AI and space safety: collision risk assessment

Abstract

The space environment has experienced drastic changes in the last years and the scenario for the next decade is expected to evolve notoriously. Mega-constellations with thousands of satellites, commercial small satellites, and low thrust propulsion spacecraft will not be the exception but the norm. Current safety strategies and collision avoidance procedures will no longer be capable to deal with the increase of information and alerts this new environment will imply. In this context, Artificial Intelligence is presented as the alternative approach for space safety, due to its ability to deal with a great amount of information, support decision-making and automatizing.

Full text

AI and space safety: collision risk assessment
Luis Sanchez∗
, University of Strathclyde
Massimiliano Vaisile, University of Strathclyde
Edmondo Minisici, University of Strathclyde
Abstract
The space environment has experienced drastic changes in the last years and the scenario for
the next decade is expected to evolve notoriously. Mega-constellations with thousands of satellites,
commercial small satellites, and low thrust propulsion spacecraft will not be the exception but the
norm. Current safety strategies and collision avoidance procedures will no longer be capable to
deal with the increase of information and alerts this new environment will imply. In this context,
Articial Intelligence is presented as the alternative approach for space safety, due to its ability to
deal with a great amount of information, support decision-making and automatizing.
Introduction
These days, the space environment is under a radical transformation that aects technologies, use
of the space, mission concepts and operations. Electrical propulsion, once proved its reliability and
capabilities, has started to be used during the last decade on commercial and scientic satellites,
both in Low Earth Orbit (LEO) and Geostationary Earth Orbit (GEO), and its use is expected to
grow. Late 1990’s technology improvements have resulted in miniaturization of space components
that eventually have allowed satellites to reduce its size. Since 2003, when the rst CubeSat was
launched, the use of such small satellites by universities or for commercial usage has continuously
increased and it is expected to keep growing during the next decade. Along with this increase on
small satellites, a higher rate of launches per year, and new countries and private actors entering
the scene are also expected. Among these new actors, maybe the most relevant due to its impact
on the orbit environment, will be swarms and constellation mission. Along with small satellites,
mega-constellations will represent the most profound change in the LEO regime during the next
decade. Several of these constellations, each of them compounded by thousands of satellites, are
planned and some of them have already started the deployment stage. It is expected that in the
next years, the number of satellites in orbit will multiply by several times. Bearing in mind that the
current number is slightly below 2,000, it will push the gures to tens of thousands of operational
satellites in orbit at the same time (Hugh et al. 2017).
On top of this, the most common elements in Earth orbits are, however, space debris objects.
Space debris refers to all man-made objects in space apart of operational satellites as well as micro-
meteoroids captured by the Earth’s gravity. It includes upper stage rocket bodies, inoperative
satellites remaining in orbit, objects left by missions and fragment from old satellites due to frag-
mentation or collision. From the beginning of the Space Era in 1958, the number of space debris
objects has kept growing to reach the current state where there are in orbit more than 34,000 objects
bigger than 10cm, more than 900,000 between 1cm and 10cm and millions of them even smaller
(ESA Report 2019). These numbers are also expected to increase in the following years, not only
linked to the increase in space trac, but also due to improvements in the current tracking tech-
niques. New infrastructures are expected to start their operation in the next decade allowing the
detection of smaller objects, which have not been possible to track until now. While this increase
in the cataloged objects does not mean an increase in the actual number of objects since they are
already in orbit, it will boost the number of conjunction alerts experienced by satellite operators
(Haimerl and Fonder 2015).
∗Corresponding author: luis.sanchez-fdez-mellado@strath.ac.uk
1

Figure 1: Evolution of number of objects in orbit: 1958 - 1984 - 2016. (Credits
Dr. Stuart Grey)
What does it mean for safe operation of satellites? First of all, the collision between operational
satellites between them or operational satellites and pieces of space debris will become a more
tangible threat. So far, there has been only one accidental collision event involving an operational
satellite, the Iridium-33/Cosmos-2251 collision in February 2009, which occurred when the number
of space objects was small compared with the expected future gures and while a conjunction
alert system was already operational. An event of this characteristics is likely to happened again,
especially regarding the increasing space population.
In the second place, the operation of those new satellites, both small satellites with reduced
propulsion capabilities and mega-constellations operating in already congested regions, will bring an
increment on the work load and dedicated time for operators to manage the possible conicts. Not
only during their operational life, but also during deployment and disposal, when these thousands
of satellites will cross other populated regions below their operational altitude. In addition to
this, new satellites are likely to be equipped with low thrust engines that on the one hand are
more ecient but on the other require more time for executing maneuvers. Longer maneuver
times means more time for planning a collision avoidance maneuver (CAM) and longer periods
crossing populated regions during deployment, disposal and routine maneuvers. In this context,
it is worth to mention the CAM the ESA’s satellite Aeolus was forced to implement to avoid one
of the StarLink constellation satellites. It was the rst one implemented by any ESA’s satellites
for avoiding a mega-constellation satellite in which gaps on the communications between operators
and lack of protocols were notorious. It has occurred even when the constellation is not completely
deployed, highlighting the impact of future New Space’s satellites and the precariousness of the
system at critical stages when agreeing a common avoidance strategy between operators.
On the third place, the increase on cataloged objects, operational satellites, and the potential
new space debris objects will translate in an unmanageable number of conjunction alerts received by
operators. The occurrence of these alerts does not necessarily mean the collision is going to happen,
but they are high resource-consuming for operators, since a detailed risk conjunction assessment is
required. With current levels of space trac, ESA’s Space Debris Oce has to deal with hundreds
of conjunction alerts per week only aecting the near 20 satellites they operate. Among these alerts,
just a small percentage are actionable alerts (alerts which require a more detailed analysis or the
collection or better quality data), and just one among 5-10 of these actionable alerts required an
avoidance maneuver (Merz et al. 2017). The expected boost of alerts during the next years, even if
they not require any actions, can collapse the current system, taking into account the work load and
time it requires and the coordination it implies. If besides, the Probability of Collision, the metric
used for evaluating events as high-risk or low-risk conjunctions, presents important limitations
(note that the aforementioned Iridium-33/Cosmos-2251 collision event presented a Probability of
Collision not classied as high risk by several operators), the system leads to a catastrophic result
unless major renovations are implemented (Peterson et al. 2018).
Among those renovations, automation is a major one. A shift from a system in which each
satellite is operated by several agents to a system where only one operator can manage several
satellites is desirable. However, such a situation is not possible with the current system structure,
especially considering the expected trac growth. It is at this point where Articial Intelligence
(AI) plays a crucial role. AI techniques can operate faster than current models and can take
decision considering a wider set of parameters than human operators and have the capacity to
perform better when the available data increases, which is the scenario expected for the next years
in space. If a certain set of reliable data is provided, AI systems are able to learn directly from them
and predict accurate results without the need for any physical model. In a scenario where more and
more data will be available and when time is a critical resource, using the surrogate model these
techniques provide can be the key for the automation of the Space Trac Management system
requires. While only a few examples of AI applied to Space Trac Management can be found,
they have been successfully used for predicting events, classication and decision support in other
2

engineering elds, including space and air trac management. This allows thinking that AI systems
are a promising trend for the next years.
The rest of the Chapter deepens on the application of AI in space engineering in general, and in
Space Trac Management in particular. Beginning with, a summary of the current situation of the
Space Trac Management (STM) and Space Situational Awareness (SSA) systems is presented,
highlighting the critical situation for the future regarding the expected increase in space trac.
An overview of studies about AI in the eld of trac management, collision avoidance, and space
engineering is then presented, followed by a survey of the main works on the application of AI
on the STM system. Finally, some challenges to be addressed for a good implementation of AI
techniques are stated.
AI and space safety: collision risk assessment.
Space safety system
A fundamental concept in space safety is Space Trac Management (STM), which is dened as
”the set of technical and regulatory provisions for promoting safe access into outer space, opera-
tions in outer space and returns from outer space to Earth free from physical or radio-frequency
interference” by the IAA (International Academy of Astronautics) in the Cosmic Study on Space
Trac Management. This concept includes a wide eld where dierent knowledge areas play a role
on space safety. On one hand, there are the rules, standards, and recommendations related to the
satellite operations, maneuvers, conict resolution and collision avoidance. This group also includes
the protocols to be implemented if a conjunction between two operational satellites is reported as
well as the good practices on sharing satellite’s operations information. On the other hand, are the
technical aspects whose aim is the implementation of the previous protocols and good practices for
the safe operation of the satellites, including tracking of space objects, conjunctions detection and
risk assessment as well as action for the mitigation of the risk of collision.
Another concept related to space safety is Space Situational Awareness (SSA) that involves
the actions, techniques, and technologies for the tracking, orbit determination and calculation of
ephemerides of the space objects. Both SSA and STM are closed related since the STM system
needs the knowledge provided by SSA about the state of the satellites to provide conjunction alerts,
perform the correct CAM if needed... Combined, these two systems create a more complex one that
involves information of thousands of space objects, requires the coordination of dierent operators,
satellite owners, and teams, provides alerts and recommend actions to be taken whose consequences
have to be managed in a short interval of time.
Continuous monitoring of all the trackable space objects around the Earth, both operational
and non-operational satellites, is carried out by SSA services providers. The main actor is the US’s
USSTRACOM, although commercial companies and other states are getting more relevance in the
last years. When a potential encounter between one operational satellite and a piece of space debris
or another operational satellite is detected when propagating the observable states, a Conjunction
Data Message (CDM) is created and sent to the operators in charge on the involved satellites. Since
information collected by the SSA system (USSTRACOM) presents low quality, especially for space
debris objects, the observable state and the propagated one are aected by uncertainty. A more
dedicated following can be carried out to reduce this uncertainty. With all the available CDMs
associated with a single event, a conjunction risk assessment is executed by the operators’ CARA
(Conjunction Assessment Risk Analysis) team to determine if the event represents a true threat
for the satellite and the space safety or not. In the case of a high probability of collision associated
with the event, a complex process starts. The rst step, if the event involves two operational
satellites, is agree a common strategy, relying on manual communication between operators that
delay the process. The common procedure (the two satellites moves, the biggest one moves, the
one with propulsion system moves...) is then analyzed with the payload team, ight dynamics
team, and ground stations to come up with a possible collision avoidance maneuver strategy. This
step requires a lot of coordination, time and workload as it is critical for the success of collision
avoidance. Secondly, the proposed strategy is then evaluated to ensure the risk of the current event
is reduced and no future possible collision arises: with the same object (secondary collisions) or
with other bodies (tertiary collisions). Eventually, once the maneuver is approved, the event is
closely monitored and one or two days before the Time of Closest Approach (TCA), as long as the
risk remains high, the CAM is performed. After the CAM is executed, the state of the satellite
should be monitor again to check the maneuver has been correctly performed.
It can be seen how many critical points the performance of a single CAM associated with only
3

one event presents. In the rst place, information of conjunction is provided just a few days in
advance what gives a tight interval of time for the whole process to be performed. Since CDMs
are available for the last 7 days before TCA, this is the time window operators have. However,
the actual time interval is shorter since rst CDMs present high uncertainty and better quality
data are usually required. In the second place, the computationally expensive and time consuming
some of these step are, something that is added to the tight time window where the process is
carry out. If better quality data are demanded for the conjunction risk assessment, sensors require
time for providing accurate orbit determination information. Not only that, but also accurate orbit
propagation is time-consuming, and it is an operation that has to be implemented in several stages:
using actual orbits for obtaining the risk of collision, for evaluating dierent CAM proposals or
for assessing future collisions once the CAM is implemented. It is not just a time issue. Besides,
coordination eort is a key aspect of the process. Flight control, ight dynamics, ground station
teams restrictions and mission requirements have to be considered when evaluating the possible
collision avoiding strategies. The coordination tasks would be even more critical if the potential
collision involves two operational satellites when teams of both missions have to agree a common
strategy, a problem that worsens due to the lack of protocols and specic regulations (Peterson
2019).
On top of that, STM is not just responsible for performing the CAM, but also it has to manage
all the conjunction alerts received before the collision risk assessment process stipulates the event
represents high probability of collision or not. All those alerts that do not need a conjunction
risk assessment and those that after the assessment do not require a CAM are considered as false
negatives. They do not give any information about real collisions but increase the operators’
workload. There is a greater number of not actionable alerts than actual high-risk events, which
means that an important part of the resources is spent on events that are not relevant for space
safety. Contrary to these false alarms (false positives) there is the possibility of false negatives to
occur. False negatives are those high-risk events that are misclassied, which can lead to collision
or risky events not noticed by operators in advance. As mentioned before, the collision between
Iridium-33 and Cosmos-2251 was a situation like this (Peterson et al. 2018). The root of these events
resides, partially, in the bad quality of initial position data, especially for the space debris objects,
what makes the acquisition of better quality information essential, bringing more information to
be manage by the system.
Note that the situation presented shows the current state of the system, where the trac of
space objects has not experienced the expected next years growth. The implementation of a CAM
explained above involves only a pair of space objects, however it involves multi-disciplinary teams
to coordinate a lot of information in a very constrained interval of time. False alerts and false
positives mentioned in the previous paragraphs currently happen. The increase on launches rate
programmed for the next decade leads to the question of the scalability of the system, the nal
issue STM and the space safety system will face scalability. The increase on space trac will
make operators struggle in managing all the available information and sub-optimal decisions are
likely, with eects on space safety. If currently, hundreds of alerts are triggered, the future space
environment will push this number to limits that the system may not cope with. Since more
resources should be put on ltering false alarms, the assessment of collision risk and mitigation
strategies will suer from this increase of alerts. Besides, future operators’ systems would not be
based on a team formed by several operators taking care of one or a few satellites, but smaller
teams controlling a whole constellation with several satellites each. Such a situation is not possible
unless a greater level of automatizing is implemented and the use of a decision support system is
used to replace most of the operators’ tasks (Nag et al. 2018).
4

Figure 2: Operative, on deployment and planned constellation on LEO region.
SSA, a fundamental part of the system, also introduces critical points to the process. It is
responsible for obtaining position information about all the objects orbiting the Earth, satellites,
rocket bodies or pieces of debris, as well as the ephemerides of those bodies. As can be expected due
to a large number of objects orbiting the Earth, the amount of information the system has to deal
with is already enormous, not counting for the expected growth of space trac of the next decade.
Furthermore, the next years will witness the start of the Space Fence system operations, the new
SSA system developed by the US for improving the monitoring of space objects. The expected
increase in sensitivity will allow the track of smaller objects, invisible at the moment, including
bodies in the range of 1 to 10cm, making it possible to include up to 200,000 orbiting objects in the
catalogs. It means that information related to space objects position and ephemerides will increase
even more at a rate much bigger than the numbers of launches since most of the new objects that
will enter in the catalogs are already in orbit. Another contributor to the SSA system are the
commercial providers, that independently to the traditional sources, carry out track campaigns,
whose information has to be merged with those from the US and agencies’ catalogs (Crosier 2016).
Figure 3: Current LEO catalogue versus expected catalogue when Space Fence
is operative.
The New Space environment presents similarities with Air Trac Management (ATM) and
Unmanned Air Trac Management (UTM) systems, where the increased on trac population
have forced them to adapt themselves to the new circumstances. ATM system is a well established
system which has coordinated an increasing air trac population for several years. Key aspects
of this system is the clear distribution of responsibilities, a proper set of protocols and common
5

practises, and the eectiveness to control several objects under only one control center, which has
facilitated the automation of activities previously carried out by human operators. However, as
in STM, the population growth, especially in certain regions, has forced actors involved on the
system to develop a more automatic system (Kochenderfer et al. 2011). UTM is another example
where automation has been implemented to handle the rapid trac increase and the necessity of
a quick decision-making process. Some studies show the eciency of implementing automation on
the UTM and the possibility to adapt the proposed system structure to the STM (Murakami et al.
2019). Furthermore, Decision Support Systems (DSS) based on AI has started to be implemented
on unmanned aerial vehicles (UAVs) control systems not for automatizing, but for supporting
operators on the decision-making stage. These approaches use AI, fuzzy logic and other related
techniques for rapidly taking into account a wide set of parameters and compute a ranking of the
best options to implement under a conict, automatizing tasks previously done by operators and
speeding up the whole process. Operators can then select the appropriate actions based on the
ranked list of alternatives based on certain criteria, reevaluating alternatives under new criteria
or recomputing the list if more information is available. While STM presents its particularities
respect UTM or ATM, it is clear that when trac management systems have experienced the
congestion of the environment, they have tended to the automation of the system, usually relying
on AI techniques.
The successful examples of applying AI on other engineering elds, including space engineering
and trac management have boosted the interest on using these techniques on the STM system for
automatizing tasks, speeding up the process, or supporting operators on taking optimal decisions
in an environment that is overcoming the capacity of human operators since more and more data
and variables have to be considered. Among the actors interested on implementing AI for STM
ESA and NASA can be named, both with programs to study the availability and applicability of AI
methods onto real missions and scenarios (Benjamin Bastida et al. 2019), (Mashiku et al. 2018).
ESA has identied three main issues to be addressed by AI for facing the population increment on
the orbital environment: reducing operators’ workload (automation), lowering the decision-taking
time on risk conjunction assessment and collision avoidance planning, and scaling down the number
of false alerts.
Articial Intelligence in engineering
Articial Intelligence is referred to as the ability of computers for learning from data, reasoning,
acquire knowledge, react to the environment and corrected themselves to imitate human intelligence
or behavior without being specically programmed to do it. It is a wide knowledge area including
Machine Learning, Natural Language Representation, Computer Vision, Data Mining among many
others (Russell and Norvig 2009). It has been studied for some decades, but only during the
last years, with faster and more capable computers and the availability of big dataset, it has
been possible its implementation into real applications in a broad range of disciplines, including
engineering.
Figure 4: Articial Intelligence areas.
Among some of the applications of AI in engineering, one interesting eld related to space safety
6

is trac management and collision avoidance. An important trend in recent years is the application
of AI on autonomous cars. Image recognition, intelligent decision systems, and autonomous collision
avoidance are issues presented in this eld and addressed by AI. However, the applicability of those
techniques to a completely dierent environment as it is space it is not a straightforward task and
it is currently under research. The development of robotics has also brought some improvements
in autonomous collision avoidance algorithms. Regarding the increasing autonomous of satellites,
bringing them closer to the general idea of what a robot is, some attempts of extrapolating those
algorithms to the space environment have been analyzed and it is an interesting research area where
promising results are expected on the next years.
The See Trac Management (SeeTM) system presents also some examples of the application
of articial intelligence on collision avoidance. While space and maritime environment presents
notorious dierences, there is also some similarities, like an initial sparse and wide operational space
which has experimented an increase on trac density, having led to the necessity of implementing
autonomy on the trac management systems, or regions where this density is reaching current
system limits, like ports on the see environment and the LEO region in space. In this sense, it
is interesting the work presented in Statheros et al. (2008) for applying an intelligence system for
ships collision avoidance, combining physical models with AI methods.
In the eld of trac management, there are also examples of using AI in the Aircraft Trac
(ATM) system and Unmanned Aircraft Trac Management (UTM) systems. Autonomy, both on
the vehicles and on the operators’ activities, is spread on these systems, although not necessarily by
using AI. Nevertheless, the increase in air trac and the irruption of commercial UAVs interacting
with the convectional aerial trac have forced the system to implement AI techniques for supporting
the operators on the management of the system. (Kochenderfer et al. 2011), (Julian and Lopez
2016) and (Ramirez Atencia 2017).
The space sector is getting interested in AI too, having incorporated techniques and methods
in dierent areas. Natural language processing (Berquand et al. 2018), knowledge representation,
automated reasoning, computer vision (Jasiobedski et al. 2001), trajectory optimization and nav-
igation (Izzo et al. 2018), satellite autonomy (Anderson et al. 2009) or robotics are some of the
elds in space engineering where AI have made interesting contributions.
Articial Intelligence in space safety
Seems clear there is a well-established eld of research and application of AI in dierent elds
of engineering, including dealing with conict, managing trac and supporting decision-making.
Based on the studies presented in the previous section, it is reasonable that AI and Machine
Learning (ML) methods can be applied also in STM. As mentioned previously, space agencies
and other actors involved on space safety have started to implement lines of investigation on this
direction and it is worth to explain in more detail the three main issues AI is expected to solve
according to ESA (Benjamin Bastida et al. 2019):
• Reducing the tasks operators currently carry out by implementing automation. Future in-
crease in space trac will translate in growth on the time and eort operators will spend just
dealing with alerts, classifying events, performing detail conjunction risk assessment, planning
and executing maneuvers, collecting better data or managing end-of-life strategies. Currently,
some of these activities present a certain degree of automation, while others require several
dedicated hours. Investing in the automation of most of these activities will allow operators
to focus on the decision-making stage, on the nominal operation of satellites or the handling
of more satellites simultaneously. Another important area where automatizing can liberate
much of the operator’s time is on the coordination between teams and other operators in the
event of a conict, switching for the current manual procedures for a much more automatic
one, with clear protocols and standardized steps.
• Lowering decision-taking time. Automation of operators’ tasks will allow them to spend more
time and eort on the critical steps of decision-taking in collision risk assessment, collision
avoidance maneuvers or disposal strategies evaluations. However, the expected rise in space
population will imply the number of satellites to be controlled and the amount of information
to be considered will exceed human operators’ capacities. AI-based systems for supporting
on the decision-making stages, like DSS agents, will be able to handle all this information
and propose alternatives strategies to operators in much less time than current approaches
taking into account a wider range of variables. Besides, surrogate models provided by AI
techniques for skipping computational expensive propagator or dynamical models, or the uses
7

of databases with predened maneuvers examples to automatically nd the optimal one are
other AI-based options for reducing time in the future STM system.
• Reduce false alarms. Currently, the vast majority of conjunction alerts reported to operators
correspond to events that not required any additional action (neither avoidance maneuvers
nor a more detailed evaluation). While triggering alerts, this kind of events do not imply true
collision scenarios, but consumes time and resources unnecessarily. In the next decade, when
smaller objects can be tracked and more satellites will be in orbit, the number of such events
will boost and more resources would be needed only to lter the actual collision encounters or
high-risk events from all the non actionable cases. Correctly selecting events without missing
the high-risk ones (false negatives) nor wasting resources on false alerts (false positives) will
be as essential as challenging for future STM, regarding current databases are dominated by
those less interesting low-risk events.
While AI has been used in other areas of space engineering, its application on Space Trac
Management and Space Situational Awareness is limited. However, it is possible to nd some
pioneer works on this subject. While scarce, they cover dierent aspects of the STM and SSA
system, addressing some of the previous aspects highlighted by ESA as priorities.
Some of those works are focus on improving orbit determination by the implementation of ML. In
Peng and Bai (2018a), Support Vector Machine is used for reducing the positional error of satellites
after orbit determination and orbit propagation processes. In Peng and Bai (2018b), they continued
with this line of research, switching from SVM to Articial Neural Networks (ANN). What they
proposed in those works is the use ML for improving orbital determination parting from the idea
that classical models keep unused certain embedded information from historical data. Using both
SVM and ANN, they tested the models for predicting a satellite’s position and velocity error caused
by measurement and dynamic propagation model limitations. Using the historical information of
a certain resident space object (RSO) during a interval of time, they expected to nd the relation
between them and the aforementioned error in three circumstances: for the same RSO in the same
interval of time of the historical data, but at epochs not including in the training set, for the same
RSO but for times after those included on the historical data, and for near RSOs, both in the
same interval as the training data and posterior epochs. They demonstrated by dierent numerical
experiments the possibility of using ML for reducing orbit determination error and thus, improving
orbit position knowledge. The benets of this method for the SSA system are clear. While SSA
is responsible to keep track of all the thousands of RSOs orbiting the Earth, the accuracy of
observations and models is restricted due to the great number of objects and limited knowledge of
the environment when building the models. Being able to correct the errors associated with them,
especially derived from imperfect modeling of the dynamics (drag, solar activity...) and limitations
of the observation sensors will automatically provide a better position for detecting conjunctions
and evaluating their risk. Nevertheless, the same authors are aware of some of the limitations
aecting this approach as reported on another publication (Peng and Bai 2017). Lack of real data
for propagating, time window limitations on the predictions and restricted generalization to other
objects dierent than the ones used for training are some of them.
Figure 5: ANN general structure Sanchez et al. (2019)
The previous approach correct orbit determination and propagating errors, but it is still limited
8

to the orbit propagation of each of the object of interest, which is time consuming when several
bodies are considered. Sanchez et al. (2019) proposes an ML-based system for predicting collision
encounters by using a set of ANNs for predicting the equinoctial parameters of a satellite during
an interval of time, by providing exclusively the initial Keplerian parameters. By comparing the
predicted orbits of a couple of objects, the equinoctial parameters obtained during the whole interval
of time are good enough for estimating potential conjunctions by calculating the impact parameter
(B-parameter) between the two bodies. In the end, the proposed method it is not anything else but
a surrogate orbit propagator that substitutes the dynamic models by a surrogate model based on
data. Some interesting aspects explain the importance of this approach and summarize some of the
general advantages of ML. First of all, this method provides a surrogate model of the underlying
problem (the two bodies perturbed movement) that does not rely on any dynamic model nor uses
any integration method (nor analytical nor numerical). Since no integration is involved in the
propagation, it performs faster. We are moving towards an environment where thousands of pieces
can be a threat to the operational satellites and where operators will be responsible not only for
one but for several of them, including constellations. Moreover, the tracking system will struggle
on providing good positional data from every piece of space debris at any time. Possessing a fast
and accurate model able to compute the propagate orbit of these thousands of satellites becomes
crucial for the future of STM. The second advantage this approach presents is that the model relies
on the data used for training. As in Peng and Bai (2018a) and Peng and Bai (2018b), dynamic
model errors are avoided since ML does not use any physical model, but builds one based on the
available data. In this way, by using the historical real position data, the uncertainties associated
with drag, solar radiation pressure and any other physical eects dicult to model simply do not
inuence the nal result. As can be seen on the results proposed in Sanchez et al. (2019), the error
is not dependent on the closeness to the initial epoch, as it usually happens on dynamic based orbit
propagators, since an independent set of six ANN has been trained for each epoch based on the
real orbital parameters of the training RSOs. This work is presented as a rst step towards the use
of ML in STM and, therefore, also presents some limitations: data used for training (assumed as
real position) comes from a virtual database obtained by using a high delity propagator and the
conjunction events prediction is made assuming the Keplerian propagation of one of the satellites
involved on the conjunction. However, despite these limitations and using a relatively simple ANN
model, it can provide accurate results for equinoctial parameters and detection of conjunction events
for RSOs dierent from those used during training. In addition, it performs quickly compared to
orbital propagators when several object’s orbits are propagated. Despite providing preliminaries
results, it sets a promising path for using ML in orbit determination and orbit propagation.
Figure 6: ANN for orbit propagation and conjunction event prediction. Sanchez
et al. (2019)
Other approaches have been followed for applying AI in STM. In Sanchez et al. (2020), ML
algorithms have been tested for classifying conjunction events based on a new approach for evaluat-
ing the risk assessment. The new approach pretends to overcome some limitation of a common risk
assessment metric, Probability of Collision, by using the Belief and Plausibility concepts coming
from Evidence Theory, accounting thus for epistemic uncertainty on collision risk assessment. This
approach takes the conjunction geometry between the two objects involved in a conjunction and
includes the uncertainty from the point of view of the Evidence Theory. Assuming one or more sets
of statistical distributions, each parameter dening those distributions is provided by the dierent
sources (i.e. sensors or experts) as an interval or intervals, without assuming any distribution of
the parameters but only the true value is included on one of them. Each of this intervals is as-
sociated a basic probability assignment accounting for the reliability of each source, which allows
9

taking into account aleatory and epistemic uncertainty independently. The classication criteria
proposed for conjunction events is then based on the time to the encounter and Belief and Plau-
sibility thresholds. Some ML methods, like Articial Neural Networks, Random Forests, Support
Vector Machine, and K-Nearest Neighbours have been tested for creating two dierent intelligent
classication systems, one using as inputs values of Belief and Plausibility as well as time, the other
considering time and geometry, allowing skipping the time consuming step of computing the Belief
and Plausibility curves. Each of the classes are related with an actions that would be suggest to
operators in the decision-making process. The results proposed in this work the potential used of
ML for decision-making support. Another intelligent decision support system is presented in Vasile
et al. (2017). The idea of the proposed method is supporting operators in the planning and imple-
mentation of collision avoidance maneuvers when needed. An interesting contribution of this work
is the creation and exploitation of a database of possible predened maneuvers to be implemented
in a conjunction event scenario. A virtual satellite position dataset was created to obtain conjunc-
tion events and later, computing the optimal maneuvers, which were stored in a new database.
The new orbits generating after the CAM were also storage in the initial database and analyzed
for detecting future encounters and thus, obtaining a wider range of CAMs for feeding the ML
algorithms. The availability of a database with these characteristics, with a broad variety of pos-
sible maneuvers, provides fundamental information to an intelligent decision system for providing
alternative proposals based on certain criteria. The criteria selected on this worked considered the
risk of not executing the collision compared with the risk associated with future possible collisions.
The ideas presented on those works lay the foundations for future intelligent DSS for supporting
operators. Other criteria can be implemented on the AI-based DSS to elaborate more sophisticated
ranked lists of proposals to the operator like condent on the sources, the inherent risk of executing
a maneuver, the cost of the maneuver versus the cost of the satellite itself, restriction due to mission
requirement or fuel usage limits. Sophisticated DSS system accounting for several variables and
proposing alternatives in relative short interval if time has already proposed in other elds of trac
management (Ramirez Atencia 2017).
There is another aspect of the satellites’ mission crucial for space safety: the disposal and re-
entry stages. The end of life of a satellite aects space safety in several aspects. First of all, it
is essential for decrease the rate of space objects in orbit, since it is the easiest way of removing
bodies from space, second, during the decay stage satellites have to cross highly populated regions,
something that will become more critical when mega-constellations are completely disposed of.
Finally, it is an extremely uncertain stage since atmosphere drag start to be the dominant eect and
density models are imprecise, solar activity is still not well modeled and knowledge on the behavior
of satellites during re-entry is hard to predict. Minisci et al. (2017) presented a study for uncertainty
propagation during the last stages of a GOCE mission. Besides the uncertainty quantication and
characterization study, the use of High Dimensional Model Representation (HDMR) methods and
the creating of large databases have set the path for the future use of AI on re-entry time windows
prediction. In the same work, meta-models based on AI were preliminary studied for mapping initial
stated and model uncertainties to re-entry time windows. Initial results suggest that it is possible
to estimate the re-entry window by this method and without any propagation. Further analysis is
being carried out for a better implementation of this idea and results suggest the potentiality of
this approach.
Some other works and studies relating AI with space safety, STM and SSA have been carried out
recently. Furfaro et al. (2019) used a Recurrent Neural Network (RNN) and Convolution Neural
Networks (CNN) for classifying and characterizing RSOs based on their curve of light for STM.
In Mashiku et al. (2019), supervised and unsupervised ML algorithms and Fuzzy Logic have been
implemented for predicting close approaches by using not only the classical probability of collision
but other parameters as well. Finally, Shabarekh et al. (2016) uses a ML approach for predicting
where and when may maneuver will be executed in the future to improve SSA capabilities.
Challenges for the future
AI is a promising approach for being implemented in STM to face the challenges of the new space
environment expected for the next years. Space agencies, operators and commercial agents have
shown interest in these techniques to ensure the future of space satellites, and there are already
ongoing researches for addressing the issues. However, being a new approach meaning it has to
face several challenges before we can talk about a space safety system based on AI.
As can be seen from current and past studies, a common problem is the lack of appropriate
datasets for training the models. AI techniques are based on the availability of enough informa-
tion to t the models, extract information or capture the patrons relating data. However, actual
10

information from real satellites is not always available in the desired format or with the required
quality. Indeed, orbiting satellites are periodically tracked, allowing accessing to a great amount of
historical data, however, some of these objects are not tracked with good enough quality to allow AI
techniques to extract reliable information or more accurate results than traditional methods. On
the other hand, some information is not available at all, like maneuvers implemented by satellites,
or the information is not enough to allow the AI models extracting patterns. Therefore, current AI
techniques relay on simulated databases, that have the advantages of creating a broad casuistic.
However, an important challenge for the coming years regarding the implementation of AI is the
creating of databases with information coming from real scenarios: real CDMs, information about
implemented collision avoidance maneuvers, uncertainties associated with measurements and state
propagation...
Articial Intelligence involves a wide set of branches. So far, space safety have just scratch
the surface on the application of those techniques in STM. Most of the methods implemented and
studied are centered on the Machine Learning branch, more specically on supervised learning.
However, there is a wide range of possibilities in AI where STM can take techniques from. Intelligent
Problem Solving, including Evolutionary Computing and Constraint Satisfaction Programming, can
be an interesting branch for DSS development along with Fuzzy logic, Automating Reasoning or
Knowledge Representation. Computer vision and image recognition are also open areas where STM
can benet from, besides Data Mining. The implementation of AI in STM is still a new research
area, but the potential for solving some of the problems already identify is huge. The advantage
is that AI is a more tested technology in other elds, including engineering. As has been seen,
trac management has already beneted from AI, and Space Engineering has already been used
AI techniques for some years. STM system has now the possibility of taking that experienced and
apply for its issues.
There is still another challenge to face, as it is the implementation of these kinds of techniques
onto real applications. The work carried out so far is focused on proving the capability of these
techniques to improve the STM system and ensure space safety in the oncoming scenario. However,
there is still a long way for being able to implement those techniques on real missions or in the
actual system. More research has to be done for really understating the relation between training
AI models and the physical laws ruling the data, more detailed studies for optimizing techniques
should be performed as well as adapting the system for gradually incorporated the proven methods.
It is now a perfect time for testing new approaches since the space environment is changing and new
techniques are not advisable but mandatory for the sustainability of the system, but at the same
time, it is critical to implement reliable methods in order not to collapse the system. This leads to
the last and main challenge the implementation of AI in STM has to face: the lack of standards
on STM. Several AI-based approaches can be suggested, but as long as there are no protocols of
actuation and standardized actions in conict situations, the problem of a congested space will still
be there. AI techniques as a way for supporting operators and moving to an automated scenario
will work as long as a set of common rules and practices are shared by the dierent agents using
the space.
Conclusions
New Space will bring great challenges to space safety in the next decades. The implementation
of new technologies, new concepts of satellites and new kinds of missions, like low-thrust engines,
small satellites or mega-constellations, will push the limits of the space system to its limits. On
top of all of this, the problem of space debris, which is going to become worse with the increase in
space trac, will make it completely necessary to carry out drastic changes on the system in order
not to collapse it.
Although these changes can come from dierent approaches, there is a consensus on the space
community that automation of the Space Trac Management and Space Situational Awareness
systems is one of them. To achieve the required level of automation, AI techniques arise as the
most promising tool due to a series of factors. Their ability to deal with huge amount of data, and
not only that but also learning from them and improving performances when more information
is available, the advances on computer systems that allow its implementation both in the ground
segment and in-orbit computers, the wide range of elds of application and task they can be applied
to or the possibility to speed up the process where they are used and the capacity for automation
and decision-making support are just some of their advantages.
While used in other engineering elds, like trac management or computer vision among many
others, the application in space engineering started near in the past, focused on image recognition,
11

autonomous navigation, satellite autonomy, orbit trajectories or robotics. However, it is only in
recent years where space safety has started to implement AI techniques, where only a few promising
studies have been carried out. However, the trend followed by agencies and space actors points in an
increasing relevance of AI for STM since it may be the only tool able to handle all the information
the congestion environment expected for the next decade will generate.
Three main issues are expected to be addressed with the implementation of AI on space safety,
space trac management, and collision avoidance: automation of certain task to reduce operators
man workload, minimize time between decision (conjunction risk assessment or collision avoidance
planning and implementation) and reduce the number of false alerts in relation of potential high-risk
conjunction events.
However, as a starting technique on the eld, there are still some challenges to be overcome. A
common limitation already faced is the lack of proper database based on real scenarios. AI tech-
niques are based on the availability of representative data. The creation of appropriate databases
with information coming from real satellites, events, and scenarios, or at least, a database of vir-
tual scenarios closely similar to real situations is vital for obtaining the better performance of these
techniques. AI is a wide area with several elds. At this moment, only some of them have been
preliminary studied, mainly focused on the Machine Learning area. Studying dierent approaches
and performing analyses to determine to determine the best AI branch to solve each problem re-
lated to space safety is highly recommendable to obtain the maximum benets from AI. Finally,
lack of protocols and standardized practiced is a drag for obtaining the best performances of some
of these methods. A promising area on AI is the development of intelligent agents or intelligent
decision support systems. However, these methods required a series of clear rules to provide the
appropriate advice to operators. Agreeing on common rules and practices for all space actors is
essential for the proper implementation of AI in space safety.
References
Anderson J.L, Kurfess F.J, Puig-Suari, J (2009), A Framework for Developing Articial Intelligence
for Autonomous Satellite Operations. Proceeding IJCAI-09 Workshop on Articial Intelligence
in Space, Pasadena, California, US, 17-18 July 2009.
Benjamin Bastida V, Flohrer T, Krag, H, Merz K, Lemmens S (2019), CREAM - ESA’s Pro-
posal for Collision Risk Estimation and Automated Mitigation. 1st International Orbital Debris
Conference, Sugar Land, Texas, USA, 9-12 December 2019.
Berquand A, Murdaca F, Riccardi A, Soares T, Gerené S, Brauer N, Kumar K (2018), Towards an
Articial Intelligence based Design Engineering Assistant for the Early Design of Space Missions.
68th International Astronautical Conference, Bremen, Germany, 2018.
Crosier C (2016), United States Strategic Command Space Situational Awareness Sharing Program
Update. United Nations Committee on the Peaceful Uses of Outer Space, Scientic and Technical
Subcommittee, 54th Session, Viena, Austria, 3 February 2016.
ESA SPace Debris Oce (2019), ESA´s Annual Environment Report. Technical Report, ESA,
2019, Available at https:\\www.sdo.esoc.esa.int/environment_report/
Furfaro R, Linares R, Reddy V (2019), Space Objects Classication and Characterization via
Deep Learning and Light Curves: Applications to Space Trac Management. Space Trac
Management Conference, Austin, Texas, USA, 26 February 2019.
Haimerl J, Fonder G (2015), Space Fence System Overview. Advanced Maui Optical and Space
Surveillance Technologies Conference Proceedings, September 2015.
Lewis H, Radtke J, Rossi A, Beck J, Oswald M, Anderson P, Bastida Virgili B, Krag H (2017),
Sensitivity of the space debris environment to large constellations and small satellites. Journal
of the British Interplanetary Society, 70, 105-117.
Izzo D, Martens M, Pan B (2018), A Survey on Articial Intelligence Trends in Spacecraft Guidance
Dynamics and Control. December 2018.
Jasiobedski P, Greenspan M, Roth G (2001), Pose Determination and Tracking for Autonomous
Satellite Capture. 6th International Symposium on Articial Intelligence, Robotics and Automa-
tion in Space, Montreal, Quebec, Canada, June 2001.
12

Julian K, Lopez J (2016), Policy compression for aircraft colli-sion avoidance systems. IEEE/AIAA
35th Digital Avionics Systems Conference (DASC), September 2016.
Kochenderfer M.J, Chryssanthacopoulos J.P (2011), Robust Air-borne Collision Avoidance through
Dynamic Programming. Lincoln Laboratory, Massachusetts Institute of Technology, January
2011.
Mashiku A, Frueh C, Memarsadeghi N (2018), Supervised-machine Learning for Intelligent Collision
Avoidance Decision-making and Sensor Tasking. 2018 NASA Goddard Workshop on Articial
Intelligence, November 2018.
Mashiku A, Frueh C, Memarsadeghi N, Gizzi E, Zielinki M, Burton A (2019), Predicting Satellite
Close Approaches in the Context of Articial Intelligence. AAS/AIAA Astrodynamics Specialist
Conference, Portland, Maine, United States, 11-15 August 2019.
Merz K, Bastida Virgili B, Braun V, Flohrer T, Funke Q, Krag H, Lemmens S (2017) Current
collision avoidance service by ESA´s Space Debris Oce. Proceedings 7th European Conference
on Space Debris, Darmstadt, Germany, 2017.
Minisci E, Serra R, Vasile M, Riccardi A, Grey S, Lemmens S (2017), Uncertainty Treatment in the
GOCE re-entry. 1st IAA Conference on Space Situational Awareness, Orlando, Florida, USA,
Nov 2017.
Murakami D, Nag S, Lifson M, Kopardekar P (2019), Space Trac Management with a NASA UAS
Trac Management (UTM) Inspired Architecture. AIAA SciTech Forum, San Diego, California,
7-11 January 2019. doi: 10.2514/6.2019-2004.
Nag S, Murakami D, Lifson M, Kopardekar P (2018) System Autonomy for Space Trac Manage-
ment. IEEE/AIAA Digital Avionics and Systems Conference, London, United Kingdom, Septem-
ber 2018.
Peng H, Bai X (2017) Limits of Machine Learning Approach on Improving Orbit Prediction Accu-
racy. Advanced Maui Optical and Space Surveillance Technologies (AMOS) Conference, Maui
Economic Development Board, Inc. Wailea Marriott, September 2017.
Peng H, Bai X (2018) Improving Orbit Prediction Accuracy through Supervised Machine Learning.
Advances in Space Research, 61(10):2628–2646, May 2018, doi:10.1016/j.asr.2018.03.001.
Peng H, Bai X (2018) Articial Neural Network–Based Machine Learning Approach to Improve Or-
bit Prediction Accuracy. Journal of Spacecraft and Rockets. 55:1248-1260. doi:10.2514/1.A34171.
Peterson G (2019), Establishing Space Trac Management standards, guidelines, and best prac-
tices. Center for Space Policy and Strategy, The Aerospace Corporation, September 2019.
Peterson G, Sorge M, Ailor W (2018), Space Trac Management in the age of New Space. Center
for Space Policy and Strategy, The Aerospace Corporation, April 2018.
Ramirez-Atencia C, Bello-Orgaz G, R-Moreno M.D, Camacho D (2017), Solving complex multi-
UAV mission planning problems using multi-objective genetic algorithms. Soft Computing,
21(17):4883-4900, September 2017, doi=10.1007/s00500-016-2376-7.
Russell S, Norvig P (eds) (2009), Articial Intelligence. A Modern Approach. 3rd Edition, Prentice
Hall Press Upper Saddle River, NJ, USA.
Shabarekh C, Kent-Bryant J, Keselman G, Mitidis A, (2016), A Novel Method for Satellite Ma-
neuver Prediction. Advanced Maui Optical and Space Surveillance Technologies Conference, 24
Wailea, Maui, Hawaii, USA, September 2016.
Statheros T, Howells G, McDonald-Maier K (2008), Autonomous ship collision avoidance navigation
concepts, technologies and techniques. Journal of Navigation, 61(1): 129-142, 2008.
Sanchez L, Vassile M, Minisci E (2019) AI to Support Decision Making in Collision Risk Assessment.
Paper presented at the 70th International Astronautical Conference, Washington, D.C., USA,
21–25 October 2019.
13

Sanchez L, Vassile M, Minisci E (2020) On the Use of Machine Learning and Evidence Theory to
Improve Collision Risk Management. Paper presented at the 2nd IAA International Conference
in Space Situational Awareness, Washington, D.C., USA, 14-16 January 2020.
Vasile M, Rodriguez-Fernandez V, Serra R, Camacho D, Riccardi A (2017) Articial Intelligence in
Support to Space Trac Management.Paper presented at the 68th International Astronautical
Conference, Adelaide, Australia, September 2017.
14