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

net/publication/260940014

Model predictive control application to spacecraft rendezvous in Mars Sample

Return scenario

Article · December 2013

DOI: 10.1051/eucass/201306137

CITATIONS READS

16 296

8 authors, including:

Alberto Bemporad Jan Maciejowski

IMT School for Advanced Studies Lucca University of Cambridge
352 PUBLICATIONS   26,477 CITATIONS    320 PUBLICATIONS   11,802 CITATIONS   

SEE PROFILE SEE PROFILE

Arthur Richards
University of Bristol
95 PUBLICATIONS   4,853 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Automotive Control View project

Wound Brushless Doubly Fed Machines View project

All content following this page was uploaded by Arthur Richards on 26 May 2016.

The user has requested enhancement of the downloaded file.

Progress in Flight Dynamics, GNC, and Avionics 6 (2013) 137-158
DOI: 10.1051/eucass/201306137
© Owned by the authors, published by EDP Sciences, 2013

MODEL PREDICTIVE CONTROL APPLICATION

TO SPACECRAFT RENDEZVOUS IN MARS
SAMPLE RETURN SCENARIO

M. Saponara1 , V. Barrena2 , A. Bemporad3 , E. N. Hartley4 ,

J. Maciejowski4 , A. Richards5 , A. Tramutola1 ,
and P. Trodden5
1
Thales Alenia Space (TAS) Italia
253 Strada Antica di Collegno, Torino 10146, Italy
2
GMV
11 Isaac Newton, PTM Tres Cantos, Madrid 28760, Spain
3
Department of Mechanical and Structural Engineering University of Trento
Trento, Italy
4
Department of Engineering, University of Cambridge
Cambridge, U.K.
5
Department of Aerospace Engineering, University of Bristol
Bristol, U.K.

Model Predictive Control (MPC) is an optimization-based control strat-

egy that is considered extremely attractive in the autonomous space
rendezvous scenarios. The Online Recon¦guration Control System and
Avionics Architecture (ORCSAT) study addresses its applicability in
Mars Sample Return (MSR) mission, including the implementation of
the developed solution in a space representative avionic architecture
system. With respect to a classical control solution High-integrity Au-
tonomous RendezVous and Docking control system (HARVD), MPC al-
lows a signi¦cant performance improvement both in trajectory and in
propellant save. Furthermore, thanks to the online optimization, it al-
lows to identify improvements in other areas (i. e., at mission de¦nition
level) that could not be known a priori.

1 INTRODUCTION

Within AURORA programme, the MSR mission is the main planned objective in
the international e¨ort on the Solar System exploration. Its main goal is to bring
back to the Earth a sample of Martian soil. A number of new technologies will be

This is an Open Access article distributed under the terms of the Creative Commons Attribution License 2.0, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Article available at http://www.eucass-proceedings.eu or http://dx.doi.org/10.1051/eucass/201306137

PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

required to carry out this pioneering mission and one of them is the rendezvous
and capture system, which will be able to detect, approach, and capture the
sample of Martian soil, previously put in a prede¦ned orbit by the Mars Ascent
Vehicle (MAV).
Although autonomous docking is now a well established technology, au-
tonomous capture (with a poorly cooperative target) is more delicate. The de-
velopment of a Guidance, Navigation and Control system (GNC) for rendezvous
and capture has been addressed in the European Space Agency (ESA) study
named HARVD. This study has been separated into two parallel activities, one
of them lead by GMV in collaboration with TAS France and Italia. The devel-
oped solution shows that, with classical control techniques, it is possible to have
an automated rendezvous and capture control system with preplanned operations
able to ful¦ll the MSR capture requirements.
Starting from HARVD experience, a further study has been de¦ned, named
ORCSAT. The objective of the study is to improve the HARVD GNC by means of
optimization-based control strategies such as MPC. The work on this study was
supported by the ESA under contract No. 22421.
Model predictive control (see, for example, [1 3]) is an advanced control
technique which uses a prediction model and numerical optimization methods
to obtain a sequence of control inputs that minimizes a function of the control
inputs and predicted plant state trajectory over a given time horizon, subject
to constraints. At each sampling instant, the optimization performed based on
new measurement data, and the ¦rst control input of the sequence is applied.
The remainder of the sequence is discarded and the process is repeated at the
next sampling instant in a ¤receding horizon¥ manner. Whilst MPC has its
origins in the chemical process industries [4], there is increasing interest in its
application to vehicle manoeuvre problems [5 7], including spacecraft trajectory
control [8 11] and attitude control [12 14]. Essentially, the application of MPC
builds upon the ideas of fuel and time-optimal trajectory planning by bringing
the optimization onboard, providing a natural framework for increased autonomy
and recon¦gurability, whilst accounting for physical and operational constraints
such as ¦nite control authority, passive safety, and collision avoidance.
The ORCSAT study considers also the developing of an MPC Framework
software (SW) tool (MPCTOOL) for supporting the design, analysis and sim-
ulation of MPC-based control systems as well as the development of embedded
model predictive controller for autonomous rendezvous control systems. Fur-
thermore, another key point of the ORCSAT study is the implementation of the
developed MPC control system into a space representative avionic architecture
system.
The paper will brie§y present the HARVD study. Afterwards, it will concen-
trate on the MPCTOOL description, the MPC design, and the avionic architec-
ture system. Finally, simulation results will be shown in comparison with the
HARVD ones.

138
 GUIDANCE AND CONTROL

2 THE HIGH-INTEGRITY AUTONOMOUS

RENDEZVOUS AND DOCKING CONTROL
SYSTEM STUDY

In the last years, the number of studies considering rendezvous and docking/
capture missions around Mars or other planets/asteroids has signi¦cantly in-
creased. As a consequence, it is surely worth dedicating e¨ort to consolidate
maturity of GNC technologies for such missions, in order to have onboard sys-
tems with a higher and higher level of autonomy, robustness, and safety, with
the ¦nal objective of decreasing costs and increasing the probability of mission
success. Following this tendency, a team led by GMV and including, among
others, TAS, has developed HARVD, an ESA-funded activity implementing a
GNC / Autonomous Mission Management (AMM) / Fault Detection, Isolation,
and Recovery onboard SW for rendezvous and docking/capture scenarios around
Mars, Earth, or potentially other planets [15 17]. The HARVD, based on radio
frequency (RF), camera, and LIDAR (light detecting and ranging) measure-
ments, includes design, prototyping, and veri¦cation at three di¨erent levels:
algorithms design and veri¦cation in a High-Fidelity Functional Engineering
Simulator, SW demonstrator to be veri¦ed in Real Time (RT) Avionics Test
Benching and Dynamic Test Benching. Rendezvous and capture on an elliptic
orbit have been specially addressed, demonstrating the technical feasibility and
the potential propellant saving.
The HARVD stepwise development and veri¦cation approach is shown in
Fig. 1.
The Development, Veri¦cation, and Validation (DVV) approach in the
HARVD activity relies on the use of COTS SW tools:

  Matlab/Simulink/State§ow from Mathworks, including associated tool-

boxes, for design, analysis, simulation, and validation of system models
and algorithms;

  TargetLink from dSPACE for automatic generation of production code (C

code) straight from the above graphical development environment; and

  dSPACE simulator for RT development/simulation environment.

The development and integration of the High-Fidelity Functional Engineer-

ing Simulator have been successfully completed, and an intensive test campaign
has been carried out. Interesting results for di¨erent MSR scenarios have been
obtained, demonstrating how the strict mission requirements on performances,
autonomy, safety, and robustness have been ful¦lled with high margins. A spe-
cial attention has been dedicated to contingency scenarios (including di¨erent
onboard system failures and collision risks detection and avoidance), for which

139
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

Figure 1 The HARVD stepwise development and veri¦cation approach and GMV£s
PLATFORM dynamic test bench (DKE ¡ domain knowledge engine)

140
 GUIDANCE AND CONTROL

the results obtained are very encouraging for the consolidation of higher Tech-
nology Readiness Levels. Mars ascent vehicle circularization failures have been
also taken into account, resulting in a number of elliptic target orbit rendezvous
scenarios for which HARVD has demonstrated to be fully ready.
The development of RT test bench has been concluded and the acceptance
RT test campaign has been successfully completed. The RT test bench is based
on a LEON board GR-PCI-XC2V @45 MHz, and computational load margins
of 32% have been achieved for the Worst Case Execution Time (WCET).
Recently, the tailoring of the GMV Dynamic Test Bench (PLATFORM, see
Fig. 1) has already started, and the dynamic tests are foreseen to be executed
in the next few months.

3 THE MPCTOOL

MPCTOOL is a MATLAB/Simulink toolbox providing all major features for the

design, analysis, and simulation of model predictive controllers based on linear
time-invariant (LTI) or linear time-varying (LTV) models, as well as for auto-
matic code-generation of embedded model predictive controllers. MPCTOOL is
tailored (although not limited) to the synthesis of autonomous rendezvous con-
trol systems. The inclusion of LTV capability is a key enabler for rendezvous,
since elliptical orbits and J2 e¨ects introduce time variation into the dynamics.
MPCTOOL extends the Model Predictive Control Toolbox from The Math-
works, Inc. [18] to introduce new features, modifying existing MATLAB objects,
adding new functions (MATLAB methods) based on them, introducing new ob-
jects and their methods, extending the C code of the S-Function behind the
basic LTI-MPC controller, and introducing new Simulink blocks coded in Em-
bedded MATLAB (EML) for LTV-MPC. Model predictive controllers designed
for LTI systems can be converted to explicit form [19] via the direct link be-
tween MPCTOOL and the Hybrid Toolbox for MATLAB [20]. Furthermore,
a new Dual-Simplex solver has been developed to manage optimization prob-
lems expressed as a Linear Programming (LP) problem [21]. The new features
introduced by MPCTOOL on top of the existing MPC Toolbox are the following:
  the ability to set terminal weights and constraints in LTI-MPC (including
in¦nite-horizon MPC);
  handle variable-horizon MPC problems in which the horizon length is op-
timized online;
  handle quantized inputs in LTI-MPC problems;
  return the optimal sequence of MPC (both in MATLAB scripts and in
Simulink), for example, to check a posteriori the enforcement of complex
constraints not modeled in the MPC optimization model;

141
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

  return the optimal cost of MPC for comparing and choosing the best action
among a set of model predictive controllers;
  allow the speci¦cation of convex piecewise a©ne stage costs (such as abso-
lute values) on inputs and outputs;
  handle arbitrary linear constraints on combinations of inputs and outputs;
and
  handle arbitrary linear time-varying models, weights, constraints, and hori-
zons by providing two Simulink blocks based on EML code, supporting
both quadratic programming (QP) and LP problem formulations.
The latter feature, namely, LTV-MPC based on LP, was employed in the
studies described in this paper and will be detailed next.
The LTV model predictive controller relies on the following rather general
linear time-varying prediction model:
⎫
x(j + Ts ) = A(j, x(t))x(j) + B(j, x(t))u(j) + f (j, x(t)) ; ⎪
⎪
⎪
⎪
⎪
⎪
y(j) = C(j, x(t))x(j) + D(j, x(t))u(j) + g(j, x(t)) ; ⎪
⎪
⎪
⎪
⎬
z(j) = Ez (j, x(t))(y(j) − r(j)) + Hz (j, x(t))(u(j) − ur (j)) (1)
⎪
⎪
⎪
⎪
+ Pz (j, x(t))–u(j) ; ⎪
⎪
⎪
⎪
⎪
⎪
⎭
c(j) = Ec (j, x(t))x(j) + Hc (j, x(t))u(j) + Pc (j, x(t))–u(j)
where Ts is the sampling time; k is the prediction step; t is the current time,
j = t + kTs is the prediction time; x is the state vector; u is the input vector; y
is the output vector; –u(j) = u(j) − u(j − Ts ) is the input increment; r is the
output reference vector; ur is the input reference; z is the ¤performance vector¥
to be optimized; c is the ¤constrained vector;¥ and A, B, f , C, D, g, E, H,
and P are (possibly time-varying and state-dependent) matrices.
The MPC optimal problem to be optimized at each time t is
⎫
4
N (t)−1 ⎪
⎪
min ρ1 1 + ρ2 2 +  z(j) 1 ; ⎪
⎪
⎪
⎪
k=0 ⎪
⎪
⎬
s.t. –umin (j) ≤ –u(j) , k = 0, . . . , N (t) − 1 ;
(2)
⎪
⎪
c(j) ≤ cmax (j) + Vc ρ1 , k = 0, . . . , N (t) − 1 ; ⎪
⎪
⎪
⎪
⎪
⎪
⎭
C (t)x(t + N (t)T ) ≤ d (t) + V ρ
N s N N 2

where N (t) ≤ Nmax is the prediction horizon; and ρ1 and ρ2 are the slack
variables used to soft constraints. Constraints are hardened by zeroing the cor-
responding entry in vectors Vc ≥ 0 and VN ≥ 0.

142
 GUIDANCE AND CONTROL

The optimal control problem (2) is mapped into the LP

⎫
4 4
N (t)−1 l ⎪
min ρ1 1 + ρ2 2 + di (j) ; ⎪
⎪
⎪
k=0 i=1 ⎬
s.t. di (j) ≥ ±zi (j) , di (j) ≥ 0 , (3)
⎪
⎪
⎪
⎪
⎭
+ MPC constraints

which has (m + l)N (t) + 1 optimization variables and, besides the nonnegativity
constraints –u − –umin ≥ 0,  ≥ 0, di (j) ≥ 0, 2lN (t) constraints to express the
1-norm in (3), plus as many constraints as the ones that are optionally de¦ned
in (2).
The user can exploit the maximum §exibility o¨ered by the EML language to
de¦ne the prediction model (1) and all the parameters appearing in the MPC op-
timization problem (2) in an EML module, which is then used by the LTV-MPC
Simulink block to construct and solve problem (3). Accordingly, as depicted in
Fig. 2, the block contains an LP builder function and a Dual Simplex LP solver
coded in EML code, implementing the LTV-MPC formulation described above.
The block is §exible enough to allow an arbitrary number of parameters enter-
ing the EML prediction model from the Simulink diagram as RT varying signals,
to vary online prediction and control horizons, to limit a priori the maximum
number of LP iterations.

4 THE ORCSAT MPC DESIGN

4.1 Control System Architecture and Choice of Prediction Model

There exist a large number of well-researched models for the prediction of the
relative dynamics of one spacecraft with respect to another [9, 22 29]. Whilst
a nonlinear model would provide the highest ¦delity predictions, for the pur-
pose of this study, it was judged that the possible gains would not be worth
the additional complexity in the optimizer. Similarly, integer decision vari-
ables have been avoided as the resulting integer program would also have ex-
cessive complexity. Discrete decision making is instead handled by solving mul-
tiple instances of continuous optimizations at each control step. Therefore,
the consideration is restricted to linear prediction models from which one can
form the MPC optimization problem as a convex quadratic or linear program.
However, out of the models considered, only the Hill Clohessy Wiltshire equa-
tions [22] are linear time-invariant, and these only apply to circular, or very-
near circular orbits. The other models are linear parameter-varying with re-
spect to the true anomaly of the target νtgt . However, because the target

143
144
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

Figure 2 Simulink diagram underlying the LP-based LTV-MPC block

 GUIDANCE AND CONTROL

Table 1 Rendezvous phases

Phase Requirement
Orbit Synchronization Transla- To bring chaser from a distance of approx-
tional Guidance (OSTG) imately 300 km into the same orbit as the
target, with an in-track separation of be-
tween 5 and 30 km on either side of the
target
Impulsive Nominal Translational To perform passively safe impulsive trans-
Guidance (INTG) fers between a sequence of prede¦ned hold-
ing points in the same orbit as the tar-
get until an in-track separation of 100 m
is reached
Forced Terminal Translational To track a straight-line trajectory from
Guidance (FTTG) 100- to 3-meter separation from the target
such that a subsequent free-drift trajectory
captures the target with a 20-centimeter
tolerance
Collision Avoidance Manoeuvre To bring the chaser to a safe distance, fur-
(CAM) ther than 5 km from the target within 3 or-
bits, avoiding collision in the process

is passive, νtgt can be calculated as a function of time using Kepler£s equa-

tion [30], thus allowing a linear time-varying representation of the relative dy-
namics.
The objective of the MPC control system designed during this study is to
bring the chaser craft from the point of target detection at a range of approxi-
mately 300 km, via a sequence of holding points in the same orbit as the target,
to a ¤blinding point¥ approximately 3 m from the target, at which point it
should be moving towards the target at an in-track velocity of 0.1 m/s. Target
capture is then completed on a passive drift trajectory. The MPC system pro-
vides both guidance and control and is not restricted to tracking predetermined
trajectories.
To achieve this objective using a single model predictive controller would re-
quire a prediction horizon su©cient to predict a trajectory at least one orbital
period into the future, a su©ciently complex prediction model to perform accu-
rate trajectory propagations over long periods of time, and a sampling period
short enough to allow target capture within a 20-centimeter tolerance. Given
¦nite computational resources, this is not a practical solution. The rendezvous
is, therefore, divided into three phases similar to those used in HARVD [31, 32],
with an additional controller to perform a collision avoidance manoeuvre (CAM)
in case of a fault during the ¦nal moments of the rendezvous (Table 1).

145
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

4.1.1 Orbit synchronization translational guidance

The ¦rst phase, Orbit Synchronization Translational Guidance (OSTG), has the
objective of bringing the chaser from a distance of approximately 300 km into
the same orbit as the target using thrusters, with an in-track separation of be-
tween 5 and 30 km on either side of the target, whilst minimizing propellant
consumption and manoeuvre time. At these ranges, short-term control accuracy
is not critical; so, a relatively long prediction time can be used. However, long-
term prediction accuracy is important in order to perform optimal manoeuvres.
For these reasons, the J2 -modi¦ed Gauss£s variational equation (GVE) predic-
tion model of [9] is chosen. This predicts the relative trajectory between the
chaser and target in terms of the relative Keplerian orbital elements rather than
relative positions and velocities in a rectangular or cylindrical coordinate frame,
whilst using the Gim Alfriend [29] approach of incorporating the e¨ects of J2 to
account for variations in gravity due to the oblateness of the central body of the
orbit. Because the relative orbital elements are small, despite large Euclidean
separations, the e¨ects of linearisation error are small in comparison to predic-
tion models such as those of [22,28], which use rectangular or cylindrical relative
coordinates. The system input is assumed to be an impulsive change in velocity
(–V ) in a local orbital reference frame centered on the chaser.

4.1.2 Impulsive nominal translational guidance

The second phase, Impulsive Nominal Translational Guidance (INTG), must per-
form a sequence of passively safe impulsive transfers between a sequence of pre-
de¦ned holding points until an in-track separation of 100 m is reached. Greater
control accuracy is required during this phase, necessitating a shorter sampling
period. In addition, collision avoidance constraints must be more ¦ne-grained.
However, as the OSTG phase will have reduced much of the radial and out-
of-plane separation between chaser and target, the e¨ect of linearization error
on the Yamanaka Ankersen [28] equations is no longer a problem, as long as a
cylindrical relative coordinate system is used [26]. This model is less complex
than the J2 -modi¦ed GVEs and allows objectives and constraints to be directly
speci¦ed in the cylindrical frame without requiring a linearized geometric trans-
formation (with inevitable loss of accuracy) from the relative orbital elements.
The prediction model input is assumed to be an impulsive –V in the cylindrical
target orbital frame.

4.1.3 Forced terminal translational guidance

The third phase, Forced Terminal Translational Guidance (FTTG), is tasked

with bringing the chaser from its ¦nal holding point at 100 m from the target

146
 GUIDANCE AND CONTROL

to a position 3 m from the target from where it can capture the target on a free
drift trajectory. Radial, in-track, and out-of-plane separation are small during
this phase. Control accuracy is critical due to the tight capture tolerances,
and a much higher sampling rate is required than for other phases. As for the
INTG phase, the Yamanaka Ankersen [28] equations are used for the trajectory
prediction model.
In addition, to maintain target pointing, the model predictive controller
must also handle attitude regulation to an externally provided setpoint, using
thrusters. A linearized quaternion-based prediction model [13] extended to con-
sider the elliptical orbital dynamics is used for the relative attitude control.
The attitude reference frame used for control is chosen depending on the direc-
tion of approach, and the attitude setpoint in the inertial frame to avoid the
predicted trajectory crossing the discontinuity at ±180◦ in the quaternion repre-
sentation [33]. Because the prediction matrices are rebuilt at each time step due
to the LTV prediction model, the opportunity is taken to relinearize the attitude
dynamics about the current measured attitude at each time step.

4.1.4 Collision avoidance manoeuvre

The CAM must safely move the chaser away from the target, to a distance
of 500 m within three orbits without collision with the target. Essentially, this
objective is similar to that of INTG, except traveling away from the target instead
of towards it, and with a less speci¦c terminal objective. It, therefore, makes
sense to use the Yamanaka Ankersen prediction model for this phase also.

4.2 Model Predictive Control Subsystem Design

Each of the model predictive controllers is designed independently, but with a

common interface and a common output function to convert the –V into ¦nite-
duration thrust pulses in the inertial frame. The core MPC function of each
control subsystem is implemented using the blocks from the MPCTOOL, with
the linear time-varying prediction models implemented as EML functions called
by the MPCTOOL blocks. Any additional logic or reference-frame changes are
implemented using Simulink blocks.

4.2.1 Orbit sinchronization translational guidance model predictive

controller

The OSTG model predictive controller must bring the chaser into the same or-
bit as the target in a timely manner, whilst minimizing propellant consumption.
Rather than using the more commonly used quadratic cost function, to correctly

147
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

Figure 3 The OSTG safety (a) and terminal (b) constraints

encode the minimization of total propellant consumption, the model predictive

controller must minimize the absolute sum of –V applied over the prediction
horizon [34]. Furthermore, to balance this with time to completion, a termi-
nal constraint enforcing the completion criteria is imposed at the end of the
prediction horizon, and the prediction horizon itself is included as a decision
variable in the cost function [6, 35, 36]. Letting N be the prediction horizon,
u = [u(t + Ts |t)T , · · · , u(t + (N − 1)Ts |t)T ]T and α be a parameter determining
constraints that will be described later, the cost function is:

#
N −1
JOSTG (α, u, N ) = N +  wu u(t + kTs |t) 1 .
k=1

Note that the summation is from k = 1 not k = 0, implying that the input
calculated at the current time step is applied at the next time step to allow
su©cient time duration for computation to occur. The terminal constraint, which
will be described later, ensures that the predicted trajectory ends in the correct
orbit, with an acceptable separation from the target.
In order the predicted trajectories do not collide with the target, constraints
are placed on the predicted trajectories to ensure that they do not enter a safety
sphere of radius Rs (t), surrounding the target. In addition, as proposed in [10],
unforced drift trajectories emanating from each point in the prediction horizon
are also constrained to ensure passive safety. Collision avoidance is a manifestly
nonconvex constraint, but it is approximated by a half-space constraint with
angle relative to the in-track direction parameterized by α (Fig. 3). The value
of α then determines on which side of the target the terminal constraint places
the end of the predicted trajectory.
The optimization is implemented using the ¤LP-based LTV model predictive
controller¥ block from MPCTOOL, which allows prediction horizon N and user-
de¦ned parameters to be passed in as signals. Given an angle α0 calculated as

148
 GUIDANCE AND CONTROL

the current angle between the chaser and the zcrf axis, rounded to the nearest
45◦ , by solving 3N convex optimizations, varying N between 1 and Nmax , for
each α ∈ {α0 − 45◦ , α0 , α0 + 45◦ } using two nested Simulink ¤For-iterator¥
subsystems, the control sequence can be found that minimizes the overall cost
function. A sampling period TS = 600 s was chosen, along with a maximum
prediction horizon Nmax = 25.

4.2.2 Impulsive nominal translational guidance model predictive

controller

The INTG model predictive controller must transfer the chaser between a se-
quence of invariant holding points on V (i. e., the in-track axis in the cylindrical
orbital frame) until a separation of 100 m is achieved. Because release from these
holding points must be governed by an external signal, there is no point predict-
ing further ahead than the end of a single transfer. It is su©cient to design a
controller to perform a transfer, parameterized by the distance from the target
of the next holding point.
The design is similar to that of the OSTG model predictive controller in
that a 1-norm cost function is used in conjunction with a variable horizon im-
plemented by solving multiple convex optimizations. However, the cost function
includes distance instead of time to re§ect that fuel consumption is proportional
to distance traveled rather than time when carrying out passively safe hopping
trajectories. The holding points are scheduled by an external algorithm and
parameterized by distance xhp . The cost function is

#
N −1
JINTG (u, N ) =  Ec (xcrf (t + kTs |t) − r(t + kTs )) 1 +  wu u(t + kTs |t) 1
k=1

T
where r(t + kTs ) = [±xhp (1 + etgt cos νtgt (t + kTs )), 0, 0, 0, 0, 0] depends on
the direction of approach; xcrf is the state vector in the cylindrical reference
frame; etgt is the eccentricity of the target orbit; νtgt is the true anomaly of the
target; and  
100000
Ec = .
010000
As for the OSTG model predictive controller, passive safety constraints are
imposed over a period of one orbit from each prediction in the control horizon.
In addition, to ensure passive safety over a longer period, an additional passive
drift constraint is imposed to make sure that long-term secular drift is away from
the target (thus avoiding collision in subsequent orbits). Letting Aorb (νtgt ) be
the propagation matrix for a whole orbit,

[−sgn (xcrf (t)) 0 0 0 0 0] (Aorb (νtgt (t + kTs ))) xcrf (t + kTs |t) ≤ 0 .

149
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

The terminal set for the INTG model predicted controller is de¦ned as a box
with side-length 2(etgt +0.1), centered on a point xhp away from the target on V ,
with an additional constraint that the chaser should be on a periodic trajectory
and also be inside the box after 1 /4 , 1 /2 , and 3 /4 orbits of free drift. A sampling
period Ts = 300 s and a maximum prediction horizon of Nmax = 20 were chosen.

4.2.3 Forced terminal translational guidance model predictive

controller

During the FTTG phase, trajectory and attitude tracking accuracy becomes
more important than long-term fuel minimization. The navigation uncertainty is
of a similar order of magnitude to the expected tracking errors; so, a conventional
quadratic cost function is appropriate. The controller is implemented using the
¤QP-based LTV model predictive controller¥ block from MPCTOOL, with a
sampling period Ts = 3 s and a prediction horizon N = 15. Letting x(j|t) be
the combined position, velocity, attitude quaternion, and angular velocity states,
r(j) the corresponding reference setpoint, and u(j|t) the vector of thruster inputs,
the cost function is:

#
N −1
JFTTG = (x(t + kTs |t) − r(t + kTs ))T Q(x(t + kTs |t) − r(t + kTs ))
k=1
+ –u(t + kTs |t)T R–u(t + kTs |t) .

Changes in input (–u) are penalized instead of the absolute input value to
enable o¨setfree tracking of forced-equilibrium setpoints [37]. Positivity and
saturation constraints are applied to inputs. The reference trajectory r(j) and
cost function weightings Q ≥ 0 and R ≥ 0 are chosen so that the controller
tracks an attitude setpoint, a position in the radial and out-of-plane directions,
and an approach velocity in the in-track direction.

4.2.4 Collision avoidance manoeuvre model predictive controller

The CAM model predictive controller is based on a modi¦ed version of the INTG
model predictive controller. However, in order for rapid response, a delay of Ts
is not assumed in the model. Instead, it is assumed that the calculation of the
control move will complete as fast as possible. To facilitate the fast computation,
a variable horizon is not used for CAM, and the trajectory is constrained so
that only one impulsive –V may be applied at the beginning of the prediction
horizon. This is applied open-loop on the assumption that navigation error
may increase outside nominal operational levels following the fault triggering
the CAM, especially if attitude pointing is lost. The terminal constraint is

150
 GUIDANCE AND CONTROL

chosen so that under the speci¦ed worst-case navigation error, the chaser will
be further than 500 m away from the target in three orbits. The INTG model
predictive controller can then hold the chaser in a periodic §y-around orbit or
restart approach via its sequence of holding points once the CAM is complete.

5 THE ONLINE RECONFIGURATION CONTROL

SYSTEM AND AVIONICS ARCHITECTURE

The avionic architecture considered in HARVD is based on the ¤Aurora Avionics

Architecture¥ ESA study, which is the avionics reference for future exploration
vehicles. From this starting point, the ORCSAT study includes the design of an
Avionic Architecture System allowing the implementation of embedded MPC-
based systems. The MPC concept is based on the optimization of a cost function
under some constraints, which usually is carried out using quite complex iter-
ative algorithms, requiring high computational capability. Therefore, the main
challenge of the avionic architecture design is to de¦ne a Central Data Man-
agement Unit (CDMU) able to cope with the MPC needs. In particular, the
selection of the Central Processing Unit (CPU) is the key for the MPC embed-
ded implementation.
Since the beginning of the design, it was evident that a CPU composed of a
processor and a coprocessor has to be considered as baseline (distributed archi-
tecture) taking into account available space quali¦ed processor computational
performances (Fig. 4). This solution allows a distribution of the complete on-
board SW on the two processors leaving to the coprocessor the execution of
GNC algorithms requiring signi¦cant computational throughput (MPC) and to
the processor the handling of the system units, the other parts of the GNC, etc.
The next step was the selection of the processors. Currently space-quali¦ed
processors are based on LEON2 FT, with performance of 86 MIPS, 23 MFLOPS
at 100 MHz: taking into account also the HARVD experience, it has been con-
sidered adequate for the central processor of the CDMU.
Regarding the coprocessor, the MPC computational throughput is the driver
for the selection. To support this task, pro¦ling of the MPC algorithm was
performed using Simulink features, to evaluate the time needed for the execution
of the MPC algorithms. Afterwards, these timing values have been scaled to
the selected processor exploiting the Whetstone benchmark. The processors
selected for the trade-o¨ are the LEON2 FT and PowerPC750FX, able to perform
1650MIPS at 733 MHz which has been used to design space-quali¦ed boards like
the Maxwell SCS750.
The pro¦ling results are summarized in Table 2. It can be seen that the
LEON2 FT cannot be selected as coprocessor, since the FTTG would take more
than 4 s for the computation of the control action against a theoretical con-

151
152
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

Figure 4 Central data management unit with distributed architecture

 GUIDANCE AND CONTROL

Table 2 Model predictive control pro¦ling

Maximal time, s Time, s
Mode
Workstation Scaling to LEON2 FT Scaling to PowerPC 750FX
OSTG 7.8575 410.9473 23.7401
INTG 1.9028 99.5164 5.7490
FTTG 0.0885 4.6286 0.2674
CAM 0.1249 6.5323 0.3774

Figure 5 The HARVD (1) vs. MPC (2) performance comparisons

trol step of 3 s. Instead, the PowerPC 750FX shows timings which are widely
compatible with the MPC design and expected computational capabilities and,
therefore, it has been selected as baseline for the CPU coprocessor.

6 SIMULATION RESULTS AND COMPARISONS

Figure 5 shows the comparison between the simulation results obtained with
the HARVD GNC solution and the ones obtained with the MPC in the case
of rendezvous circular orbit. Di¨erences are visible since the beginning of the
rendezvous, where the MPC trajectory remains closer to the target with respect
to HARVD, but the most signi¦cant result is the propellant save, which in this
case is about 35 kg.
The main di¨erences can be found analyzing the trajectory during the OSTG.
In this phase, the MPC design is such that the chaser is left at a relative dis-
tance with respect to the target between 5 and 30 km: with the ¦nal MPC
tuning, it has been noted that the chaser is left at the end of OSTG usually at
15 km from the target. The latter ¦nding suggested a di¨erent de¦nition of the
holding points, which in the HARVD initial solution started from 50 km: there-

153
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

Figure 6 The HARVD (1) vs. MPC (2) performance comparisons with new holding
points de¦nition

Figure 7 The CAM simulation results: 1 ¡ §y-around and 2 ¡ CAM

fore, the HARVD simulation has been repeated with the ¦rst holding point at
20 km. Figure 6 shows that HARVD performance improved a lot, in particular,
on the propellant consumption: in this case, the di¨erence is reduced to about
10 kg.
This aspect is quite important: the online optimization performed by MPC
has permitted the detection of a possible improvement in the nominal mission
scenario (i. e., de¦nition of the holding points) that would be di©cult to clearly
identify a priori.
Figure 7 shows the trajectories obtained in three simulations where the CAM
is triggered at di¨erent relative distances from the target and the following §y-
around. Performances are very good, since MPC allows avoiding the collision
also at very short distance (10 m) with a single manoeuvre.

154
 GUIDANCE AND CONTROL

Figure 8 Overall trajectories during OSTG and INTG of 50 cases of the Monte-Carlo
campaign: 1 ¡ initial position; 2 ¡ no control; 3 ¡ OSTG; and 4 ¡ INTG.

Figure 9 Capture performance of 400 test cases of the Monte-Carlo campaign: 1 ¡

20-centimeter requirement; and 2 ¡ target center

The MPC solution has been also validated and veri¦ed by means of a Monte-
Carlo simulation campaign composed by 800 test cases, in order to test the
performance of the control in di¨erent scenarios (circular and elliptic orbit) and
starting from di¨erent initial relative positions and dynamics with respect to the
target. The obtained results are very good, since the capture has been always
achieved with margins.
Figure 8 shows a typical result of this campaign, summarizing 50 test cases
trajectories during OSTG and INTG. Instead, Fig. 9 shows the aggregated cap-
ture accuracy results of 400 cases.

155
PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

7 CONCLUDING REMARKS

Optimization-based control techniques like MPC are considered extremely at-

tractive for applications which require high level of autonomy, optimal path
planning, and dynamic safety margins. The ORCSAT study is addressing the
usage of the MPC techniques on the rendezvous and capture scenarios of the
MSR mission. The results obtained after the design phase are encouraging,
since with respect to classical control techniques (HARVD), it is possible to have
a signi¦cant improvement, in particular, in the propellant consumption. As side
e¨ect, but not less important, online optimization could drive the de¦nition of
higher level mission aspects that could not be easy to address in the earlier phase
of GNC design.
The MPC design has been veri¦ed and validated throughout a wide Monte-
Carlo simulation campaign which considers plant mismatch, sensors and actua-
tors failures, and di¨erent initial dynamic conditions. Obtained results con¦rm
the robustness of the design and the very good performance.
The next step of the ORCSAT study will be the implementation of the MPC
algorithms in the selected avionic architecture, with the objective to test the RT
performance of the developed solution on a §ight-representative avionic. In the
end, the GMV Dynamic Test Bench will be enhanced with the model predictive
based control and the selected avionics for the ¦nal dynamic test campaign.

REFERENCES

1. Maciejowski, J. M. 2002. Predictive control with constraints. Pearson Education.

2. Camacho, E. F., and C. Bordons. 2004. Model predictive control. London: Springer-
Verlag.
3. Rawlings, J. B., and D. Q. Mayne. 2009. Model predictive control: Theory and
design. Nob Hill Publishing.
4. Qin, S. J., and T. A. Badgwell. 2003. A survey of industrial model predictive control
technology. Control Eng. Practice 11(7):733 64.
5. Shim, D. H., H. J. Kim, and S. Sastry. 2003. Decentralized nonlinear model pre-
dictive control of multiple §ying robots. 42nd IEEE Conference on Decision and
Control Proceedings. 4:3621 26. Maui, Hawaii, USA.
6. Richards, A., and J. P. How. 2006. Robust variable horizon model predictive control
for vehicle maneuvering. Int. J. Robust Nonlinear Control 16(7):333 51.
7. Almeida, F. A. 2008. Waypoint navigation using constrained in¦nite horizon model
predictive control. AIAA Guidance, Navigation and Control Conference and Ex-
hibit Proceedings. Honolulu, Hawaii.
8. Larsson, R., S. Berge, P. Bodin, and U. J onsson. 2006. Fuel e©cient relative or-
bit control strategies for formation §ying and rendezvous within PRISMA. 29th
Annual AAS Guidance and Control Conference Proceedings.

156
 GUIDANCE AND CONTROL

9. Breger, L., and J. P. How. 2007. Gauss£s variational equation-based dynamics and
control for formation §ying spacecraft. J. Guidance Control Dyn. 30(2):437 48.
10. Breger, L., and J. P. How. 2008. Safe trajectories for autonomous rendezvous of
spacecraft. J. Guidance Control Dyn. 31(5):1478 89.
11. Bodin, P., R. Noteborn, R. Larsson, and C. Chasset. 2011. System test results
from the GNC experiments on the PRISMA in-orbit test bed. Acta Astronautica
68(7-8):862 72.
12. Manikonda, V., P. O. Arambel, M. Gopinathan, R. K. Mehra, and F. Y. Hadaegh.
1999. A model predictive control-based approach for spacecraft formation keeping
and attitude control. American Control Conference Proceedings. San Diego, CA.
6:4258 62.
13. Hegrenˆs, O., J. T. Gravdahl, and P. Tondel. 2005. Spacecraft attitude control
using explicit model predictive control. Automatica 41(12):2107 14.
14. Wood, M., and W. H. Chen. 2008. Model predictive control of low Earth orbiting
satellites using magnetic actuation. Proc. of the Institution of Mechanical Engi-
neers, Part I: J. Syst. Control Eng. 222(6):619 31.
15. Colmenarejo, P., L. Tarabini, C. Le Peuv‚edic, and A. Guiotto. 2008. HARVD devel-
opment, veri¦cation and validation approach (from traditional GNC design/V&V
framework simulator to real-time dynamic testing). 7th ESA Conference (Inter-
national) on Guidance, Navigation and Control Systems. Tralee, County Kerry,
Ireland.
16. Barrena, V., P. Colmenarejo-Matellano, D. Modrego-Contreras, C. Le Peuv‚edic,
and A. Guiotto. 2008. Integrated development, veri¦cation and validation approach
for space systems using autocoding techniques. Data System in Aerospace Confer-
ence (DASIA 2008). Palma Majorca, Spain.
17. Strippoli, L., P. Colmenarejo, T. V. Peters, C. Le Peuv‚edic, and T. Voirin. 2010.
High integrity control system for generic autonomous RVD. 61st Astronautical
Congress (Internatioanl). Prague, CZ.
18. Bemporad, A., M. Morari, and N. L. Ricker. 2009. Model predictive control
ToolboxTM 3 ¡ user£s guide. The Mathworks, Inc. http://www.mathworks.com/
access/helpdesk/help/toolbox/mpc/.
19. Bemporad, A., M. Morari, V. Dua, and E. N. Pistikopoulos. 2002. The explicit
linear quadratic regulator for constrained systems. Automatica 38(1):3 20.
20. Bemporad, A. 2009. Hybrid Toolbox v1.2.2 ¡ user£s guide. Dec. http://
www.ing.unitn.it/∼bemporad/hybrid/toolbox.
21. Bertsimas, D., and J. N. Tsitsiklis.1997. Introduction to linear optimization. Athena
Scienti¦c.
22. Clohessy, W. H., and R. S. Wiltshire. 1960. Terminal guidance system for satellite
rendezvous. J. Aerospace Sci. 27(9):653 58.
23. Tschauner, J. 1967. Elliptical orbit rendezvous. AIAA J. 5(6):1110 13.
24. Carter, T. E. 1998. State transition matrices for terminal rendezvous studies: Brief
survey and new example. J. Guidance Control Dyn. 21(1):148 55.
25. Battin, R. H. 1999. An introduction to the mathematics and methods of astrody-
namics. Revised edn. AIAA education ser. American Institute of Aeronautics and
Astronautics.

157
 PROGRESS IN FLIGHT DYNAMICS, GNC, AND AVIONICS

26. Melton, R. G. 2000. Time-explicit representation of relative motion between ellip-

tical orbits. J. Guidance Control Dyn. 23(4):604 10.
27. Schaub, H., S. R. Vadali, J. L. Junkins, and K. T. Alfriend. 2000. Spacecraft forma-
tion §ying control using mean orbital elements. J. Astronautical Sci. 48:69 87.
28. Yamanaka, K., and F. Ankersen. 2002. New state transition matrix for relative
motion on an arbitrary elliptical orbit. J. Guidance Control Dyn. 25(1):60 66.
29. Gim, D., and K. T. Alfriend. 2003. State transition matrix of relative motion for
the perturbed noncircular reference orbit. J. Guidance Control Dyn. 26(6):956 71.
30. Sidi, M. J. 1997. Spacecraft dynamics and control: A practical engineering ap-
proach. Cambridge University Press.
31. Kerambrun, S., N. Despr‚e, B. Frapard, P. Hyounet, B. Polle, M. Ganet, N. Silva,
A. Cropp, and C. Philippe. 2008. Autonomous rendezvous system: The HARVD
solution. 7th ESA Conference (International) on Guidance, Navigation and Control
Systems Proceedings. Tralee, Ireland.
32. Le Peuv‚edic, C., P. Colmenarejo, and A. Guiotto. 2008. Integrated multi-range
RDV control system ¡ autonomous RDV GNC test facility ¡ HARVD control
system trade-o¨ analysis and baseline solution. Technical Report GMV-HARVD-
TN06. GMV.
33. Bach, R., and R. Paielli. 1993. Linearization of attitude-control error dynamics.
IEEE Trans. Automatic Control 38(10):1521 25.
34. Tillerson, M., G. Inalhan, and J. P. How. 2002. Co-ordination and control of dis-
tributed spacecraft systems using convex optimization techniques. Int. J. Robust
Nonlinear Control 12(2-3):207 42.
35. Richards, A., and J. How. 2003. Performance evaluation of rendezvous using model
predictive control. AIAA Guidance, Navigation and Control Conference and Ex-
hibit. Austin, Texas.
36. Richards, A. G., and J. P. How. 2003. Model predictive control of vehicle maneuvers
with guaranteed completion time and robust feasibility. 2003 American Control
Conference Proceedings. Denver, Colorado. 5:4034 40.
37. Maciejowski, J. M. 1998. The implicit daisy-chaining property of constrained pre-
dictive control. Appl. Math. Computer Sci. 8(4):101 17.

158

View publication stats