are defined as factors
in relation to the shape-radius (cp. Table 1) of the leading aircraft. The default values
pe
provided in Table 4 resulted from expert interviews.

parameter aircraft tugs unit

maximal velocity vmax 15 11.8 m/s
maximal turn velocity vturn,max 5 5 m/s
ot

minimal velocity vmin 1.5 1.5 m/s

acceleration acc 0.25 0.25 m/s2
deceleration dec -0.75 -0.75 m/s2
cost factor cd 0.1 0.1
tn

s/m
cost factor cprio 1 1 –
safety factor for path edges σpath 3 3 –
safety factor for neighbouring edges σN E 1 1 –
positioning tolerance δpos 0 0 m
rin

Table 4: Default values for the key parameters of AS-MAMP

As key performance indicators (KPIs), we use the sum of the taxi times ttaxi and taxi
ep

distances dtaxi of all flights. However, to show the relative increase in taxi time or distance
between different parameter settings, the KPIs are normalized by the lowest value of the
tested set. Furthermore, we track the number of collisions between vehicles during the
execution of the planned routes, and ensure that none exist. In the following, we analyse
the influence of key algorithmic and operational parameters on these KPIs.
Pr

1. vmax = 11.8 m/s is the current limit for the reference TaxiBot system (2013)

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
6.2 Influence of PBS Search-Modes
As outlined in Section 5.1, the AS-MAMP algorithm can be run with different search-modes
without and with induced constraints (IC). Table 5 provides a comparison of these when
applied to the four test cases. First, we analyse the results when running PBS without IC,

iew
followed by searching with IC.

search-mode Case cost runtime prio-pairs PBS nodes SIMP states

A 4710.8 ≈160 s 9 665 241 930
B 7513.2 ≈ 21 s 13 159 27 515
best-first

ev
C 6074.9 ≈450 s 13 1356 664 924
D 5099.0 ≈ 94 s 13 412 163 044
A 101 % ≈ 3s 12 25 6 049
B 100 % ≈ 3s 13 27 4 291
depth-first

r
C 108 % ≈ 6s 15 31 12 354
D 106 % ≈ 4s 14 24 8 751

greedy
A
B
C
D
100 %
138 %
104 %
115 %
≈
≈
≈
≈
er
3s
2s
2s
3s
9
6
6
11
19
13
13
20
4 980
3 468
3 010
5 709
pe
A 4710.8 ≈190 s 10 731 283 203
B 7513.2 ≈ 28 s 13 190 31 184
best-first (IC)
C 6074.9 ≈390 s 12 1138 550 543
D 5149.8 ≈100 s 10 464 160 096
A 100 % ≈ 3s 13 27 6 694
ot

B 100 % ≈ 3s 13 27 4 466
depth-first (IC)
C 110 % ≈ 13 s 19 39 16 500
D 108 % ≈ 11 s 13 53 19 440
tn

A 108 % ≈ 3s 9 19 4 494
B 138 % ≈ 2s 6 13 3 468
greedy (IC)
C 104 % ≈ 2s 6 13 3 010
D 100 % ≈ 3s 10 18 4 660
rin

Table 5: Results per test case and search-mode without or with induced constraints (IC).
For the depth-first and greedy searches, the cost ratios w.r.t. ”best-first” are listed.

As expected, while the best-first search yields the highest solution quality, it requires a
ep

large number of explored PBS nodes, total sum of SIMP states as well as a long runtime.
Substantially less time is required for the depth-first and greedy searches to find a solution,
with only a fraction of PBS nodes being explored. Since the runtime correlates with the
total sum of SIMP states, the greedy search terminates faster (cases C and D) or slightly
Pr

faster (cases A and B). Furthermore, due to picking the child-node with the lowest number
of remaining conflict pairs, it finds solutions requiring less priority relations. However, in

ed
comparison to the solution cost of the best-first search, its solution quality varies more
between the cases than with a depth-first search.
When running PBS with IC, recall that conflict detection returns the conflict pair with
the highest number of induced constraints, and likely yields a different pair than running

iew
PBS without IC. Despite returning the optimal priority ordering among those explored,
also the best-first search does not explore all possible priority pair sequences. Therefore,
when using IC, the number of PBS nodes, total sum of SIMP states, and runtime change
in general. Furthermore, the conflict free solutions contain a different sequence of priority
pairs, and the final count of pairs may differ. While the best-first search (IC) returns
identical solution costs for cases A, B, and C, that of test case D slightly increases. When

ev
comparing the resulting priority sequences of the two best-first searches of case D, they
indeed differ. Furthermore, ”best-first (IC)” never explores the final one of ”best-first”
that led to a slightly better solution. Likewise, using induced constraints in the other
two search-modes results in different solution costs. For test cases C and D, the solution
quality slightly degrades using ”depth-first (IC)”, while it improves using IC in the greedy

r
search. Nonetheless, the returned solution quality of the greedy search continues to fluctuate
more than that of the depth-first search. Regarding runtime, using IC in the depth-first

er
search increases the number of explored PBS nodes and SIMP states, leading to a higher
computational time. In contrast, using IC in the greedy search seems to slightly improve
its consistently low runtime. Overall, the comparison among the different search-modes
without and with IC suggests that the conflict selection procedure can be further optimized
pe
to combine the advantages of a fast greedy search with a more consistent solution quality
returned by a depth-first search.

6.3 Influence of Key Operational Parameters

In this subsection we analyse how the acceleration and deceleration, minimal velocity, cost-
factor cd in the cost function in SIMP, and positioning tolerance δpos influence the KPIs.
To do so, we deploy a best-first search without IC to yield the highest solution quality.
The choice of acceleration and deceleration has a strong impact on the taxiing duration,
tn

as shown exemplary for case A in Fig. 10: the higher the acceleration and/or deceleration,
the faster the aircraft are able to complete their taxiing. In contrast, the influence on the
travelled distance is minimal. This trend holds for the other cases as well. Thus, with
the values of acc = 0.25 m/s2 and dec = −0.75 m/s2 , approximately 10 % higher taxi times
rin

result in comparison to the highest considered absolute value of 2 m/s2 for both parameters.
Since the maximal velocities are operational and/or technological limits, we did not
vary them. Nonetheless, it is self-evident that changing these values will affect the taxiing
duration. When varying the minimal velocity vmin between 0.5 m/s and 5.0 m/s, the taxi
time increased by more than 7 % for vmin ≥ 2.0 m/s in test case D. While case C showed
ep

a small increase up to 1 % at vmin = 5.0 m/s, test cases A and B were unaffected by the
different minimal velocities. This suggests that the AS-MAMP algorithm is capable of
creating trajectories without the need for agents to stop at arbitrary locations to resolve
conflicts, as long as vmin ≤ 1.5 m/s. The operational benefit is that pilots have to apply
Pr

breakaway-thrust less often, resulting in fuel savings and reduced engine-blasts that lower
the associated risks.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
iew
r ev
er
pe
Figure 10: Sensitivity analysis of case A depicting how acceleration acc and deceleration
dec influence the taxi time and distance, shown as increase in [%] w.r.t. the lowest sum
over all flights.
ot

Regarding the cost function in SIMP (cp. Eq. (2)) our hypothesis was that small values
of cd have a positive impact on the KPIs: while for cd = 0 the algorithm may choose long
detours to save a few seconds in taxi time, slightly including the travelled distance in the
tn

objective function should mitigate these. In contrast, high values for cd should result in
shorter paths but increasingly long taxi times, since the secondary objective of optimizing
for distance may outweigh the primary one of low taxi times. When varying cd between
0 and 0.5 s/m, the hypothesis is supported: setting cd = 0, the taxi distance increases by
rin

1.7 % for test case A and 3.1 % for test case C, while cases B and D are unaffected. For
cd = 0.5, especially case D is affected: both taxi time and taxi distance increased by 3.9 %
and 0.8 %, respectively. For cases A and C, the taxi distances were unaffected, and the taxi
times showed a small increase of 1.7 % (case A) and 3.1 % (case C).
Increasing the positioning tolerance δpos of each agent from 0 m to 50 m, showed a
ep

negligible influence on test cases A and B. In contrast, the taxi time abruptly increased to
> 6 % for δpos >= 20 m in test case C, and > 4 % for δpos >= 10 m in test case D. In both
cases, the taxi time further increased with rising positioning tolerances adding around 2 % at
δpos = 50 m, while the increase in taxi distance remained < 0.8 %. This suggests that higher
Pr

positioning tolerances likely impact confined passages in airport layouts. Nonetheless, the
overall impact on the KPIs should be reasonably small for δpos < 20 m.

ed
6.4 Influence of Planning Window
Up to this point, we have considered flight schedules that consist of a limited amount of
flights and can be routed within a single planning round. To assess how the AS-MAMP
algorithm deals with continuous schedules and thus the necessity to re-plan, we extended

iew
the Queuing schedule in Table 3a to over 90 minutes by repeating the seven flights: for
each replicate, the parameter tcycle is added to the fictive reference time T . Note that we
only use case A and case C in the analysis below, as the bay areas (case B) and the narrow
section (case D) would result in unrealistic bottleneck-situations for continuous schedules
on the layout from Fig. 8. Furthermore, as the number of agents exceeded the feasibility
of a best-first search, we used a best-of greedy search as slight modification in PBS: after

ev
finding the first solution, we store it, let PBS pick the lowest-cost search-node, and continue
greedily searching for an alternative solution. This procedure is repeated until a maximum
number of solutions (in this case 24) is found, returning the one with the lowest cost.
We varied the planning window wplng and replanning period hplng from which the set

r
with hplng = 50 % ∗ wplng of case A is depicted in Fig. 11. As expected, for both test
cases, the runtime as sum over all planning-rounds is lower for smaller planning windows:
AS-MAMP has to assign less priority-relations since conflicts occurring after the planning

er
window are ignored. For test case A, the average taxi time remained unaffected, neglecting
the outlier in Fig. 11. Likewise, for test case C, the average taxi time did not change
significantly for all considered values of wplng and hplng when tcycle ≤ 10 min. With tcycle >
pe
10 min, the simulations showed stronger taxi time variations across varying wplng and hplng
while increasing in general for a lower value of tcycle . This is further discussed in the next
subsection.
ot
tn
rin
ep

Figure 11: Sensitivity analysis depicting how the planning window wplng influences the
maximal runway throughput per 15min and 60min frame for the scenario of case A and a
Pr

relative replanning period hplng = 50 % ∗ wplng .

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
6.5 Stress-test: Potential Runway Throughput
In this part, we focus on the potential maximal runway throughput to demonstrate the
applicability of AS-MAMP for busy airports. To do so, we fix wplng = 30 min and hplng =
15 min, vary tcycle from 15 min to 7.5 min, and compare multi-engine taxiing (MET, test

iew
case A) to tug-enabled taxiing (TET, test case C). Fig. 12 shows the average taxi times,
average taxi distances as well as the maximal runway throughput and occupancy rate for
increasing traﬀic loads (i.e. lower tcycle ) for both test cases.
In test case A, the taxi distance remains unaffected across the tested traﬀic loads, while
the taxi time increases only for the most congested case leading to an average delay of
around 2.5 min. However, this also maxes out the achievable runway throughput with 56

ev
flights within 60 min as well as the occupancy rate reaching its limit of around 100 % within
both a 15 min and 60 min frame.
For TET, both taxi time and taxi distance increase for tcycle ≥ 10 min. At tcycle = 8 min,
the KPIs reach their respective maximal values: the average delay rises to around 5 min,

r
while the runway throughput reaches 47 flights per hour equal to an occupancy rate of
around 94 %. The occupancy rate for a 15 min frame shows that its highest value of around

er
98 % is approached for tcycle ≥ 10 min, matching the increase in taxi time. For lower
traﬀic loads, the TET case consistently shows higher runway occupancy rates, while having
identical runway throughput compared to the MET case. This means that the AS-MAMP
algorithm cannot optimize the runway sequence of TET as much as in the MET case.
pe
This is also reflected in the most congested TET scenario with tcycle = 7.5 min which the
simulation failed to complete. As visualized in Fig. 13, aircraft queue almost until the bay
area when the simulation crashes, around one hour into the 90 min flight schedule. This is
also due to the choice of simulation parameters: with a minimal velocity of vmin = 1.5 m/s,
the agents eventually run out of options, and AS-MAMP cannot find a solution anymore.
ot

In contrast, without the necessity to decouple, minimal queuing occurs in test case A.
As mentioned above, the chosen set of parameters results in a maximal runway through-
put within 60 min of 56 takeoffs and 47 takeoffs for MET and TET, respectively. To put this
into perspective, Table 6 lists the maximal runway throughput of all possible runway config-
tn

urations for the five largest airports in Europe (Eurocontrol, 2020): the highest throughput
per runway is reached by London Heathrow with 54 departing flights per hour. Therefore,
when aircraft use MET, the multi-agent system is capable of matching or exceeding the
maximal runway throughput seen at these large European airports. For TET, the runway
rin

throughput in the simulated setting still exceeds those of most European airports. Note
that the potential throughput can be increased when the taxiway network is adapted: with
either better accessible or more decoupling locations, the AS-MAMP algorithm would have
more options to optimize the takeoff sequence and thus the potential runway capacity.
Regarding runway occupancy rates, our previous studies on the operational consequences
ep

of automated MET and TET operations (von der Burg & Sharpanskykh, 2023, 2024) re-
sulted in similar rates, also with respect to the historic ones. Based on the flight schedules
of two of the busiest days at Amsterdam Airport Schiphol (AMS), the maximal hourly
occupancy rate did not exceed 79 % regardless of the operations and runway configuration.
Pr

In turn, we believe that this validates that the two cases analysed above are suitable stress-
tests for AS-MAMP: all scenarios with tcycle ≥ 10 min resulted in higher occupancy rates

ed
iew
r ev
er
Figure 12: KPIs w.r.t. tcycle for continuous schedules of test cases A and C
pe
(a) multi-engine taxiing (MET), test case A
ot
tn

(b) tug-enabled taxiing (TET), test case C

rin
ep
Pr

Figure 13: Simulation snapshots at ≈ 60 min into continuous schedule with tcycle = 7.5 min

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
EDDF EHAM LFPG EGLL LEMD case A case C
FRA AMS CDG LHR MAD MET TET
departure runways 2 1 2 1 2 1 1
max. throughput 68 46 75 54 70 56 47

iew
max. occupancy n/a n/a n/a n/a n/a 100 % 94 %

Table 6: Comparison of maximal runway throughput and occupancy rate for departures
within 60 min between the 5 largest European airports and the simulated continuous cases
A and C

ev
than those found in the historic flight schedules. This underscores that the AS-MAMP
algorithm is able to coordinate ASM Ops in an eﬀicient way, and only returns planned
trajectories when these are indeed conflict-free.

r
7. Future Work

er
Developing realistic models for real-world applications such as ASM Ops realistically is a
challenging and labour-intensive task. Furthermore, as MAPF-related research is a very
active field, many new algorithms and extensions are continuously published. However,
to use these algorithms for such realistic applications, they usually need to be extended,
pe
refined, and tailored to specific operational details and conditions, which is often not a
trivial task as well. Therefore, we deem the comparison to recent alternative algorithms,
e.g. to MAPF-LNS2 (Li et al., 2022) or (Li, Ruml, & Koenig, 2021), to be out of scope
of this paper, and leave it for future work. These high-level algorithms will likely require
changes to the SIMP algorithm as well that need to be carefully tested.
ot

Regarding the AS-MAMP algorithm, the presented comparison between the depth-first
and greedy search-modes in its high-level solver did not result in a clear winner. Thus,
we plan to investigate further options to build a priority ordering that quickly returns a
conflict-free solution close to the one returned by the best-first search. When assessing the
tn

impact of these search strategies on the trade off between solution quality and computational
time, we plan to use the layout and flight schedules from our previous studies as well. Low
runtimes of the AS-MAMP algorithm are important to eventually allow for human-in-
the-loop experiments. Thus, to speed up SIMP, we want to explore improvements to the
rin

heuristics as well as pruning strategies to lower the number of explored states.

On the operational side, we deem important to study how local trajectory adjustments
during plan execution that exceed the positioning tolerances of an agent can be implemented
in a robust way: the initial conditions for replanning in AS-MAMP must be carefully
designed so that SIMP returns a valid path whenever one is operationally feasible. Another
ep

direction for future research is to use the simulator for further operational studies, e.g. to
investigate the impact of current ATC procedures such as standard taxiway directions on
the performance indicators of taxi time, taxi distance, and algorithmic runtime. Embedding
such procedures in an eﬀicient way will likely entail changes to the cost definition, heuristics,
Pr

and search in SIMP. Additionally, the MAS model can be further expanded to integrate
other aspects of ASM Ops like the task assignment of tugs as well as routing tugs directly

ed
to the location of the new task. However, the latter will require adjustments to the airport
layout as the land-sides of stands are currently not connected to service roads. Other
research focuses on analysing ADS-B data on service roads that could be helpful to yield
more realistic modelling of the tug movements between task assignments.

iew
8. Conclusions
In this paper, we proposed a hybrid multi-agent system model to facilitate autonomous
airport surface movement operations, while sustaining safety levels. In the MAS, vehicle
paths are planned centrally by the Routing Agent. Based on the resulting routes, the

ev
distributed Localized Agents instruct the Moving Agents which then execute the commands
accordingly.
To deconflict all concurrent routes, we developed the Multi-Agent Motion Planning
on Airport Surfaces (AS-MAMP) algorithm. This two-level solver plans with priorities
based on the Priority-Based Search (PBS) algorithm in its high-level search. We combined

r
this with the new Safe Interval Motion Planning (SIMP) algorithm as low-level solver that
accounts for the kinodynamic properties and two-dimensional, circular shapes of the taxiing

er
vehicles. As such, it further generalizes the ideas from SIPP with Interval Projection (SIPP-
IP) to deal with agents that cannot stop instantaneously, and who have an initial velocity.
Furthermore, to enable collision-free queuing, SIMP detects and resolves any overlaps of
the shapes on edges in the travel direction of the Moving Agents. The routing algorithm
pe
thus yields conflict-free paths with the maintained safety distances at real-world airports.
Moreover, SIMP uses a novel activity-based search to account for the various operations
such as pushback and coupling/decoupling of tugs along the route. With the three activities
“go-to”, “follow”, and “wait”, all relevant operations for both inbound and outbound flights
can be expressed. Since new aircraft continuously land and depart from airports, the AS-
ot

MAMP algorithm uses a rolling-horizon scheme with a planning window and replanning
period.
We simulated four different cases corresponding to the potential bottlenecks in taxiway
networks. In the utilized layout, all important features of an airport surface were included.
tn

Using sensitivity analyses, we showed the influence of the main algorithmic and operational
parameters such as different acceleration and deceleration, minimal velocities, positioning
tolerances of agents, as well as the planning window and replanning period. Furthermore,
we evaluated the influence of six search-modes in PBS on the trade-off between solution
rin

quality and runtime.

We demonstrated that the multi-agent system model coordinated the ASM Ops safely
and eﬀiciently. Safe operations were ensured directly by the MAS: it tracks the number
of collisions during execution of the planned routes, and verifies that none exist. Eﬀicient
operations were shown by analysing the maximal runway throughput for both multi-engine
ep

taxiing (MET) and tug-enabled taxiing (TET): on the exemplary layout used, the MAS
achieved a runway throughput of 56 and 47 aircraft per hour for MET and TET, respec-
tively. These results match or exceed the maximal hourly runway throughput at the largest
European airports. This underlines the usability of the hierarchical MAS model with the
Pr

embedded AS-MAMP algorithm to study airport surface movement operations in detail,

and highlights its potential for future autonomous operations.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Acknowledgments

This work has received funding from the SESAR Joint Undertaking under grant agreement
No 892869 under European Union’s Horizon 2020 research and innovation programme.

iew
References
Ali, Z. A., & Yakovlev, K. (2023). Safe Interval Path Planning with Kinodynamic Con-
straints. Proceedings of the AAAI Conference on Artificial Intelligence, 37 (10),
12330–12337. https://doi.org/10.1609/aaai.v37i10.26453

ev
Andreychuk, A., Yakovlev, K. S., Atzmon, D., & Stern, R. (2019). Multi-agent pathfinding
(MAPF) with continuous time. Proceedings of IJCAI-19, 39–45.
Andreychuk, A., Yakovlev, K. S., Boyarski, E., & Stern, R. (2021). Improving Continuous-
time Conflict Based Search. Proceedings of AAAI-21, 11220–11227.
Ashok, A., Balakrishnan, H., & Barrett, S. R. H. (2017). Reducing the air quality and CO2

r
climate impacts of taxi and takeoff operations at airports. Transportation Research
Part D: Transport and Environment, 54, 287–303. https://doi.org/10.1016/j.trd.
2017.05.013
er
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, 8.
pe
Atzmon, D., Stern, R., Felner, A., Wagner, G., Barták, R., & Zhou, N.-F. (2020). Robust
multi-agent path finding and executing. Journal of Artificial Intelligence Research,
67, 549–579. https://doi.org/10.1613/jair.1.11734
Bakowski, D. L., Hooey, B. L., & Foyle, D. C. (2017). Flight deck surface trajectory-
based operations (STBO): A four-dimensional trajectory (4DT) simulation. 2017
IEEE/AIAA 36th Digital Avionics Systems Conference (DASC), 1–10. https://doi.
ot

org/10.1109/DASC.2017.8102008
Bakowski, D. L., Hooey, B. L., Foyle, D. C., & Wolter, C. A. (2015). NextGen Sur-
face Trajectory-Based Operations (STBO): Evaluating Conformance to a Four-
tn

dimensional Trajectory (4DT). Procedia Manufacturing, 3, 2458–2465. https : / /

doi.org/10.1016/j.promfg.2015.07.506
Balakrishnan, H., & Jung, Y. (2007). A Framework for Coordinated Surface Operations
Planning at Dallas-Fort Worth International Airport. AIAA Guidance, Navigation
and Control Conference and Exhibit. https://doi.org/10.2514/6.2007-6553
rin

Barer, M., Sharon, G., Stern, R., & Felner, A. (2014). Suboptimal variants of the Conflict-
Based Search algorithm for the multi-agent pathfinding problem. Proceedings of
SoCS-14, 19–27.
Bennell, J. A., Mesgarpour, M., & Potts, C. N. (2011). Airport runway scheduling. 4OR,
ep

9(2), 115–138. https://doi.org/10.1007/s10288-011-0172-x

Bernatzky, T. (2019). Automation Concept for Cockpit Crew Integration into Trajectory-
Based Dispatch Towing [Doctoral dissertation, TU Darmstadt].
Brownlee, A. E. I., Weiszer, M., Chen, J., Ravizza, S., Woodward, J. R., & Burke, E. K.
(2018). A fuzzy approach to addressing uncertainty in Airport Ground Movement
Pr

ed
optimisation. Transportation Research Part C: Emerging Technologies, 92, 150–175.
https://doi.org/10.1016/j.trc.2018.04.020
Chan, S.-H., Stern, R., Felner, A., & Koenig, S. (2023). Greedy Priority-Based Search for
Suboptimal Multi-Agent Path Finding. Proceedings of the International Symposium

iew
on Combinatorial Search, 16(1), 11–19. https://doi.org/10.1609/socs.v16i1.27278
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, 1–8. https://doi.org/10.1145/2950112.2964589
Cohen, L., Uras, T., Kumar, T. K. S., & Koenig, S. (2019). Optimal and bounded-suboptimal
multi-agent motion planning. Proceedings of SoCS-19, 8.

ev
Cohen, L., Uras, T., Kumar, T. K., Xu, H., Ayanian, N., & Koenig, S. (2016). Improved
solvers for bounded-suboptimal multi-agent path finding. In IJCAI, 3067–3074.
Dhief, I., Wang, Z., Liang, M., Alam, S., Schultz, M., & Delahaye, D. (2020). Predicting
Aircraft Landing Time in Extended-TMA Using Machine Learning Methods.
Eurocontrol. (2016). A-CDM impact assessment: Final report. Eurocontrol.

r
Eurocontrol. (2018). European aviation in 2040 - challenges of growth. Eurocontrol Statis-
ticsForecast Service. Retrieved December 14, 2021, from https://www.eurocontrol.

er
int/sites/default/files/2019-07/challenges-of-growth-2018-annex1_0.pdf
Eurocontrol. (2020). Airport capacity imbalance. Eurocontrol, Performance Review Com-
mission. Retrieved May 19, 2023, from https://www.eurocontrol.int/sites/default/
files/2020-12/eurocontrol-prc-technical-note-airport-capacity-imbalance-11122020.
pe
pdf
Eurocontrol. (2021). Data snapshot #22 on lower summer taxi-out times. Retrieved Novem-
ber 30, 2021, from https : / / www . eurocontrol . int / sites / default / files / 2021 - 11 /
eurocontrol-data-snapshot-22-taxi-time-affect-ops.pdf
Eurocontrol. (2023). Data snapshot #40 on taxi times. Retrieved April 4, 2023, from https:
ot

//www.eurocontrol.int/sites/default/files/2023- 03/eurocontrol- data- snapshot-

40.pdf
Felner, A., Stern, R., Shimony, S. E., Boyarski, E., Goldenerg, M., Sharon, G., Sturtevant,
N. R., Wagner, G., & Surynek, P. (2017). Search-based optimal solvers for the multi-
tn

agent pathfinding problem: Summary and challenges. Proceedings of SoCS-17, 29–

37.
Gao, J., Li, Y., Li, X., Yan, K., Lin, K., & Wu, X. (2024). A review of graph-based multi-
agent pathfinding solvers: From classical to beyond classical. Knowledge-Based Sys-
rin

tems, 283, 111121. https://doi.org/10.1016/j.knosys.2023.111121

Gerdes, I., & Schaper, M. (2015). Eleventh USA/Europe Air Traﬀic Management Research
and Development Seminar (ATM2015): Management of Time Based Taxi Trajecto-
ries coupling Departure and Surface Management Systems.
Goldenberg, M., Felner, A., Stern, R., Sharon, G., Sturtevant, N., Holte, R. C., & Scha-
ep

effer, J. (2014). Enhanced Partial Expansion A*. Journal of Artificial Intelligence

Research, 50, 141–187. https://doi.org/10.1613/jair.4171
González, D., Pérez, J., Milanés, V., & Nashashibi, F. (2016). A Review of Motion Planning
Techniques for Automated Vehicles. IEEE Transactions on Intelligent Transporta-
tion Systems, 17 (4), 1135–1145. https://doi.org/10.1109/TITS.2015.2498841
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Guépet, J., Briant, O., Gayon, J., & Acuna-Agost, R. (2016). The aircraft ground routing
problem: Analysis of industry punctuality indicators in a sustainable perspective.
European Journal of Operational Research, 248(3), 827–839. https://doi.org/10.
1016/j.ejor.2015.08.041

iew
Guépet, J., Briant, O., Gayon, J.-P., & Acuna-Agost, R. (2017). Integration of aircraft
ground movements and runway operations. Transportation Research Part E: Logis-
tics and Transportation Review, 104, 131–149. https://doi.org/10.1016/j.tre.2017.
05.002
Hayashi, M., Hoang, T., Jung, Y. C., Malik, W., Lee, H., & Dulchinos, V. L. (2015). Eval-
uation of Pushback Decision-Support Tool Concept for Charlotte Douglas Interna-

ev
tional Airport Ramp Operations. Retrieved June 18, 2024, from https://ntrs.nasa.
gov/citations/20160008912
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,
56.

r
Hoenig, W., Kumar, T. K. S., Cohen, L., Ma, H., Xu, H., Ayanian, N., & Koenig, S. (2016).
Multi-agent path finding with kinematic constraints. Proceedings of ICAPS-16, 9.

er
Hönig, W., Preiss, J. A., Kumar, T. K. S., Sukhatme, G. S., & Ayanian, N. (2018). Trajectory
Planning for Quadrotor Swarms. IEEE Transactions on Robotics, 34(4), 856–869.
https://doi.org/10.1109/TRO.2018.2853613
Hooey, B. L., Cheng, V. H. L., & Foyle, D. C. (2014). A Concept of Operations for Far-Term
pe
Surface Trajectory-Based Operations (STBO) (NASA/TM-2014-218354). Retrieved
May 13, 2024, from https://ntrs.nasa.gov/citations/20190002687
IATA. (2021). 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/
ot

ICAO. (2004). Advanced Surface Guidance and Control Systems (A-SMGCS) Manual.
ICAO. (2016). Aerodrome design and operations. In Annex 14 to the convention on inter-
national civil aviation (7th ed.).
Jiang, Y., Liu, Z., Hu, Z., & Zhang, H. (2021). A Priority-Based Conflict Resolution Strategy
tn

for Airport Surface Traﬀic Considering Suboptimal Alternative Paths. IEEE Access,
9, 606–617. https://doi.org/10.1109/ACCESS.2020.3045586
Jung, Y., Hoang, T., Montoya, J., Gupta, G., Malik, W. A., Tobias, L., & Wang, H.
(2011). Performance Evaluation of a Surface Traﬀic Management Tool for Dallas
rin

/ Fort Worth International Airport. Retrieved June 18, 2024, from https://www.
semanticscholar . org / paper / Performance - Evaluation - of - a - Surface - Traffic - Tool -
Jung-Hoang/e95cc72f4286ee1d748204a50039fecc4ed7ac55
Kottinger, J., Almagor, S., & Lahijanian, M. (2022). Conflict-Based Search for Multi-
Robot Motion Planning with Kinodynamic Constraints. 2022 IEEE/RSJ Interna-
ep

tional Conference on Intelligent Robots and Systems (IROS), 13494–13499. https:

//doi.org/10.1109/IROS47612.2022.9982018
Lane, R., Dubuisson, S., Ellejmi, M., & Cerasi, E. (2020). EUROCONTROL specification
for A-SMGCS services.
Pr

ed
Lee, H. (2014). Airport surface traﬀic optimization and simulation in the presence of un-
certainties [Thesis]. Massachusetts Institute of Technology. Retrieved May 6, 2024,
from https://dspace.mit.edu/handle/1721.1/87480
Li, J., Chen, Z., Harabor, D., Stuckey, P. J., & Koenig, S. (2022). MAPF-LNS2: Fast

iew
Repairing for Multi-Agent Path Finding via Large Neighborhood Search. Proceed-
ings of the AAAI Conference on Artificial Intelligence, 36(9), 10256–10265. https:
//doi.org/10.1609/aaai.v36i9.21266
Li, J., Gange, G., Harabor, D., Stuckey, P. J., Ma, H., & Koenig, S. (2020). New Tech-
niques for Pairwise Symmetry Breaking in Multi-Agent Path Finding. Proceedings
of ICAPS-20, 30, 193–201. https://doi.org/10.1609/icaps.v30i1.6661

ev
Li, J., Gong, M., Liang, Z., Liu, W., Tong, Z., Yi, L., Morris, R., Pasearanu, C., & Koenig,
S. (2019). Scheduling and Airport Taxiway Path Planning under Uncertainty. In
AIAA Aviation 2019 Forum. American Institute of Aeronautics and Astronautics.
https://doi.org/10.2514/6.2019-2930
Li, J., Hoang, T. A., Lin, E., Vu, H. L., & Koenig, S. (2023). Intersection Coordination

r
with Priority-Based Search for Autonomous Vehicles. Proceedings of AAAI-23, 37,
11578–11585. https://doi.org/10.1609/aaai.v37i10.26368

er
Li, J., Ruml, W., & Koenig, S. (2021). EECBS: A Bounded-Suboptimal Search for Multi-
Agent Path Finding. Proceedings of the AAAI Conference on Artificial Intelligence,
35(14), 12353–12362. https://doi.org/10.1609/aaai.v35i14.17466
Li, J., Surynek, P., Felner, A., Ma, H., Kumar, T., & Koenig, S. (2019). Multi-agent path
pe
finding for large agents. Proceedings of AAAI-19, 33, 7627–7634. https://doi.org/
10.1609/aaai.v33i01.33017627
Li, J., Tinka, A., Kiesel, S., Durham, J. W., Kumar, T. K. S., & Koenig, S. (2021). Lifelong
multi-agent path finding in large-scale warehouses. Proceedings of AAAI-21.
Lieder, A., Briskorn, D., & Stolletz, R. (2015). A dynamic programming approach for the
ot

aircraft landing problem with aircraft classes. European Journal of Operational Re-
search, 243(1), 61–69. https://doi.org/10.1016/j.ejor.2014.11.027
Lukic, M., Giangrande, P., Hebala, A., Nuzzo, S., & Galea, M. (2019). Review, Challenges,
and Future Developments of Electric Taxiing Systems. IEEE Transactions on Trans-
tn

portation Electrification, 5(4), 1441–1457. https : / / doi . org / 10 . 1109 / TTE . 2019 .
2956862
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, 7643–7650.
rin

Ma, H., Hönig, W., Kumar, T. K. S., Ayanian, N., & Koenig, S. (2019). Lifelong Path Plan-
ning with Kinematic Constraints for Multi-Agent Pickup and Delivery. Proceedings
of AAAI-19, 33, 7651–7658.
Ma, H., Koenig, S., Ayanian, N., Cohen, L., Honig, W., Kumar, T. K. S., Uras, T., Xu,
H., Tovey, C., & Sharon, G. (2016). Overview: Generalizations of multi-agent path
ep

finding to real-world scenarios. In IJCAI-16 Workshop on Multi-Agent Path Finding

(WoMAPF).
Ma, H., Kumar, T. K. S., & Koenig, S. (2017). Multi-agent path finding with delay proba-
bilities. Proceedings of AAAI-17, 3605–3612.
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
Ma, H., Wagner, G., Felner, A., Li, J., Kumar, T. K. S., & Koenig, S. (2018). Multi-agent
path finding with deadlines. Proceedings of IJCAI-18, 417–423. https://doi.org/10.
24963/ijcai.2018/58
Morris, R., Pasareanu, C. S., Luckow, K., Malik, W., Ma, H., Kumar, T. K. S., & Koenig,

iew
S. (2016). Planning, Scheduling and Monitoring for Airport Surface Operations. In
AAAI-16 Workshop on Planning for Hybrid Systems.
Okuniek, N., Gerdes, I., Jakobi, J., Ludwig, T., Hooey, B. L., Foyle, D., Jung, Y. C., &
Zhu, Z. (2016). A Concept of Operations for Trajectory-based Taxi Operations.
16th AIAA Aviation Technology, Integration, and Operations Conference. https :
//doi.org/10.2514/6.2016-3753

ev
Okuniek, N., Zhu, Z., Jung, Y. C., Gridnev, S., Gerdes, I., & Lee, H. (2018). Performance
Evaluation of Conflict-Free Trajectory Taxiing in Airport Ramp Area Using Fast-
Time Simulations. 2018 IEEE/AIAA 37th Digital Avionics Systems Conference
(DASC), 1–10. https://doi.org/10.1109/DASC.2018.8569228
Phillips, M., & Likhachev, M. (2011). SIPP: Safe Interval Path Planning for dynamic en-

r
vironments. Proceedings of ICRA-11, 5628–5635. https://doi.org/10.1109/ICRA.
2011.5980306

er
Ravizza, S., Atkin, J. A. D., & Burke, E. K. (2014). A more realistic approach for airport
ground movement optimisation with stand holding. Journal of Scheduling, 17 (5),
507–520. https://doi.org/10.1007/s10951-013-0323-3
Rieth, M., & Hagemann, V. (2022). Automation as an equal team player for humans? – A
pe
view into the field and implications for research and practice. Applied Ergonomics,
98(4). https://doi.org/10.1016/j.apergo.2021.103552
Rooseleer, F., & Treve, V. (2018). RECAT-EU: European wake turbulence categorisation
and separation minima on approach and departure. Retrieved April 5, 2023, from
https : / / www . eurocontrol . int / sites / default / files / 2021 - 07 / recat - eu - released -
ot

september-2018.pdf
Rowland, B. (2022). Which Part of a Flight Uses the Most Fuel? Aviation Sustainability.
Retrieved April 5, 2023, from https://www.oag.com/blog/which-part-flight-uses-
most-fuel
tn

Semiz, F., & Polat, F. (2021). Incremental multi-agent path finding. Future Generation
Computer Systems, 116, 220–233. https://doi.org/10.1016/j.future.2020.09.032
Sharon, G., Stern, R., Felner, A., & Sturtevant, N. R. (2012). Conflict-Based Search for
optimal multi-agent path finding.
rin

Sharon, G., Stern, R., Goldenberg, M., & Felner, A. (2013). The increasing cost tree search
for optimal multi-agent pathfinding. Proceedings of IJCAI-13, 195, 470–495. https:
//doi.org/10.1016/j.artint.2012.11.006
Silver, D. (2005). Cooperative pathfinding. Proceedings of the Artificial Intelligence and
Interactive Digital Entertainment Conference (AIIDE-2005), 117–122.
ep

SKYbrary. (2024a). Accident and Serious Incident Reports: RI. Runway Incursion. Retrieved
June 23, 2024, from https://skybrary.aero/articles/accident-and-serious-incident-
reports-ri
SKYbrary. (2024b). Ground Collision. Ground Operations. Retrieved June 23, 2024, from
https://skybrary.aero/articles/ground-collision
Pr

ed
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. (2019). Multi-agent path finding – an overview. In G. S. Osipov, A. I. Panov,

iew
& K. S. Yakovlev (Eds.), Artificial Intelligence: 5th RAAI Summer School, Dolgo-
prudny, Russia, July 4–7, 2019, Tutorial Lectures (pp. 96–115). Springer Interna-
tional Publishing. https://doi.org/10.1007/978-3-030-33274-7_6
Stern, R., & Sturtevant, N. R. (2019). Multi-agent pathfinding: Definitions, variants, and
benchmarks. Proceedings of SoCS-19, 151–158.
Švancara, J., Vlk, M., Stern, R., Atzmon, D., & Barták, R. (2019). Online multi-agent

ev
pathfinding. Proceedings of AAAI-19, 33(01), 7732–7739. https://doi.org/10.1609/
aaai.v33i01.33017732
TaxiBot: The world’s only certified and operational Taxiing alternative. (2013). TaxiBot.
Retrieved May 10, 2023, from https://www.taxibot-international.com
Thayer, J., Dionne, A., & Ruml, W. (2011). Learning Inadmissible Heuristics During Search.

r
Proceedings of ICAPS-11, 21, 250–257. https://doi.org/10.1609/icaps.v21i1.13474
Thayer, J. T., & Ruml, W. (2011). Bounded Suboptimal Search: A Direct Approach Using

er
Inadmissible Estimates. Proceedings of IJCAI-11.
Udluft, H. (2017). Decentralization in Air Transportation [Doctoral dissertation, Depart-
ment of Control and Operations, Aerospace Engineering, TU Delft].
Paris Agreement (2015). Retrieved May 17, 2023, from https://unfccc.int/sites/default/
pe
files/english_paris_agreement.pdf
von der Burg, M., & Sharpanskykh, A. (2023). Multi-Agent Planning for Autonomous
Airport Surface Movement Operations. SESAR Innovation Days 2023. https://doi.
org/10.61009/SID.2023.1.27
von der Burg, M., & Sharpanskykh, A. (2024). Studying the Operational Consequences of
ot

Automated Engine-Off Taxiing using Multi-Agent Planning. ICRAT-24 (Accepted

for Presentation).
Wagner, G., & Choset, H. (2011). M*: A Complete Multirobot Path Planning Algorithm
with Performance Bounds. Proceedings of IROS-11.
tn

Walker, T. T. (2022). Multi-Agent Pathfinding in Mixed Discrete-Continuous Time and

Space [Doctoral dissertation, University of Denver].
Walker, T. T., & Sturtevant, N. R. (2019). Collision Detection for Agents in Multi-Agent
Pathfinding. arXiv: 1908.09707 [cs, eess]. Retrieved December 4, 2023, from http:
rin

//arxiv.org/abs/1908.09707
Walker, T. T., Sturtevant, N. R., & Felner, A. (2018). Extended Increasing Cost Tree Search
for Non-Unit Cost Domains. Proceedings of the Twenty-Seventh International Joint
Conference on Artificial Intelligence, 534–540. https://doi.org/10.24963/ijcai.2018/
74
ep

Walker, T. T., Sturtevant, N. R., & Felner, A. (2020). Generalized and Sub-Optimal Bipar-
tite Constraints for Conflict-Based Search. Proceedings of the AAAI Conference on
Artificial Intelligence, 34(05), 7277–7284. https://doi.org/10.1609/aaai.v34i05.6219
Walker, T. T., Sturtevant, N. R., & Felner, A. (2024). Clique Analysis and Bypassing in
Continuous-Time Conflict-Based Search. Proceedings of the International Sympo-
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
sium on Combinatorial Search, 17, 152–160. https://doi.org/10.1609/socs.v17i1.
31553
Wang, X., Brownlee, A. E. I., Weiszer, M., Woodward, J. R., Mahfouf, M., & Chen, J.
(2021). A chance-constrained programming model for airport ground movement op-

iew
timisation with taxi time uncertainties. Transportation Research Part C: Emerging
Technologies, 132, 103382. https://doi.org/10.1016/j.trc.2021.103382
Wang, Z., Liang, M., & Delahaye, D. (2018). Automated data-driven prediction on aircraft
Estimated Time of Arrival. 88, 101840. https://doi.org/10.1016/j.jairtraman.2020.
101840
Woo, C. J., Goh, S. K., Alam, S., Ferdaus, M., & Ellejmi, M. (2022). A Runway Exit

ev
Prediction Model with Visually Explainable Machine Decisions.
Yakovlev, K., & Andreychuk, A. (2017). Any-Angle Pathfinding for Multiple Agents Based
on SIPP Algorithm. Proceedings of ICAPS-17, 27, 586–594. Retrieved December 23,
2021, from https://ojs.aaai.org/index.php/ICAPS/article/view/13856
Yakovlev, K., & Andreychuk, A. (2021). Towards Time-Optimal Any-Angle Path Planning

r
With Dynamic Obstacles. Proceedings of the International Conference on Automated
Planning and Scheduling, 31, 405–414. https://doi.org/10.1609/icaps.v31i1.15986

er
Yakovlev, K., Andreychuk, A., & Vorobyev, V. (2019). Prioritized Multi-agent Path Finding
for Differential Drive Robots. 2019 European Conference on Mobile Robots (ECMR),
1–6. https://doi.org/10.1109/ECMR.2019.8870957
Yin, J., Zhang, M., Ma, Y., Wu, W., Li, H., & Chen, P. (2024). Prediction and Analysis
pe
of Airport Surface Taxi Time: Classification, Features, and Methodology. Applied
Sciences, 14(3), 1306. https://doi.org/10.3390/app14031306
Zaninotto, S., Gauci, J., & Zammit, B. (2021). A Testbed for Performance Analysis of
Algorithms for Engineless Taxiing with Autonomous Tow Trucks. 2021 IEEE/AIAA
40th Digital Avionics Systems Conference (DASC), 1–9. https://doi.org/10.1109/
ot

DASC52595.2021.9594411
Zheng, Y., Ma, H., Koenig, S., Kline, E., & Kumar, T. K. S. (2023). Priority-Based Search
for the Virtual Network Embedding Problem. Proceedings of ICAPS-23, 33, 472–
480. https://doi.org/10.1609/icaps.v33i1.27227
tn

Zou, X., Cheng, P., Liu, W., Cheng, N., & Zhang, J. (2018). A two-stage taxi scheduling
strategy at airports with multiple independent runways. Transportation Research
Part C: Emerging Technologies, 95, 165–184. https://doi.org/10.1016/j.trc.2018.07.
005
rin

Zou, Y., & Borst, C. (2024). Zeta*-SIPP: Improved Time-Optimal Any-Angle Safe-Interval
Path Planning. Proceedings of IJCAI-24.
Zoutendijk, M., Mitici, M., & Hoekstra, J. M. (2023). An investigation of operational man-
agement solutions and challenges for electric taxiing of aircraft. Research in Trans-
portation Business & Management, 49, 101019. https://doi.org/10.1016/j.rtbm.
ep

2023.101019
Pr

ed
Appendix A. Additional Pseudocode of Algorithms

Algorithm 3: High-Level Search of AS-MAMP as adaptation of PBS from (Ma,

Harabor, et al., 2019) (changes marked by red line-numbers)

iew
Input: MAMP instance, searchMode, rootConstraints, rootPriorities (= ∅ by default)
## root-node: SAMP-plan for all agents, adhere to root-constraints
1: root ← SearchNode(type="root") # create root-node as class-instance
2: root.plan ← ∅
3: root.rootConstraints ← rootConstraints
4: root.priorities ← rootPriorities

ev
5: foreach agent ai do
6: success ← root.updatePlan(ai ) # invokes SIMP, updates plan in-place
7: if not success then
8: return "no solution"

r
9: root.cost ← cN # see Eq. (1)
## child-nodes: detect and resolve conflicts between agents
stack ← root
10:
11:
12:
13:
14:
while stack ̸= ∅ do

conflict ← N.detectConflict()
if no conflict then
er
N ← pick and remove node from stack according to searchMode
# get conflict pair, see Algorithm 4
pe
15: break # N contains conflict-free solution

16: N.children ← ∅ # conflict exists: resolve it for both agents

17: foreach ai in conflict do
18: N’ ← SearchNode(type="child")
19: N’.plan ← N.plan
N’.rootConstraints ← N.rootConstraints
ot

20:
21: N’.priorities ← N.priorities ∪ {aj ≺ ai } # agent aj has priority over ai
22: success ← N'.updatePlan(ai ) # updates ai and all ak for which ai ≺ ak
23: if success then
tn

24: N’.cost ← cN # see Eq. (1)

25: N.children ← N.children ∪ {N’}

26: stack ← stack ∪ {N.children in order of decreasing cost}

27: return N.plan of solution-node N OR "no solution"
rin

Output: conflict-free solution to MAMP-instance if existent and found

ep
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874
von der Burg, Kamphof, Soomers, & Sharpanskykh

ed
iew
Algorithm 4: Conflict detection: without or with induced constraints (IC)

ev
Input: graph reservations of all agents
1: conflicts ← ∅
2: foreach agent ai do
3: foreach agent aj do
## check if conflict detection is needed between ai and aj

r
4: skipDetection ← checkConflictDetectionNeeded(ai , aj )
5: if skipDetection then
6:

7:
8:
continue

er # conflict detection not required

## detect path-conflicts between ai and aj
foreach edge in plan of ai do
if edge in plan of aj then
pe
9: conflict ← checkMotionOverlap(motions of ai , aj along edge)
10: conflicts ← conflicts ∪ {conflict}

## detect neighbourhood- and WTC-conflicts between ai and aj

11: foreach edge in plan of ai do
12: foreach reservation of aj on edge do
13: if motion of ai overlaps with reservation then
ot

14: conflict ← store conflict-information

15: conflicts ← conflicts ∪ {conflict}
tn

16: if without IC then

17: conflict ← conflict in conflicts with min(tstart )
18: else
19: conflict ← conflict with max(IC); breaking ties with min(IClower ), then min(tstart )
20: return conflict
rin

Output: agent pair in conflict

ep
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4916874

Navigation Assessment
100% (6)
Navigation Assessment
77 pages
Robot Testing
No ratings yet
Robot Testing
22 pages
Air Taxi Service For Urban Mobility A Critical Rev
No ratings yet
Air Taxi Service For Urban Mobility A Critical Rev
33 pages
Concept of Operations For Uncrewed Urban Air Mobility v2.0 1
No ratings yet
Concept of Operations For Uncrewed Urban Air Mobility v2.0 1
68 pages
Time of Arrival
No ratings yet
Time of Arrival
35 pages
Prediction and Analysis of Airport Surface Taxi Ti
No ratings yet
Prediction and Analysis of Airport Surface Taxi Ti
26 pages
Phases of Flight Poster
100% (2)
Phases of Flight Poster
1 page
Smart Aviation Towards Secure and Collaborative Flight Delay Prediction Using Machine Learning
No ratings yet
Smart Aviation Towards Secure and Collaborative Flight Delay Prediction Using Machine Learning
12 pages
TBO-Systems and Actors by Phase of Flight
No ratings yet
TBO-Systems and Actors by Phase of Flight
51 pages
Air Taxi Service For Urban Mobility A Critical Rev
No ratings yet
Air Taxi Service For Urban Mobility A Critical Rev
34 pages
Highlights: Machine Learning-Enhanced Aircraft Landing Scheduling Under Un-Certainties
No ratings yet
Highlights: Machine Learning-Enhanced Aircraft Landing Scheduling Under Un-Certainties
43 pages
Avionics Systems
No ratings yet
Avionics Systems
37 pages
Perspective Article Mishra Et Al PREPRINT
No ratings yet
Perspective Article Mishra Et Al PREPRINT
34 pages
Segmented Standard Taxi Routes-A New Way To Integr
No ratings yet
Segmented Standard Taxi Routes-A New Way To Integr
22 pages
ATM Ops Review V1.00 Public
No ratings yet
ATM Ops Review V1.00 Public
108 pages
Segmented Standard Taxi Routes-A New Way To Integr
No ratings yet
Segmented Standard Taxi Routes-A New Way To Integr
22 pages
1 Simulation - Models
No ratings yet
1 Simulation - Models
81 pages
Huỳnh Trần Xuân Nguyên _ 2431310306
No ratings yet
Huỳnh Trần Xuân Nguyên _ 2431310306
14 pages
Artificial Intelligence Systems in Aviation
No ratings yet
Artificial Intelligence Systems in Aviation
27 pages
Efficient Passenger Boarding Powered Fast Airplane Turnaround
No ratings yet
Efficient Passenger Boarding Powered Fast Airplane Turnaround
18 pages
Analyzing Air Taxi Operations From A Sys
No ratings yet
Analyzing Air Taxi Operations From A Sys
16 pages
Automation in The Aviation Industry
No ratings yet
Automation in The Aviation Industry
5 pages
Airport Taxiway Optimization
No ratings yet
Airport Taxiway Optimization
44 pages
Algorithms 06 00494
No ratings yet
Algorithms 06 00494
18 pages
Collocation Methods To Minimum-Fuel Trajectory Problems With Required Time of Arrival in ATM
No ratings yet
Collocation Methods To Minimum-Fuel Trajectory Problems With Required Time of Arrival in ATM
23 pages
Master - Fatima Ali Alleem
No ratings yet
Master - Fatima Ali Alleem
1 page
Science
No ratings yet
Science
23 pages
4D Trajectory Planning of Aircraft Taxiing Conside
No ratings yet
4D Trajectory Planning of Aircraft Taxiing Conside
12 pages
1 (32) 1
No ratings yet
1 (32) 1
17 pages
A Reinforcement Learning Approach To Vehicle Coor - 2024 - Green Energy and Inte
No ratings yet
A Reinforcement Learning Approach To Vehicle Coor - 2024 - Green Energy and Inte
18 pages
AR & VR Projects Portfolio by SmartTek Solutions
No ratings yet
AR & VR Projects Portfolio by SmartTek Solutions
23 pages
Predicting Taxi-Out Time at Congested Airports With Optimization Based Support Vector Regression Methods
No ratings yet
Predicting Taxi-Out Time at Congested Airports With Optimization Based Support Vector Regression Methods
12 pages
Aircraft Taxi Time Prediction
No ratings yet
Aircraft Taxi Time Prediction
23 pages
RPL 15
No ratings yet
RPL 15
14 pages
Rathinam ICAS2008
No ratings yet
Rathinam ICAS2008
14 pages
Aircraft Taxi Time Prediction Comparisons and Insights
No ratings yet
Aircraft Taxi Time Prediction Comparisons and Insights
10 pages
Using Computer Simulations To Enhance Science Teaching and Learning
100% (1)
Using Computer Simulations To Enhance Science Teaching and Learning
10 pages
A4 (S1) IT Consultancy Presentation
No ratings yet
A4 (S1) IT Consultancy Presentation
16 pages
Throughput Maximizing Takeoff Scheduling For EVTOL Vehicles in On-Demand Urban Air Mobility Systems
No ratings yet
Throughput Maximizing Takeoff Scheduling For EVTOL Vehicles in On-Demand Urban Air Mobility Systems
14 pages
rMulticriteria-Optimized Trajectories Impacting Today - Osenow2018
No ratings yet
rMulticriteria-Optimized Trajectories Impacting Today - Osenow2018
8 pages
Will Trajectory Based Operations
No ratings yet
Will Trajectory Based Operations
20 pages
1 s2.0 S2405844020324245 Main
No ratings yet
1 s2.0 S2405844020324245 Main
8 pages
1 (32) 1
No ratings yet
1 (32) 1
11 pages
Icas2024 0963 Paper
No ratings yet
Icas2024 0963 Paper
13 pages
01-Altair Technology For Digital Product Developments
No ratings yet
01-Altair Technology For Digital Product Developments
67 pages
Software Architecture and Design
No ratings yet
Software Architecture and Design
18 pages
Keynote 2
No ratings yet
Keynote 2
19 pages
Aircraft Taxi Route Planning For A SMGCS Based On Discrete Event
No ratings yet
Aircraft Taxi Route Planning For A SMGCS Based On Discrete Event
5 pages
Optimization of Airport Ground Operations
No ratings yet
Optimization of Airport Ground Operations
23 pages
Tool For Analysis of Traffic of Vertical Take-Off and Landing Aircraft in Urban Agglomerations
No ratings yet
Tool For Analysis of Traffic of Vertical Take-Off and Landing Aircraft in Urban Agglomerations
11 pages
Greener Aircraft Taxiing
No ratings yet
Greener Aircraft Taxiing
8 pages
Collision Risk Modeling of Air Traffic: September 2003
No ratings yet
Collision Risk Modeling of Air Traffic: September 2003
7 pages
1 s2.0 S2210539523000779 Main
No ratings yet
1 s2.0 S2210539523000779 Main
14 pages
A Study of Aircraft Taxi Performance For
No ratings yet
A Study of Aircraft Taxi Performance For
16 pages
Wuolah Free Lecture 7
No ratings yet
Wuolah Free Lecture 7
8 pages
Total Airport Management Platform With ASI DSS 1730053412
No ratings yet
Total Airport Management Platform With ASI DSS 1730053412
10 pages
Presentation Teknia RD 2013
No ratings yet
Presentation Teknia RD 2013
16 pages
GRC 10 Online Training
No ratings yet
GRC 10 Online Training
75 pages
19 and 20 hw
No ratings yet
19 and 20 hw
8 pages
2D2012-1-Tutorial Plaxis PDF
100% (1)
2D2012-1-Tutorial Plaxis PDF
0 pages
PLAXIS Introductory 2014 - Tutorial Manual 11
No ratings yet
PLAXIS Introductory 2014 - Tutorial Manual 11
104 pages
Time Based Fuel Efficient Aircraft Descents Thrus - 2023 - Transportation Resea
No ratings yet
Time Based Fuel Efficient Aircraft Descents Thrus - 2023 - Transportation Resea
19 pages
Computer Simulation of Fast Hydraulic Actuators
No ratings yet
Computer Simulation of Fast Hydraulic Actuators
12 pages
ICAO DOC 9870 App PDF
No ratings yet
ICAO DOC 9870 App PDF
10 pages
Whitepaper - The Power of Digitalization in Battery Cell Manufacturing
No ratings yet
Whitepaper - The Power of Digitalization in Battery Cell Manufacturing
48 pages
Ansys All You Need To Know About Hardware For Simulation
No ratings yet
Ansys All You Need To Know About Hardware For Simulation
36 pages
Self-Assessment Form Instructions For Applicants: Page 1 of 12
100% (1)
Self-Assessment Form Instructions For Applicants: Page 1 of 12
12 pages
MSC Programme: School of Engineering and Design
No ratings yet
MSC Programme: School of Engineering and Design
198 pages
UK566220 - Pro-Series 8000i Artemis - Operation
No ratings yet
UK566220 - Pro-Series 8000i Artemis - Operation
27 pages
Enabling Green Approaches by FMS-AMAN Coordination
No ratings yet
Enabling Green Approaches by FMS-AMAN Coordination
20 pages
Inventory Management Applications For Healthcare Supply Chains
No ratings yet
Inventory Management Applications For Healthcare Supply Chains
7 pages
Python Project: How To Manage A Speed Sensor With A Labjack U3 HV
0% (1)
Python Project: How To Manage A Speed Sensor With A Labjack U3 HV
9 pages
The Past, Present and Future of CFD For Agro-Environmental Applications - 2013
No ratings yet
The Past, Present and Future of CFD For Agro-Environmental Applications - 2013
16 pages
RFPro Circuit Simulation
No ratings yet
RFPro Circuit Simulation
13 pages
MIL - Module 15 16
No ratings yet
MIL - Module 15 16
25 pages
Verification and Validation of Simulation Models
No ratings yet
Verification and Validation of Simulation Models
37 pages
State of Mechanical Engineering Today and Beyond
No ratings yet
State of Mechanical Engineering Today and Beyond
24 pages
A Virtual Reality Platform For Safety Training in Coal Mines With
No ratings yet
A Virtual Reality Platform For Safety Training in Coal Mines With
7 pages
FEM Reporting Guidelines-2014
No ratings yet
FEM Reporting Guidelines-2014
5 pages
Rps Mini Project Final - Removed
No ratings yet
Rps Mini Project Final - Removed
11 pages
Eletric Lighting Energy Consumption Simulation
No ratings yet
Eletric Lighting Energy Consumption Simulation
11 pages
IT Innovations in Aviation Operation
No ratings yet
IT Innovations in Aviation Operation
29 pages
STRATEGIC
No ratings yet
STRATEGIC
7 pages
Electrical, Electronics and Instrumentation
No ratings yet
Electrical, Electronics and Instrumentation
2 pages
A Framework For EVTOL Performance Evaluation in Urban Air Mobility Realm
No ratings yet
A Framework For EVTOL Performance Evaluation in Urban Air Mobility Realm
7 pages
Efficient Selection of Agitators: Sulzer Pumps
No ratings yet
Efficient Selection of Agitators: Sulzer Pumps
2 pages
Resume Ala Ghardaddou
No ratings yet
Resume Ala Ghardaddou
1 page