Multi-objective Multi-perspective Numerical Optimization of Collision Avoidance

Maneuvers Using Differential Evolution

Naman M Ladhad
Digantara Research and Technologies Pvt Ltd, Bengaluru, India
Rithwik Neelakantan
Digantara Research and Technologies Pvt Ltd, Bengaluru, India
Tanveer Ahmed
Digantara Research and Technologies Pvt Ltd, Bengaluru, India

ABSTRACT

The design of collision avoidance maneuvers in real case scenarios involves intricate decision-making processes,
demanding varying fidelity of data and processes at different stages. Mission constraints, propellant constraints,
reliability of collision risk estimation, nature of secondary objects and even operator’s schedules contribute to the
process of decision making. Therefore, it is imperative to adopt a multi-perspective approach to the problem
formulation involving many (if not all) of the above-mentioned aspects. In this context, the maneuver design for
collision avoidance is formulated as a heuristic multi-objective multi-perspective optimization problem in this research
and the solution is obtained using Differential Evolution (DE), an evolutionary optimization technique. The objective
functions to minimize in the problem formulation are a) mass of fuel used b) the collision probability after maneuver(s)
c) the deviation of the maneuvered trajectory from the non-maneuvered nominal trajectory and d) disruption time of
routine payload operations (defined as the time span for which the spacecraft deviates from its nominal orbit). The
solution search space consists of the number of maneuvers, start times and durations of each maneuver and the
components of each maneuver. Altitude constraints are modelled as boundary conditions to the optimization problem.
The optimal solutions found indicate that the primary objective of collision avoidance is accomplished, reducing the
value of collision probability to commonly accepted limits. The penalties of conducting sub-optimal maneuvers are
quantified by comparison with the optimal solution. When the objective function doesn’t include the propellant
consumed, it is found that there is a significant saving in the optimal solution equivalent to the fuel used for collision
avoidance maneuver. Similarly, when the objective function doesn’t consider the deviation of the maneuverd
trajectory, unacceptably large deviations are incurred. The proposed methodology allows the satellite operators to
explore a number of design scenarios and tune the solutions to the exact requirements.

1. INTRODUCTION AND BACKGROUND

Humanity’s utilization of space has grown exponentially in the last two decades. There is a rapid increase in the
number of Resident Space Objects (RSO) due to the launch of many new satellites and their associated debris. The
rate at which close approaches of RSOs are being reported has alarmed the space community at large, calling for a
robust infrastructure to monitor the RSOs, to replace legacy systems of sensors and for international cooperation
towards sustainable and judicious use of space. Any event of an actual impact spells catastrophe, as a single collision
not only leads to the loss of the objects, but also generates thousands of fragments of debris, dispersed practically
randomly in new trajectories, further increasing space junk which poses a severe danger of future collision to other
RSOs. Thus, a predicted conjunction with a quantified high level of risk warrants the need to consider and plan a
Collision Avoidance Maneuver (CAM).

SpaceX alone performed more than 25,000 CAMs between Dec 1, 2022, and May 31, 2023 [1]. These numbers
doubled to nearly 50,000 between Dec 1, 2023, and May 31, 2024, after SpaceX decreased the threshold for
performing evasive maneuvers by an order of magnitude, while adding more satellites to its mega-constellation,
increasing the count to 6,200 spacecraft [1]. The updates in collision risk metrics and increase in RSOs all point to the
fact that the number of CAMs performed by operators are only going to increase moving forward.

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
Performing a CAM, however, is an undesirable event from a satellite operator's perspective, as it consumes available
propellant from a satellite affecting the total mission lifetime and consequently, the revenue generated. Furthermore,
performing a CAM requires an interruption of normal operations of the payload and deviation from the nominal
trajectory. Studies have also shown that performing CAM for a predicted conjunction event increases the risk of
collision downstream in the future, due to inadequate methods of calculating orbital uncertainties in the aftermath of
a CAM, rendering subsequent collision predictions to be less than accurate for several days [2].

In practice, real case scenarios involve intricate decision-making processes, demanding varying fidelity of data and
processes at different stages. Mission constraints, propellant constraints, reliability of collision risk estimation, nature
of secondary objects involved and even operator’s schedules contribute to the process of decision making. Therefore,
it is imperative to adopt a multi-perspective approach to the problem formulation involving many (if not all) of the
above-mentioned aspects.

Taking into consideration the multitude of decisions and objectives to monitor for performing a CAM, tackling it as
an optimization problem is beneficial to obtain maneuver solutions that minimize the amount of fuel used while also
optimally reducing the Collision Probability (Pc), having a great significance on the mission lifetime. Formulation of
an optimization problem allows for modelling of mission operations as well.

2. LITERATURE REVIEW

There is immense wealth of literature for the optimization of collision avoidance maneuvers and most of them adopt
the reduction of collision probability and minimization of delta-V as the objective function. A gradient based
optimization technique was employed in [3] to determine the optimal CAM solution and the determination of optimal
maneuver direction and magnitude were decoupled. This allowed the optimal maneuvering direction to be determined
using the gradient of the collision probability with respect to the maneuver direction and the optimal maneuver
magnitude to be determined using simple root finding numerical techniques. The design of optimal CAM is posed as
a convex optimization problem and solved using linearized dynamics in [4]. The maneuvers are modeled as a set of
impulses allowing it to be used with propulsive systems across any thrust range and tackles the objective of propellant
consumption with constraints either on collision risk or miss distance. An analytical solution to the problem of finding
miss distance and collision probability is described in [5] wherein the optimal maneuver direction is determined as
solution to the linearized eigenvalue problem. The solutions for maximizing the miss distance and minimizing the
collision probability are presented and the differences are highlighted. The solved numerical examples indicate that
when the maneuver is carried out less than several orbits ahead, the optimal direction is different from being purely
tangential and involves a significant out-of-plane component [5].

As a contrast to treating collision probability and miss distance separately, [6] tackles these two aspects simultaneously
as constraints to a non-linear optimization problem in addition to the maneuver time and time duration between the
maneuver and TCA. A two-step process is described wherein the first step is to conduct a grid search for forming the
initial guesses to the problem and the second step is the nonlinear optimization process. The optimality of the solutions
is compared against a Monte-Carlo based simulation as the reference.

Apart from the above-mentioned studies, there have been several efforts to incorporate the actual mission constraints
into the objective function formulation and increase the fidelity of the solution to better suit the requirements of the
satellite owner/operators. Because no two orbital missions are essentially the same, the problem formulation approach
is heuristic in nature and is typically solved using evolutionary optimization techniques. A multi-objective collision
avoidance maneuver design strategy for satellites in both GEO and LEO orbital regimes is proposed in [7]. The
different aspects considered are: 1) the fuel to be minimized, 2) the deviations of the maneuvered trajectory from the
nominal trajectory and 3) reduction in the collision probability post maneuver. The objective function includes a
scalarized approach where these three aspects are assigned different weights. Specific mission related constraints
such as allowable deviations from the GEO orbital slot are modeled as constraints to the optimization problem. A
genetic algorithm-based heuristic optimization process is used to derive optimal solutions. The flexibility of the
evolutionary algorithms to model the intricacies of the CAM design is demonstrated. Another alternate approach to
forming a multi-objective optimization problem is to consider the conflicting objective functions as separate,
independent formulations. The solution space will then consist of pareto-optimal fronts, with the possibility of

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
obtaining a number of optimal solutions rather than a single optimal solution in the case of the scalarization approach.
A CAM design strategy based on this approach was proposed in [8] and the optimization problem is solved using a
swarm particle optimization approach. The possibility of secondary and tertiary conjunctions was considered for the
design of CAM for objects in GEO and solutions were ensured to avoid these.

A multi-objective optimization using the NSGA-II algorithm was performed in [9]. In this study multiple objects are
considered for a collision threat with a satellite. The optimizing factors modelled were the delta-v and maneuver cycle.
Maneuver cycle here is defined as the time interval within which ground track error is maintained. Three different
burn strategies were employed for each simulation and their results compared in terms of different objective function
contributions.

Optimization of collision avoidance maneuvers was performed using a formulation based on self-adaptive differential
evolution in [10]. The single best minimal solution also includes re-insertion maneuvers to return to the nominal orbit
using Lambert’s problem, by exploring fuel-optimum transfer trajectories for re-insertion numerically. Miss distance
and collision probability are tackled as constraints in the study. The solutions are presented for both the two and three
impulse approaches for several cases of constraint profiles. The discussion of the solution also highlights the trade-off
between ∆𝑉 and the maneuver cycle. Maneuver cycle here is defined as the duration the satellite spends away from
it’s nominal orbit.

Based on the above-mentioned literature, the current work proposes a multi-perspective multi-objective formulation
of the design of collision avoidance maneuvers. Similar to the work described in [7], a single objective function is
formulated combining different aspects to be optimized using a scalarization approach. The solution approach employs
an evolutionary algorithm known as differential evolution. The aim is to determine a single-best optimal solution
which can be valuable to the satellite operator. The next section describes the methodology for the problem
formulation.

3. METHODOLOGY
3.1 Differential Evolution

Differential Evolution (DE) is a stochastic direct search method which mimics the evolution of living species [11].
For a problem of 𝐷 unknown parameters, the respective search domains are to be defined (for a vector space, the
corresponding components are to be defined as well). From these search bounds, an initial population of size 𝑁𝑃 is
built randomly, following uniform distribution and the objective function is evaluated for each member of the
population. The members of this population are tested for violation of path constraints, if any. A new member is
formed using three operations of mutation, crossover and selection. The new member will replace the existing member
if the corresponding objective function value is lesser than that of the existing member. The process is repeated till a
predefined convergence criterion is met. This basic variant of DE is denoted as DE/rand/1/bin.

3.2 Objective function

The motive of the work is to design optimal collision avoidance maneuvers considering the operational constraints
of owner/operators of the satellites. The different factors considered are:

1. Mass of the propellant used for the maneuver(s).

2. The probability of collision before and after the maneuver(s)
3. Payload interruption time: defined as the time from the start of the first maneuver to the end of the last
maneuver. This assumes that the satellite cannot perform its payload operations when it is maneuvering.
4. Deviation of the maneuvered trajectory from the nominal, non-maneuvered trajectory: quantified as the
difference of the states between the non-maneuvered and maneuvered trajectory in the final half orbit at the
end of the analysis duration.

Towards incorporating the above-mentioned aspects, the objective function to be minimized is formulated using a
scalarization approach as follows:

𝐹 = 𝐹1 + 𝐹2 + 𝐹3 + 𝐹4 + 𝐹5 (1)

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
where

𝐹1 = 𝑊1 ∗ (𝑙𝑜𝑔10 (𝑃𝑐 + 1)) (2)

𝐹2 = 𝑊2 ∗ 𝑚𝑝𝑟𝑜𝑝 (3)
𝐹3 = 𝑊3 ∗ 𝑡𝑖𝑛𝑡𝑒𝑟𝑟𝑢𝑝𝑡𝑖𝑜𝑛 (4)
𝑁

𝐹4 = 𝑊4 ∗ ∑|𝑃⃗𝑁𝑜𝑚𝑖𝑛𝑎𝑙_𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦 − 𝑃⃗𝑀𝑎𝑛𝑒𝑢𝑣𝑒𝑟_𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦 | (5)

𝑖=1
𝑁
⃗ 𝑁𝑜𝑚𝑖𝑛𝑎𝑙_𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦 − 𝑉
𝐹5 = 𝑊5 ∗ ∑|𝑉 ⃗ 𝑀𝑎𝑛𝑒𝑢𝑣𝑒𝑟_𝑇𝑟𝑎𝑗𝑒𝑐𝑡𝑜𝑟𝑦 | (6)
𝑖=1

In these equations, 𝐹 is the composite objective function. 𝑊1 to 𝑊5 represent the weights assigned to the individual
components of the objective function because the values of physical quantities have different orders of magnitudes.
𝑃𝑐 is the value of the collision probability, as read from the Conjunction Data Message (CDM). The formulation of
the component 𝐹1 is inspired by the objective function formulated in [8]. 𝑚𝑝𝑟𝑜𝑝 is the mass of the propellant in the
satellite and 𝑡𝑖𝑛𝑡𝑒𝑟𝑟𝑢𝑝𝑡𝑖𝑜𝑛 is the payload operations interruption time caused due to the maneuver(s). Terms 𝐹4 and
𝐹5 correspond to the deviation of the maneuvered trajectory from the nominal, non-maneuvered trajectory.

3.3 Design variables

The search space for the optimization problem consists of the following design variables: 1) the number of maneuvers
(𝑁𝑚𝑎𝑛 ) 2) the start times of each of the maneuvers 3) burn duration for each of the maneuvers and 4) the three
components of each maneuver accelerations in radial (𝑅), along-track (𝑆) and cross-track (𝑊) directions [13].

3.4 Search bounds and boundary conditions

In order to derive meaningful solutions from the design of collision avoidance maneuvers, the number of maneuvers
(𝑁𝑚𝑎𝑛 ) is restricted to lie between 2 and 5. It is ensured that at least one of the maneuvers is always conducted before
the time of closest approach (TCA) to avoid illogical solutions where the conjunction is not avoided, but other
objectives are met. Further, the analysis duration is restricted to span for two days; one day prior to TCA and one day
post the TCA. The duration of each burn is restricted to lie between 1 and 30 seconds. The thrust for each burn is
maintained constant at 5 N (with a specific impulse of 250s), resulting in search bounds for the acceleration
components. The mass of the primary (maneuvering satellite) is constantly updated during the numerical integration
of equations of motion. An assumed operational altitude constraint of 10km is also introduced, where the position
deviation of the maneuvered trajectory from the nominal trajectory is checked at every time-step of the numerical
integration process.

3.5 Computational algorithm

A step-by-step computational algorithm is mentioned below:

1. Re-creating the collision scenario: the purposes of this step are two-fold: a) it checks the conformance of
the close approach to the (different) data source and (different) processes used in-house compared to those
used for generating the CDM and b) generation of the non-maneuvered trajectory for comparison with the
maneuvered trajectory. First, using the data ingested from the CDM, initial states one day prior to TCA are
obtained through back propagation of states and covariances for both the objects involved in the conjunction.
These initial states are then propagated numerically for two days, thereby obtaining the nominal, non-
maneuvered trajectories for the entire analysis duration for both the objects.
2. Generation of initial population: An initial population of size 𝑁𝑃 is built. Each member of the population
consists of 5 ∗ 𝑁𝑚𝑎𝑛 scalar components (𝑁𝑚𝑎𝑛 for start times, 𝑁𝑚𝑎𝑛 for burn durations and 𝑁𝑚𝑎𝑛 ∗ 3 for
acceleration components) and the corresponding objective function values. The values for these unknowns

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
 are chosen randomly from their respective bounds. Evaluation of the objective function involves propagation
from the initial state with the simulation of maneuvers and the subsequent re-evaluation of terms in the
objective function (c.f. Eq. 1). For this, the collision probability is evaluated using the DCA of the
maneuvered trajectory at TCA and the interruption time is evaluated as the difference between the epoch at
the end of the last maneuver and the epoch the start of the first maneuver. The propellant used is calculated
based on the burn duration and the assumed mass flow rate of the propulsion system. Finally, the error in the
end states is obtained by the summation of the differences of the maneuvered trajectory from the nominal
trajectory.
3. Generation of trial population: A trial member from the search bounds is generated for each member of the
current population through the process of mutation, crossover and selection.
a. Mutation: A mutant vector is generated using some randomly selected members from the current
population such that they are not the same as the member under testing. A scaling factor denoted by
𝐹 is used for the mutation process, and the mutant member 𝑉 is generated according to the relation
𝑉𝑖 = 𝑈𝑅1 + 𝐹 ∗ (𝑈𝑅2 − 𝑈𝑅3 ). Here 𝑅1 , 𝑅2 and 𝑅3 are three distinct random integers chosen from
[1, 𝑁𝑃] and the variable 𝑖 varies between 1 and 𝑁𝑃. These members are chosen such that they are
different from the element under testing (𝑖 member), that is 𝑅1 , 𝑅2 and 𝑅3 must not be equal to 𝑖.
b. Cross over: The member of the current population under testing and the mutant member together
generates the trial member. A parameter ‘crossover frequency’ (𝐶𝑅) is used to generate a trial
member [11]. A random number 𝑟𝑎𝑛𝑑(𝑗) is generated between 0 and 1, for each component of the
𝑖 th member U for which trial member is to be generated. For each of the component (𝑗),
𝑖𝑓 𝑟𝑎𝑛𝑑(𝑗) > 𝐶𝑅, the 𝑗th component of the 𝑖th member of the current population is retained for
the trial vector and if 𝑟𝑎𝑛𝑑(𝑗) ≤ 𝐶𝑅, the component in the trial vector is replaced with the 𝑗th
component of the mutant vector.
c. The objective function for the trial member is evaluated and the member under testing is replaced
by this trial member if the objective function value is lower.
4. Generation of trial members and subjecting them to the above three operations is carried out for all the
members in the current population and thus a new population is generated.
5. The above-mentioned steps are repeated till a pre-defined convergence criterion is met. For the current
problem, the convergence criterion chosen is that the difference between the maximum and minimum values
of objective functions between two successive generations should be less than 1.E-3.

A high-level depiction of the above-mentioned computational algorithm is presented in Fig. 1.

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
 Fig. 1. A high level flowchart of the design of collision avoidance maneuvers. The aspects involving differential
evolution are marked in green.

4. RESULTS

A real-world conjunction scenario is chosen for the analysis and is generated using proprietary data ingested into the
in-house developed conjunction analysis tool. A close approach was identified between two active satellites 558xx
and 394xx in LEO, predicted at 08:41:00 UTC on 07/08/2024. The Distance at Close Approach (DCA) was 1.28808
km, with radial, along-track, and cross-track separations of -1.11 km, 156.58m and 631.78 m respectively. The relative
velocity at TCA was 14.09 km/s with the maximum collision probability value of 2.99598591E-04. The collision
probability technique employed is Alfano Max Probability [12]. The parameters other than the DCA used are Aspect
Ratio (𝐴𝑅) of 55 and a hard body radius of 5m. The value of aspect ratio, i.e. the ratio of standard deviations of major
and minor axis of the combined covariance ellipse is derived from the computed covariances. The design variable
number of maneuver (𝑁𝑚𝑎𝑛 ) is fixed at three for illustration purposes. The DE-related parameters are set as Mutation
factor 𝐹 = 0.5, cross over ratio 𝐶𝑅 = 0.7 and 𝑁𝑃 = 75. The weights used in the objective function are: 𝑊1 = 5𝐸5,
𝑊2 = 50, 𝑊3 = 1, 𝑊4 = 10 and 𝑊5 = 100.

The optimal solution is presented in Table 1.

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
 Table 1: Optimal maneuver design for the close approach
Burn Maneuver start Maneuver Radial Along track Cross track
no: (UTC) duration (s) Acceleration Acceleration Acceleration
(m/s2) (m/s2) (m/s2)

1 24/08/06 1.543 -0.025618 -0.010933 -0.019118

16:11:45.746

2 24/08/07 1.031 -0.006495 0.032982 -0.003367

12:02:05.953

3 24/08/08 2.130 -0.015870 -0.029657 -0.003162

08:38:01.549

The first maneuver avoids the close approach leading to a Pc reduction to a value of 5.2325e-5 with the new DCA
being 3.1061 km. The second and the third maneuver performed restituting burns to minimize the error between the
maneuvered primary satellite and its nominal orbit. The propellant mass used is 9.5944 grams. The ∆𝑉 for each burn
is 5.21 cm/s, 3.48 cm/s and 7.19 cm/s bringing the total consumed ∆𝑉 to 15.89 cm/s. The evolution of the maneuverd
trajectory in RSW coordinates is depicted in the Fig. 2.

Fig. 2. Evolution of radial, along-track and cross-track differences between maneuvered and non-maneuvered trajectories

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
In order to compare the optimal solution obtained and validate its optimality, a comparative analysis is carried out.
The comparative analysis excludes one objective from the objective function at a time by setting its associated weight
to 0. The simulation is then carried out, keeping every other parameter unchanged. The solution consequently obtained
is termed as a ‘sub-optimal’ solution. The penalty incurred in the objective excluded is then quantified against the
optimal maneuver design solution obtained. This process is repeated for every objective and the results are presented
in Table 2.

Table 2 Penalty incurred by exclusion of different components of objective function

Objective excluded Penalty incurred compared to optimal
solution

Propellant mass (𝐹2 ) / 11.9 grams of propellant lost (corresponds

∆𝑉 to ∆𝑉 of 19.72 cm/s)

Interruption time (𝐹3 ) - (6 hr 44 min)

End states (𝐹4 & 𝐹5 ) 53.16 km deviation at the end of analysis

As can be observed from Table 2, a position error of about 53 kms is incurred by excluding the end states. The high
position error in the sub-optimal solution is due to the three maneuvers introduced in the optimization, where all the
maneuvers have converged to a solution of a purely along-track burn, leading to a large displacement in the along-
track direction.

By excluding the interruption time, the sub-optimal solution performed better by about 6 hours as compared to the
optimal solution. This is identified as a trade-off solution where a higher interruption time is incurred, due to the
conflicting nature of the objectives in the optimal result. Further exploration into the dynamics of the evolution of
individual objective functions can be conducted to pinpoint the reason for this result. Finally, exclusion of the
propellant mass leads to a penalty of 11.9 grams of propellant or a ∆𝑉 of 19.72 cm/s as compared to the optimal
solution.

A normalized convergence is presented based on the difference between the maxima and minima in each generation.
The Y-axis represents the objective function normalized with respect to the highest value in the difference of the
objective function. The convergence is as depicted in Fig. 3.:

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
 Fig. 3. Convergence pattern in the differential evolution algorithm

5. CONCLUSIONS

The heuristic optimization of collision avoidance maneuvers using differential evolution is presented. A number of
objective functions including the mass of propellant, collision probability after maneuver, deviation of the maneuvered
trajectory from the non-maneuvered trajectory and reduction of payload interruption time are modelled. The solution
search space consists of the number, start time and duration of each maneuver and its components. Altitude constraints
are modelled as boundary conditions to the optimization problem. The optimization process is implemented on a real-
case close approach in LEO, and the optimal three burn solution is presented. The optimal propellant used for the
collision avoidance in the investigated case is found to be 15.89 cm/s. Interestingly, when the simulation is carried out
without considering the propellant used as an objective function, a penalty of 19.72 cm/s is found to be incurred. This
shows that inclusion of multiple objectives into the problem formulation is essential to derive optimal solutions.
Similarly, when the objective function doesn’t consider the deviation of the maneuvered trajectory, unacceptably large
position deviations are found to be incurred. The applications of the proposed methodology will benefit the satellite
owner/operators with optimal maneuver recommendations and informed decision-making contexts. The modular
nature of the methodology allows for a multitude of operational constraints to be modelled, offering flexibility to the
operator to tune the optimal it’s to their precise requirements. Extended mission life, confidence and reliability in
collision avoidance maneuvers and overall financial gain are the broad outcomes of the proposed research.

Further efforts can made to extend the study towards developing pareto-optimal solutions to present a set of solutions
due to the conflicting nature of the objectives. Use of more sophisticated collision probability techniques can be
adopted, while also accounting for position and maneuver uncertainties into the problem to present a more complete
picture of the performance of the suggested evasive maneuvers.

References

[1] Pultarove, T. (2024, July 23). SpaceX Starlink satellites made 50,000 collision-avoidance maneuvers in the past 6
months. Space. https://www.space.com/spacex-starlink-50000-collision-avoidance-maneuvers-space-safety
[2] Pultarove, T. (2023, December 11). Performing evasive maneuvers increases satellites' collision risk down the
road. Space. https://www.space.com/satellites-collision-avoidance-maneuvers-increase-collision-risk
[3] Patera, R. P., & Peterson, G. E. (2003). Space vehicle maneuver method to lower collision risk to an acceptable
level. Journal of Guidance, Control, and Dynamics, 26(2), 233-237.

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com
[4] Armellin, Collision avoidance maneuver optimization with a multiple-impulse convex formulation, Acta
Astronautica 186 (2021) 347–362
[5] Bombardelli, C., & Hernando-Ayuso, J. (2015). Optimal impulsive collision avoidance in low Earth orbit. Journal
of Guidance, Control, and Dynamics, 38(2), 217-225.
[6] Kelly, B., & De Picciotto, S. (2005). Probability-based optimal collision avoidance maneuvers. In Proceedings of
Space 2005 (pp. 6775). American Society of Civil Engineers.
[7] Lee, S.-C., Kim, H.-D., & Suk, J. (2012). Collision avoidance maneuver planning using GA for LEO and GEO
satellites maintained in keeping area. International Journal of Aeronautical and Space Sciences, 13(4), 474-483
[8] Morselli, A., Armellin, R., Lizia, A., & Zazzera, M. (2014). Collision avoidance maneuver design based on multi-
objective optimization. Advances in the Astronautical Sciences, 152(4), 1819-1838.
[9] Seong, J.-D., & Kim, H.-D. (2016). Multiobjective optimization for collision avoidance maneuver using a genetic
algorithm. Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering, 230(8),
1438-1447.
[10] Rajasekar, P. E. (2017). Optimization of space debris collision avoidance maneuver (Master's thesis, McGill
University, Canada).
[11] Storn, R., & Price, K. (1997). Differential evolution—a simple and efficient heuristic for global optimization over
continuous spaces. Journal of Global Optimization, 11(4), 341-359.
[12] Alfano, S. (2005). Relating position uncertainty to maximum conjunction probability. The Journal of the
Astronautical Sciences, 53(2), 193-205.
[13] Vallado, D. A. (2001). Fundamentals of astrodynamics and applications (Vol. 12). Springer Science & Business
Media.

Copyright © 2024 Advanced Maui Optical and Space Surveillance Technologies Conference (AMOS) – www.amostech.com