2104 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO.

6, NOVEMBER 2013

Lossless Convexification of Nonconvex Control

Bound and Pointing Constraints of the Soft
Landing Optimal Control Problem
Behçet Açıkmeşe, John M. Carson III, and Lars Blackmore

Abstract— Planetary soft landing is one of the benchmark “soft lands.” The problem of designing the fuel-optimal thrust
problems of optimal control theory and is gaining renewed profile as a function of time, which delivers the vehicle as close
interest due to the increased focus on the exploration of planets as possible to a prescribed target under constraints on the thrust
in the solar system, such as Mars. The soft landing problem
with all relevant constraints can be posed as a finite-horizon and the state of the landing vehicle, is a finite-horizon optimal
optimal control problem with state and control constraints. control problem, and it is referred to as the “soft landing”
The real-time generation of fuel-optimal paths to a prescribed problem. In this paper, we present an algorithm, which is
location on a planet’s surface is a challenging problem due to the referred to as the powered descent guidance algorithm, to
constraints on the fuel, the control inputs, and the states. The compute optimal solutions of the soft landing problem based
main difficulty in solving this constrained problem is the existence
of nonconvex constraints on the control input, which are due to on a lossless convexification of the problem.
a nonzero lower bound on the control input magnitude and a Powered descent guidance algorithms minimize the land-
nonconvex constraint on its direction. This paper introduces a ing error by simultaneously satisfying the constraints such
convexification of the control constraints that is proven to be as the governing physics, thrust bounds, and position and
lossless; i.e., an optimal solution of the soft landing problem can speed constraints. Additionally, the short duration of planetary
be obtained via solution of the proposed convex relaxation of the
problem. The lossless convexification enables the use of interior powered descent requires that the guidance algorithms be
point methods of convex optimization to obtain optimal solutions executed quickly on board in real time and that they guarantee
of the original nonconvex optimal control problem. finding a solution when one exists. However, the soft-landing
Index Terms— Convex optimization, convexification, interior powered-descent guidance problem is a nonconvex finite-
point method algorithms, optimal control, planetary soft landing. horizon optimal control problem in its original form because
of the control constraints. The descent thrusters cannot be
I. I NTRODUCTION throttled off after ignition, so the guidance algorithm must
generate valid thrust vectors with a nonzero minimum and

P LANETARY soft landing is one of the benchmark prob-

lems of optimal control theory and is gaining renewed
interest due to the increased focus on the exploration of planets
maximum on the thrust magnitude. This is a nonconvex
constraint on the control input. Further, onboard sensors for
terrain-relative navigation generally require specific viewing
in the solar system. The main focus of the missions in the orientations, which constrain the allowable spacecraft orienta-
near future is to soft-land precisely at scientifically interesting tions and thus the thrust vector pointing direction, which is an
locations, which is also referred to as precision or pinpoint additional source of nonconvexity. The main contribution of
landing. The soft landing is the final phase of a planetary this paper is to formulate a convex relaxation of the original
EDL (entry, descent, and landing), and it is also referred to optimal control problem and show that an optimal solution of
as the powered descent landing stage. In the powered descent the relaxed problem is also optimal for the original one. We
stage, the vehicle is guided as close as possible to a prescribed refer to this as a lossless convexification of the original control
location on the planet’s surface by using thrusters that provide problem, since no part of the feasible region is removed and
the control authority. This phase concludes when the vehicle the optimal solutions of the relaxed problem define optimal
lands with zero velocity relative to the surface, i.e., when it solutions to the original problem. Note that many nonconvex
Manuscript received January 16, 2012; accepted December 19, 2012. optimal control problems can be convexified, but guaranteeing
Manuscript received in final form December 27, 2012. Date of publication that the convexification is lossless is not always possible. More
February 1, 2013; date of current version October 15, 2013. Recommended particularly, one can convexify the problem in two obvious
by Associate Editor N. Hovakimyan.
B. Açıkmeşe is with the Department Aerospace Engineering and Engineer- ways: 1) restrict the set of feasible solutions to a convex subset
ing Mechanics, The University of Texas, Austin, TX 78712 USA (e-mail: of the feasible set and 2) relax the set of feasible solutions to
behcet@austin.utexas.edu). a convex set containing the original set of feasible solutions.
J. M. Carson is with the Jet Propulsion Laboratory, California Institute of
Technology, Pasadena, CA 91109 USA (e-mail: jmcarson@jpl.nasa.gov). The first approach produces a feasible solution of the original
L. Blackmore is with SpaceX Company, Hawthorne, CA 90250 USA problem with an upper bound on the actual optimal cost, and
(e-mail: lars.blackmore@spacex.com). the second approach generates a lower bound on the optimal
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. cost but, in general, does not necessarily provide a feasible
Digital Object Identifier 10.1109/TCST.2012.2237346 solution of the original problem. Hence the first approach has
1063-6536 © 2013 IEEE
AÇIKMEŞE et al.: CONTROL BOUND AND CONSTRAINTS OF OPTIMAL CONTROL 2105

a loss in the optimal cost achievable and the second approach at the extreme points of a projection of the relaxed set of
has a loss since it does not necessarily provide a feasible feasible controls. This set is then shown to be contained in
solution. Our convexification follows the second approach, the original nonconvex set of feasible controls, thereby estab-
but we guarantee a feasible and, hence, optimal solution of lishing that we can obtain optimal solutions of the original
the original problem. Therefore, our convexification of control nonconvex problem via solving its convex relaxation.
constraints is lossless. Another source of nonconvexity is The theoretical development of lossless convexification for
having time-varying mass in the dynamics, which is resolved the soft landing problem allows the application of IPMs of
via a change of variables as in [1]. convex optimization [10]–[12], [19] that can solve these prob-
A number of prior works have proposed solutions to variants lems reliably with polynomial-time convergence guarantees.
of the powered descent guidance problem, including [1]–[8], Further, a surge of interest in the area of fast real-time
and [9]. In [3], a 1-D version of the soft landing problem is convex optimization [20]–[22] has demonstrated computa-
solved analytically in a closed form. However, this solution tional speedups of several orders of magnitude for IPMs. The
cannot be extended to the 3-D problem with state or con- advancements in this area will dramatically enhance the real-
trol constraints. One existing solution method is our convex time computational efficiency of IPMs, thereby enabling their
optimization approach [1], [2], which poses the problem of use for planetary soft landing.
minimum-fuel powered descent guidance as a second-order
cone program (SOCP). This optimization problem can be A. Partial List of Notations
solved in polynomial time using existing interior point method
R is the set of real numbers; a condition is said to hold
(IPM) algorithms that have a deterministic stopping criterion
almost everywhere in the interval [a, b], a.e. [a, b] if the set
given a prescribed level of accuracy. That is, the global opti-
of points in [a, b] where this condition fails to hold is in a set
mum can be computed to any given accuracy with an a priori
of measure zero; Rn is the n-dimensional real vector space;
known upper bound, i.e., a polynomial function of the problem
v is the 2-norm of the vector v; 0 is a matrix of zeros;
size, on the number of iterations required for convergence. In
I is the identity matrix; ei is a column vector with its i th
addition, IPM algorithms of SOCPs are guaranteed to find a
entry 1 and other entries zero; (v 1 , v 2 , . . . , v m ) represents a
feasible solution if one exists [10]–[13]. This is in contrast
vector obtained by augmenting vectors v 1 , . . . , v m such that
with other approaches that either compute a closed-form  
T T ; ∂S denotes the set of
solution by ignoring the constraints of the problem [5], [14], (v 1 , v 2 , . . . , v m ) := v 1T v 2T , . . . , sv m
propose solving a nonlinear optimization on board [6], [7], or boundary points and int S denotes the interior of the set S.
solve a related problem that does not minimize fuel use [8].
The closed-form solution approaches result in solutions that do II. P LANETARY L ANDING W ITH T HRUST P OINTING
not obey the constraints inherent in the problem, such as no C ONSTRAINTS
subsurface flight constraints. This means that constraints must The planetary soft landing problem searches for the thrust
be checked explicitly after a solution is generated, and any (control) profile Tc and an accompanying translational state
solution that violates the constraints is rejected. In practice, trajectory (r, ṙ) which guide a lander from an initial position
this reduces the size of the region from which return to r0 and velocity ṙ0 to a state of rest at the prescribed target
the target is possible by a factor of five or more [15], and location on the planet while minimizing the fuel consumption.
the proposed method is numerically robust as also observed The problem considers planets with a constant rotation rate
independently [16]. Nonlinear optimization approaches, on (angular velocity), a uniform gravity field, and negligible aero-
the other hand, cannot provide a priori guarantees on how dynamic forces during the powered-descent phase of landing.
many iterations will be required to find a feasible trajectory, When the target point is unreachable from a given initial
and are not guaranteed to find the global optimum, which state, a precision landing problem (or minimum landing error
limits their onboard applicability. For a comparison of the problem) is considered instead, with the objective to first find
convex optimization approach to alternative approaches, see the closest reachable surface location to the target and second
[9] and [15]. to obtain the minimum fuel state trajectory to that closest
In this paper, we unify the convex optimization approaches point. We formulate a prioritized optimization approach that
of [1], [2], [17] and extend them to handle thrust pointing handles both problems under a unified framework, which is
constraints. While convexifying the problem with nonconvex then referred to as the planetary soft landing problem.
thrust pointing constraints, we develop a geometrical insight In this problem, there are several state and control con-
into the problem that establishes a connection with “normal straints. The main state constraints are the glide slope con-
systems” [18]. A normal linear system is defined in the context straint on the position vector and an upper bound constraint
of optimal control theory where the system is said to be normal on the velocity vector magnitude. The glide slope constraint
with respect a set of feasible controls if it maximizes the is described in Fig. 1 and is imposed to ensure that the
Hamiltonian at a unique point of the set of feasible controls. In lander stays at a safe distance from the ground until it
the case when the set of feasible controls is convex, a system reaches its target. The upper bound on velocity is needed to
being normal implies that the Hamiltonian is maximized at an avoid supersonic velocities for planets with atmosphere, where
extreme point of the set [18]. Our convexification result has the control thrusters can become unreliable. Both of these
a similar geometric interpretation since it establishes lossless constraints are convex and fit well to the convex optimization
convexification by ensuring that the Hamiltonian is maximized framework considered in this paper. The control constraints,
2106 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013

Initial described by
Velocity  
Initial Position Tc (t)
ẋ(t) = A(ω) x(t) + B g +
Glideslope m(t)
Constraint Optimal ṁ(t) = −αTc (t) (1)
Trajectory
where x(t) = (r(t), ṙ(t)) : R+ → R6 ,
m(t) : R+ → R+ is
X the mass of the lander
Z  
Landing Target 0 I
A(ω) =
Origin −S(ω)2 −2S(ω)
 
0
B =
Y I
⎡ ⎤
0 −ω3 ω2
Fig. 1. Glide slope constraint in the minimum landing error powered descent
guidance problem. This constraint requires the spacecraft to remain in a cone S(ω) = ⎣ ω3 0 −ω1 ⎦ (2)
defined by the minimum slope angle γ. In the minimum landing error case, −ω2 ω1 0
the apex of the cone coincides with the landed position of the spacecraft,
rather than the original target. ω = (ω1 , ω2 , ω3 ) ∈ R3 is the vector of planet’s constant
angular velocity, g ∈ R3 is the constant gravity vector, and
Pointing
Envelope
Pointing
Envelope
α > 0 is a constant that describes the fuel consumption (mass
Intersection
depletion) rate. Here, the time derivatives of the vectorial
quantities are expressed in a planet surface fixed frame that
has the planet’s angular rotation rate and we also used the
rocket equation which relates the fuel mass consumption rate
to the applied thrust vector [24].
Intersection We use a lumped mass rigid body model of the landing
(a) (b) (c)
vehicle, where the translational dynamics are decoupled from
Fig. 2. (a) Planar representation of original thrust bounds, intersec- rotational (attitude) dynamics. This is a common assumption
tion of thrust bounds and thrust pointing limits: (b) θ ∈ (π/2, π ) and used in practice mainly because the attitude control authority is
(c) θ ∈ [0, π/2]. typically of far higher bandwidth than that of the translational
one. Specifically, any attitude maneuver required to point the
thruster in the right direction for translational control can
however, are challenging since they define a nonconvex set be done very quickly such that the interaction between the
of feasible controls. We have three control constraints (see attitude and translational control systems are minimal. As a
Fig. 2). Given any maneuver time (time of flight) t f , for all result, this is a reasonable assumption that reduces the problem
t ∈ [0, t f ]. complexity considerably.
1) Convex upper bound on thrust, Tc (t) ≤ ρ2 . Given the constraints, the dynamics, and a target location
2) Nonconvex lower bound on thrust, Tc (t) ≥ ρ1 > 0. on the surface (0, q) where q ∈ R2 denotes the coordinates of
3) Thrust pointing constraint n̂T Tc (t)/Tc (t) ≥ cos θ the target at zero altitude, the planetary soft landing problem
where n̂ = 1 is a direction vector and 0 ≤ θ ≤ π can be formulated as a prioritized optimization problem as
is the maximum allowable angle of deviation from the follows.
direction given by n̂, which is convex when θ ≤ π/2 Problem 1 (Nonconvex Minimum Landing Error Problem):
and nonconvex when θ > π/2.
min E r(t f ) − q (3)
Onboard sensors for terrain-relative navigation generally t f ,Tc
 ⎫
require specific viewing orientations, which imposes a con- Tc (t) ⎬
s.t. ẋ(t) = A(ω)x(t)+ B g +
straint on the vehicle orientation (attitude). Since we model m ∀t ∈ [0, t f ] (4)
⎭
the vehicle as a point mass with a thrust vector, the required ṁ(t) = −αTc (t)
control force is applied by pointing the thrust vector along x(t) ∈ X ∀ t ∈ [0, t f ] (5)
the desired force direction. In this framework, we can impose
0 < ρ1 ≤ Tc (t) ≤ ρ2 , n̂T Tc (t) ≥ Tc (t) cos θ (6)
constraints on the vehicle orientation by simply restricting
the directions that the thrust vector can point to. This also m(0) = m 0 , m(t f ) ≥ m 0 − m f > 0 (7)
avoids incorporating the attitude dynamics of the vehicle into r(0) = r0 , ṙ(0) = ṙ0 (8)
the problem formulation, which would otherwise increase the
e1T r(t f ) = 0, ṙ(t f ) = 0. (9)
problem complexity significantly. Considering attitude dynam-
ics explicitly and imposing the pointing constraints directly, Problem 2 (Nonconvex Minimum Fuel Problem):
rather than on the thrust direction, can be a part of future  tf
research, which can benefit from the recent convexification max m(t f ) − m(0) = min αTc (t) dt s.t. (10)
t f ,Tc t f ,Tc (·) 0
results on the constrained attitude control [23].
As mentioned earlier, the lander is modeled as a lumped dynamics and constraints given by (4)–(9)
∗
mass with a thrust vector for control, and its dynamics are E r(t f ) − q ≤ dP1 − q. (11)
AÇIKMEŞE et al.: CONTROL BOUND AND CONSTRAINTS OF OPTIMAL CONTROL 2107

In Problem 2, dP1∗ = E r ∗ (t ) ∈ R2 is the final position for

P1 f
the optimal cost obtained by solving Problem 1, i.e., dP1∗ is the

closest reachable point on the surface to the target location q,

m f is the available fuel, m 0 is the initial mass of the lander,
and
   
e2T 010
E= = .
e3T 001

We use X to define the set of feasible positions and velocities

for the spacecraft, i.e., the glide slope constraint, and the
maximum allowable speed constraint given by Vmax
   
X = (r, ṙ) ∈ R6 :  ṙ ≤ Vmax ,  E r − r(t f ) 
  
−cT r − r(t f ) ≤ 0 (12) Fig. 3. The convex relaxation of the final position in Problem 2.

where c specifies a feasible cone with its vertex at r(t f )

e1 achieve both goals, let F1 represent all feasible solutions of
c , γgs ∈ (0, π/2). (13)
tan γgs Problem 2 with (11) replaced by (15); F2 is the corresponding
Here, γgs is the minimum glide slope angle, as illustrated set with (11) replaced by (16), and Fe is the set for Problem 2
in Fig. 1. The glide slope constraint (12) ensures that the as it is. Then it is straightforward to note that F1 ⊆ F2 ⊆ Fe .
trajectory to the target cannot be too shallow or go subsurface. Further, we have F2 = Fe . This follows from the fact that the
X is a convex set, and for completeness we give the standard inequality (11) does not rule out solutions with end positions
definition of the interior of X on the circle and we cannot get inside the circle. Now, since
F1 ⊆ F2 , the second option above will produce solutions that
intX  {x∈X : ∃ ε > 0 such that y∈X if x − y< ε}. (14) use minimal fuel trajectories and end as close as possible to
the target. So the best choice is to use the second option.
The boundary of X is given by ∂X  {x ∈ X : x ∈ /
But E r(t f ) − q = d ∗P1 − q is a nonconvex constraint
intX}. Equation (7) defines the initial mass of the lander and
on the end position. Since F2 = Fe , we can simply replace
ensures that no more fuel than available is used. Equation (8)
this nonconvex constraint with the convex inequality (11).
defines the initial position and velocity of the lander, while
Consequently, we achieve the two main objectives: 1) to
(9) constrains the final altitude and the velocity. The time of
prioritize two costs, the achievable distance and fuel, explicitly,
flight t f is an optimization variable and is not fixed a priori.
and 2) to convexify both problems so that they can be solved to
The solution of the soft landing problem is obtained by
global optimality in polynomial time via IPM algorithms. Note
solving Problems 1 and 2. The motivation for this two-step
that we achieve both goals via a systematic prioritization of
prioritization approach is quite intuitive. The primary goal
costs. This kind of prioritization cost in optimization problems
of planetary landing problem considered in this paper is to
is also referred to as lexicographic goal programming [25].
land a vehicle as close to a given target as possible, i.e., to
A key challenge in solving Problems 1 and 2 is the lower
minimize the landing error as in Problem 1. There can be
bound ρ1 > 0 on the thrust magnitude in (6), which means that
multiple optimal solutions for this problem, hence we have a
the set of allowable thrust values is nonconvex (see the first
second step where we find the minimum error solution that
illustration in Fig. 2). Furthermore, when ρ1 = 0, the thrust
consumes the least fuel, as in Problem 2.
bound constraint is convex; however, the control constraints
Remark 1: In Problem 2, the constraint on the final position
can still be nonconvex due to the thrust pointing constraint in
is given by inequality (11). This constraint could have also
(6), which is nonconvex for θ > π/2 (see the second and third
been one of the following choices
illustrations in Fig. 2). These nonconvex control constraints
∗
E r(t f ) = d P1 (15) prevent the direct use of convex optimization techniques in
E r(t f ) − q = d ∗P1 − q. (16) solving this problem. Additionally, the dynamics for mass con-
sumption ṁ(t) in (4) define a nonlinear differential equation,
And the third choice is the constraint given by (11), which which when discretized leads to nonlinear equality constraints
is a convex relaxation of the second choice by including the which are also nonconvex.
end positions strictly inside this circle (see Fig. 3). Clearly, The key result in [1] includes a relaxation to the nonconvex
we will never be able to compute a solution that has final thrust-bound constraints and an approach to handle the mass
position strictly inside the circle with this further relaxation consumption dynamics that provided a relaxed version of
because the radius of this circle is the minimum achievable Problem 2. The optimal solution of this relaxed problem
distance to the target. We have two motivations to use the was shown to be also an optimal solution of Problem 2.
third option: 1) to get the minimum fuel solution that has a However, the convexification result of [1] does not hold in
final position as close as possible to the desired target and 2) the presence of any thrust pointing constraints, including when
to have all constraints of Problem 2 convex. To see how we θ ∈ [0, π/2]. This paper extends the convexification of control
2108 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013

Problem 4 (Convex Relaxed Minimum Fuel Problem):

 tf
min
(t)dt subject to (5), (7), (8), (9), (17), (18),
t f ,Tc ,
0
(19), and
∗
E r(t f ) − q ≤ dP3 − q (20)
where ∗
dP3 is the final optimal position from Problem 3. Note
Convex
Intersection that the nonconvex thrust constraints in (6) for Problems 1 and
Pointing 2 have been replaced with convex constraints (18) and (19) in
Half-Space
Problems 3 and 4. In [1], we showed that constraint relaxation
(18) on the thrust bound allows the discrete-time form of
Problem 4 to be posed as a convex optimization problem; the
same holds true with the addition of the convex thrust-pointing
constraint (19). Hence we will not discuss the discretization
of this problem in this paper and refer the reader to [1].
Definition 1: Fe and F f denote the sets of feasible solu-
Fig. 4. Relaxed pointing constraint U1 := {(Tc ,
) : n̂T Tc ≥ cos θ
} tions of Problems 1 and 2, i.e., {t f , Tc , x, m} ∈ Fe if it satisfies
is convex, which is illustrated in the first row for a planer case. The set is
a half-space in the direction on the normal vector to the planes shown for
all the state (5), control (6), and fuel (7) constraints, dynamics
θ = 180° (left), θ = 90° (center), and θ = 0° (right). Since U2 := {(Tc ,
) : (4), and boundary constraints (8) and (9) of Problem 1, and
Tc  ≤
, ρ1 ≤
≤ ρ2 } is also convex, their intersection, U = U1 ∩ U2 , similarly for F f . Fe∗ and F ∗f are the corresponding set of
defines a convex set of feasible controls for the relaxed problem in Tc −

space. This is demonstrated in the second row of illustrations for a given θ .

optimal solutions. Fre and Fr f are the set of feasible solutions
{t f , Tc ,
, x, m} for Problems 3 and 4, with Fre ∗ and F ∗ , the
rf
sets of optimal solutions.
constraints to hold under thrust pointing constraints as well. Lemma 1: The following hold:
We also prove that the convexification still holds with the 1) F f ⊆ Fe , Fr f ⊆ Fre , F ∗f ⊆ Fe∗ , and Fr∗f ⊆ Fre ∗;

planet’s rotation accounted for. 2) if {t f , Tc ,

, x, m} ∈ Fr f such that Tc (t) =
(t)
In the relaxed problem, we have the following control ∀[0, t f ] then {t f , Tc , x, m} ∈ F f ;
constraints: 3) if {t ∗f , Tc∗ ,
∗ , x ∗ , m ∗ } ∈ Fr∗f such that Tc (t)∗  =

∗ (t) a.e. [0, t ∗f ] then {t ∗f , Tc∗ , x ∗ , m ∗ } ∈ F ∗f .
1) convex upper bound on thrust, Tc (t) ≤
(t);
Proof: The first two relationships in 1) are straightforward
2) convex thrust pointing constraint n̂T Tc (t) ≥
(t) cos θ ;
to prove. The next two relationships follow from the first two.
3) convex bounds on the slack variable
, ρ1 ≤
(t) ≤ ρ2 .
Consider a subset of Fr f defined as F̄r f composed of all
The relaxed pointing constraint forms a half-space of valid solutions of Fr f such that
(t) = Tc (t), ∀t ∈ [0, t f ]. Since
thrust values in Tc -
space, with pointing in the direction {t f , Tc ,
, x, m} ∈ F̄r f , where Tc (t) =
(t), satisfies all the
of
 the outward facing normal to the half space, given by dynamics and constraints of Problem 2, {t f , Tc , x, m} ∈ F f .
n̂, − cos θ , which comes directly from the relaxed convex This proves 2). Consequently, since the cost functions are
pointing-constraint inequality above. The half-space constraint identical for Problem 4 with Tc  =
and Problem 2, the
is illustrated for several pointing angles (θ = {180, 90, and0°}) the last statement of the theorem follows.
in Fig. 4, which uses a planar representation of thrust (i.e., To make use of Lemma 1, we need to compute an optimal
2-D thrust) with a pointing vector n̂ along the Tc(1) axis. It solution of Problem 4 and check whether
(t) = Tc (t) ∀t ∈
shows that the relaxed set of constraints is convex. [0, t ∗f ] or not. But we do not know ahead of time, before
numerically computing the solution, whether this condition
III. L OSSLESS C ONVEXIFICATION will be satisfied or not. The following theorem establishes
that this condition will indeed be satisfied in general, with
We propose the following relaxed minimum-error and the caveat that the problem may be required to be modified
fuel-optimal control problems that are convex relaxations of slightly with the introduction of  and ω̂. It also provides a
Problems 1 and 2. generalization of the earlier results in [1] and [2] to handle
Problem 3 (Convex Relaxed Minimum Landing Error thrust pointing constraints as well as nonzero lower bound on
Problem): the thrust magnitude. Hence it establishes a lossless convex-
ification of the control constraints in the minimum-fuel soft
min E r(t f ) − q subject to (5), (7), (8), (9), and landing problem, Problem 2, and hence Problem 1.
t f ,Tc ,

 ⎫ Theorem 1: Consider Problem 4 with ω replaced by ω̂,

Tc (t) ⎬
ẋ(t) = A(ω)x(t) + B g + which is defined as follows:
m ∀ t ∈ [0, t f ] (17) ⎧
⎭ ⎨ ω if S(ω)n̂ = 0, NT S(ω)2 n̂ = 0
ṁ(t) = −α
(t)
where ω̂ := ω +  n̂⊥ if S(ω)n̂ = 0
Tc (t) ≤
(t), 0 < ρ1 ≤
(t) ≤ ρ2 (18) ⎩
ω +  n̂ if S(ω)n̂ = 0, NT S(ω)2 n̂ = 0
n̂ Tc (t) ≥ cos θ
(t)
T
(19) (21)
AÇIKMEŞE et al.: CONTROL BOUND AND CONSTRAINTS OF OPTIMAL CONTROL 2109

where n̂⊥ is a unit vector such that n̂T n̂⊥ = 0, N ∈ R3×2 by (23), y(t) = 0, ∀t ∈ [0, t ∗f ], or y(t) = 0 occurs at a
has its columns spanning the null space of n̂T , and  > 0 is a countable number of points in [0, t ∗f ], which follows from the
(arbitrarily small) real number. Let {t ∗f , Tc∗ ,
∗ , x ∗ , m ∗ } ∈ Fr∗f first conclusion of Lemma 2. Suppose y(t) = 0 ∀t ∈ [0, t ∗f ].
such that the corresponding state trajectory x ∗ (t) ∈ intX ∀t ∈ Note that the pair [ A(ω), B] is controllable, which follows
[0, t ∗f ]. Then {t ∗f , Tc∗ , x ∗ , m ∗ } ∈ F ∗f with ω̂. from the fact that [B A(ω̂)B] is an invertible matrix. Hence
Proof: This proof uses Lemmas 2 and 3, which are given the pair (B T , −A(ω̂)T ) is observable. Consequently, y(t) =
in the Appendix. 0 ∀t ∈ [0, t ∗f ] implies that λ(t) = 0 ∀t ∈ [0, t ∗f ]. Hence
Let q̃ := E r ∗ (t ∗f ). Then we can also consider η̇(t) = 0 ∀t ∈ [0, t ∗f ]. Since η(t ∗f ) = 0, this then implies
{t f , Tc∗ ,
∗ , x ∗ , m ∗ } as an optimal fuel solution of Problem
∗ that η(t) = 0 ∀t ∈ [0, t ∗f ]. These imply H (φ(t)) = β
(t).
4, where the constraint E r(t f ) − q ≤ dP1 ∗ − q is replaced Since H (φ(t ∗f )) = 0 and
(t) ≥ ρ1 > 0, this suggests
by E r(t f ) = q̃. So, without loss of any generality, this version that β = 0. Therefore, (β, λ(t), η(t)) = 0 ∀ t ∈ [0, t ∗f ],
of Problem 4 will be used in this proof. which is a contradiction with necessary Condition 1) above.
Since x ∗ (t) ∈ intX and m(t) > m 0 − m f for all t ∈ [0, t f ], Consequently, there are countably many number of points in
the Maximum Principle of optimal control [see [18, Sec. V.3] [0, t ∗f ] where y(t) = 0. Since a countable set has measure
or [26, Ch. 1]], there exists a constant β ≤ 0 and absolutely zero, condition (29) holds.
continuous function λ : R+ → R6 and η : R+ → R, the Since any countable set has measure zero, the second
co-state vectors, such that the following conditions hold. conclusion of Lemma 2 implies that
1) Co-state conditions: ∀t ∈ [0, t ∗f ] y(t) = −α(t)n̂ a.e. [0, t ∗f ] for α(t) > 0. (30)
(β, λ(t), η(t)) = 0 (22)
Since condition (29) holds, a.e. [0, t ∗f ] such that y(t) = 0,
λ̇(t) = −A(ω̂)T λ(t) (23) and for a given
∗ (t) an optimal control thrust Tc∗ (t) must
λ(t)T B Tc (t) satisfy
η̇(t) = . (24)
m(t)2
Tc∗ (t) = argmax
∗
y(t)T Tc = argmax ∗ y(t)T Tc (31)
2) Pointwise maximum principle: (Tc ,
(t ))∈V Tc ∈U(
)

H (φ(t)) = M(x ∗(t), m ∗(t), λ(t), η(t)) a.e. t ∈ [0, t ∗f ] (25) where U(
) := {Tc ∈ R3 : Tc  ≤
, n̂T Tc ≥
cos θ }.
Furthermore, since condition and (30) holds, a.e.[0, t ∗f ] such
where φ(t) = (t, x ∗ (t), m ∗ (t), Tc∗ (t),
∗ (t), λ(t), η(t)) H is that y(t) = 0 and y(t) = −α(t)n̂ for some α(t) > 0, the
the Hamiltonian defined by maximizing solution of (31) must be on the boundary point
  of U(
∗ ) which is also an extremal point of the set U(
∗ ),
H (φ) := β
+ λT A(ω̂)x + B(g + Tc /m) − α
η (26)
which follows from Lemma 3. This lemma also implies that
and, by letting V := {(Tc ,
) ∈ R4 : Tc  ≤
, ρ1 ≤
≤ all extremal points of the set U(
) satisfy Tc  =
, and
ρ2 , n̂T Tc ≥
cos θ } hence Tc∗ (t) =
∗ (t)
M(x ∗ , m ∗ , λ, η) = max H (φ). (27) Tc∗ (t) =
∗ (t) a.e. [0, t ∗f ] (32)
(Tc ,
)∈V

3) Transversality conditions: which implies that an optimal solution of the relaxed problem
(4) satisfies
η(t ∗f ) = 0 and H (φ(t ∗f )) = 0. (28)
0 < ρ1 ≤ Tc∗ (t) ≤ ρ2 , n̂T Tc∗ (t) ≥ Tc∗ (t) cos θ a.e. [0, t ∗f ].
The necessary conditions of optimality 1) and 2) directly
follow from the statement of the Maximum Principle. But the Consequently, (t ∗f , Tc∗ , x ∗ , m ∗ ) ∈ F f . Since for any
transversality condition requires further explanation. Transver- (t f , Tc , x, m) ∈ F f , (t f , Tc , Tc , x, m) ∈ Fr f , an optimal
sality condition implies that (see [18, p. 190, Sec. V.3]), for an solution of Problem 4 has an optimal cost which is not
optimal solution of the relaxed problem, the vector ψ, defined greater than the optimal cost of Problem 2. This implies that
by (t ∗f , Tc∗ , x ∗ , m ∗ ) ∈ F ∗f from Lemma 1.
  The above theorem states that we can find the optimal
ψ := H (φ(0)), H (φ(t ∗f )), −λ(0), −η(0), λ(t ∗f ), η(t ∗f ) solution for Problem 2 by solving its relaxation in Problem 4
for ω̂. Clearly, for ω = ω̂, we can find the exact optimal
must be orthogonal to the manifold defined by the
solution of the original problem of interest by solving its
set of feasible initial and final states described by
relaxation. When ω̂ = ω, we can find optimal solutions of
(0, t f , x0 , m 0 , (0, q̃, 0), m(t ∗f )), which is given by
a problem that can be made arbitrarily close to the problem of
span {e2 , e14}. The above follows from the fact that t f
interest by simply choosing  > 0 close to zero. Also when
and m(t f ) are the only free variables in the manifold of
θ = π, i.e., when there is no pointing constraint, we can use
boundary conditions. This then implies that e2T ψ = 0 and
T ψ = 0, i.e., H (φ(t ∗ )) = 0 and η(t ∗ ) = 0. Next we show any unit vector for n̂ such that ω = ω̂. Hence we can find the
e14 f f exact optimal solution of the original problem of interest.
that
This result has a connection with “normal systems” [18].
y(t) := B T λ(t) = 0 a.e. [0, t ∗f ]. (29)
A normal linear system is defined in the context of the
This will be done by contradiction. Suppose that the condition optimal control theory and a linear system is said to be normal
(29) does not hold. Since y is an output of the system given with respect to a set of feasible controls if it maximizes the
2110 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013

As mentioned earlier, a more benign source of noncon-

vexity in the relaxed problems is the existence of 1/m
term multiplying the thrust vector in the dynamics. This
leads to nonlinearity in the dynamics of the system, which
then leads to nonconvex equality constraints in the result-
ing optimization problem. This difficulty will be resolved
as described in [1], and is briefly explained in the next
subsection.
Both steps above require solutions of the optimal control
(a) (b) problems in continuous time. These solutions are obtained
by discretizing the problems (with the change of variables
Fig. 5. Planar representation of U(
): (a) θ ∈ (π/2, π ) and (b) θ ∈ [0, π/2]. given in the next section) to obtain their discrete-time approx-
Extreme points of U(
) satisfy that Tc  =
, and they are on the solid arc
in both figures. The optimal solutions of Problem 4 must be on these extreme imations, which are convex parameter optimization prob-
points in order to produce feasible, hence optimal, solutions of Problem 2, lems: more specifically, SOCP problems. SOCPs are solved
which is established by Theorem 1. numerically via IPMs to obtain the solutions of the optimal
control problems in both steps. This discretization is explained
in [1].
Hamiltonian at a unique point of the set of feasible controls. Theorem 1 ensures convexification will hold when the state
In the case when this set is convex, a system being normal trajectories are strictly inside the set of feasible states. We
implies that the Hamiltonian is maximized at an extreme point observed in our extensive simulations that the result still holds
of the set of feasible controls (see [18, Coroll. 7.2]). Our even when the state trajectories intersect with the boundary
convexification result has a similar geometric interpretation of the set of feasible states. A proof of this observation
since it establishes lossless convexification by ensuring that is not included here, but will be the topic of a future
Hamiltonian is maximized at the extreme points of a projection paper.
of the relaxed set of feasible controls. This set is then shown to It is noteworthy that our convexification result relies on the
be contained in the original nonconvex set of feasible controls, pair [ A(ω̂), B] being controllable. So, it still applies under
thereby establishing that we can obtain optimal solutions system parameter variations as long as this controllability
for the original nonconvex problem via solving its convex property is preserved. This is clearly not a restrictive condition
relaxation. The projected set of the relaxed set of controls since it is reasonable to expect that a planetary landing vehicle
mentioned above, U, is given by would be designed to be controllable.
Another important issue is the robustness to uncertainties to
U(
) := {Tc ∈ R3 : Tc  ≤
, TcT n̂ ≥ cos θ
}. unknown parameter variations in the dynamic model or to the
disturbance forces. Here we provide a method to compute the
Set U and its extreme points are shown in Fig. 5 for θ ∈
optimal state trajectory to be followed and the corresponding
[0, π/2] and θ ∈ (π/2, π), and one can observe that its
optimal controls. Though it is not covered in this paper, one
extreme points satisfy Tc  =
, which results in the con-
has to consider a feedback control action to handle uncertain-
vexification of Problem 2 via Problem 4. Indeed, Theorem 1
ties and disturbances. There are several approaches that we can
establishes that any optimal solution of Problem 4 must lie
propose to achieve this: 1) to have a tracking feedback con-
on the extreme points of U(
∗ ), and then the convexification
troller to follow the optimal state trajectory; 2) to recompute
follows from Lemma 1. Note that being in the interior of U(
)
the optimal trajectory and controls as the state knowledge (esti-
or on a boundary point that is not an extreme point of U(
)
mates) are updated during the maneuver; and 3) a combination
can lead to thrust vectors that are not feasible for the original
of tracking feedback and optimal control recomputations. All
problem, i.e., Problem 2. But, as noted earlier, an extreme
actions would utilize the current best estimate of the state and
point of the set satisfies Tc  =
and hence is in the set
would provide the feedback action to minimize the effects
of feasible controls for Problem 2. Since our result proves
of uncertainties and disturbances. The design of a robust
that the optimal solution of the relaxed problems occurs at the
feedback action will be another future extension to our current
extreme points, it guarantees that the resulting thrust vectors
paper.
will always be feasible. Equipped with Lemma 1 and Theorem
A natural question is whether the convexification approach
1, we propose the following two-step algorithm for the solution
here can be extended to other control constraints. Some new
of the soft landing problem:
results for linear systems are provided in [27], where this
1) solve Problem 3 with ω̂ to obtain dP3 ; convexification procedure is applied to a large class of control
2) solve Problem 4 with ω̂ and dP3. constraints. However, there the convexification used the fact
In cases when ω̂ = ω, the above algorithm will produce that optimal controls are on the boundary points of the relaxed
optimal solutions to a problem with dynamics slightly per- set of controls rather than on the extremal points. With the
turbed from the original dynamics. This is necessary for insight obtained in the current paper on the extremal points
purely theoretical reasons, and we have not yet empirically of the relaxed set, there can be further generalization to this
encountered any case requiring this perturbation of the convexification approach, which can be the subject of future
dynamics. research.
AÇIKMEŞE et al.: CONTROL BOUND AND CONSTRAINTS OF OPTIMAL CONTROL 2111

A. Change of Variables
We use the following change of variables on the thrust
vector and mass to remove the nonlinearity in the dynamics

Tc
due to Tc /m: σ  , u  , z  ln m.
m m
The mass depletion dynamics can then be rewritten as
ṁ(t)
ż = = −ασ (t) . (33)
m(t)
(a) (b) (c)
The change of variables therefore yields a set of linear equa-
Fig. 6. Simulation results with three different pointing constraints. (a) Uncon-
tions for the state dynamics. The control constraints, however, strained. (b) θ = 90°. (c) θ = 45°. Thrust pointing and magnitude constraints
are no longer convex. These are now given by of Problem 2 are satisfied for the optimal solution of Problem 4, as seen in
the first two plots. The last plot is the position trajectory of each solution.
u(t) ≤ σ (t) n̂T u(t) ≥ cos θ σ (t) ∀t ∈ [0, t f ] (34)
ρ1 e−z(t ) ≤ σ (t) ≤ ρ2 e−z(t ) ∀t ∈ [ 0, t f ]. (35)

As in [1], we use a second-order cone approximation of the

inequalities in (35) that can be readily incorporated into the
SOCP solution framework, given by
 
(z(t) − z 0 (t))2 Fig. 7. Example with one contact with the glide slope boundary and two
ρ1 e−z0 1 − (z(t) − z 0 (t)) + contacts with the speed boundary. The throttle and thrust angle from vertical
2 time profiles (second row) indicate that the thrust vector profile is feasible
−z 0 for Problems 1 and 2, and the convexification still holds, and we obtain an
≤ σ (t) ≤ ρ2 e [1 − (z(t) − z 0 (t))] ∀ t ∈ [0, t f ] (36) optimal solution to the soft landing problem.
where z 0 (t) = ln(m 0 − αρ2 t) and m 0 is the initial mass of the TABLE I
spacecraft. T IGHTENING T HRUST P OINTING C ONSTRAINTS A FFECTING THE F UEL
In [1, Lemma 2], it is shown that the approximation of the U SE AND THE F LIGHT T IME
inequalities in (35) given by (36) always produces a feasible
solution of the relaxed problems and is generally very accurate Attitude Required Fuel (kg) Flight Time (s)
Unconstrained 200.1 44.63
for both parts of the inequality and derives an analytic upper
90° constraint 201.8 46.96
bound on the approximation error. Since this paper’s focus is
45° constraint 222.3 57.29
the convexification of the control constraints, we will not go
into further details of this approximation and refer the reader
and the target landing site is at q = 0 m, which
to [1] for more details.
is the origin of the guidance frame. The Mars gravita-
tional parameters, also expressed  in the same coordinate
IV. N UMERICAL E XAMPLES frame
 are as follows: g = −3.71, 0, 0 m/s2 and ω =
−5
2.53×10 , 0, 6.62×10 −5 rad/s. Since S(ω)n̂ = 0 and
This section presents a numerical example to demonstrate
the algorithm proposed in the previous section to solve the NT S(ω)2 n̂ = 0, we have ω̂ = ω. Three simulations were
planetary soft landing problem. performed for varying pointing-constraint limits: 1) uncon-
The pointing constraints in the planetary soft landing ensure strained; 2) 90° constraint; and 3) 45° constraint. The results
that the translational maneuver does not require the spacecraft of these simulations are overlaid in Fig. 6. As seen in the plot,
to be oriented outside of a desired pointing cone, which usually the pointing angle is relative to local vertical, which aligns the
results in a reduction in performance. For instance, as the pointing cone n̂ vector along the coordinate frame x-axis. The
pointing constraints are tightened, the required fuel and flight attitude pointing plot indicates that the solution of the relaxed
time generally increase due to the restricted thrust pointing problem ensures satisfying the prescribed pointing constraints
capability. This result will be seen in the following comparison for the original problem. The throttle plot shows that the
simulations, which make use of an example with the following thrust bounds are satisfied, which indicates that the con-
properties: m 0 = 2000 kg, m f = 300 kg, ρ1 = 0.2 Tmax , straint relaxations in Problem 4 on both thrust magnitude and
ρ2 = 0.8 Tmax , Tmax = 24000 N, and α = 5 × 10−4 s/m, pointing remain valid for the original problem [Problem 2].
where Tmax is the maximum thrust magnitude. The thrust This figure provides some further insight on the tradeoff in
limits coincide with a minimum and maximum throttle level of performance that occurs as the constraints are tightened. As
20% and 80%, respectively. A frame of reference, as illustrated the pointing limit tightens, the required flight time and fuel
in Fig. 1, is used to express the Martian gravity and rotation increase, as summarized in Table I, which is also visible in the
vector, as well as the spacecraft’s initial state conditions at position trajectory. The last plot of Fig. 6 overlays the position
powered-descent ignition. The initial state of the spacecraft trajectories coinciding with the three thrust profiles from the
for the simulations is given by prior figures. The 45° case overshoots the target along the
    y-axis to satisfy the pointing constraint. Interestingly, the 90°
r0 = 2400, 450, −330 m, ṙ0 = −10, −40, 10 m/s constrained path takes a more direct route.
2112 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 21, NO. 6, NOVEMBER 2013

The next simulation (see Fig. 7) presents a case where the  S(n̂)S(ω)n̂. This has a nonzero component in the nullspace
convexification holds with the finite number of contacts with of n̂T , because  S(n̂)S(ω)n̂ is nonzero and in the nullspace
the boundary of X. In this example, we have the following of n̂T , and S(ω̂)2 n̂ is perpendicular to the nullspace of n̂T .
parameters: Consequently, we have NT v 2 = 0. Now note that v 1T n̂ = 0.
    Let v 3 be a nonzero vector such that v 3T n̂ = v 3T v 1 = 0. Hence
r0 = 2400, 3400, 0 m, ṙ0 = −40, 45, 0 m/s
v 1 and v 3 span the nullspace of n̂. This implies that we can
with a glide slope γgs = 30° and θ = 120°, with a maximum write v 2 = c1 n̂ + c2 v 3 + c3 v 1 for some scalars c1 through c3 .
velocity of 90 m/s. Since v 2T v 1 = 0 and NT v 2 = 0 we know that c3 = 0 and
c2 = 0.
V. C ONCLUSION Observe that
 t
This paper presented a lossless convexification of thrust
pointing constraints for a benchmark problem in optimal λ̇2 (t) = λ1 (t1 ) − α(s)ds v 2 − 2α(t)v 1
t1
control theory, known as soft landing. This extended and
unified our previous work in the area, such that the algorithm which implies that
could handle upper and lower bounds on thrust, thrust pointing λ̇2 (t) = λ1 (t1 ) + g1 (t)n̂ + g2 (t)v 3 − 2α(t)v 1
constraints, position and velocity constraints, and planetary t
where g1 (t) = c1 t1 α(s)ds and where g2 (t) =
rotation. This convexification enables the planetary soft land- t
ing problem to be solved optimally by using convex optimiza- c2 t1 α(s)ds = 0 for t ∈ (t1 , t2 ). Since v 1 , v 3 , and n̂ form
tion, and is hence amenable to onboard implementation with an orthogonal set of nonzero vectors, if α is a nonconstant
a priori bounds on the computational complexity and realtime function of time then NT λ̇2 (t) = 0 a.e. [t1 , t2 ]. If α is constant,
execution time. then g2 is a nonconstant, linear function of time (since c2 = 0),
and hence NT λ̇2 (t) = 0 a.e. [t1 , t2 ]. Consequently, NT λ̇2 (t) =
A PPENDIX 0 a.e. [t1 , t2 ], which implies that there exists an open finite
Lemma 2: Consider the following linear time invariant interval in [t1 , t2 ] such that λ2 will leave the subspace defined
system: by n̂, which is a contradiction caused by assuming that λ2 (t) =
−α(t)n̂ in [t1 , t2 ]. Therefore there exists no finite interval
λ̇(t) = −A(ω̂)T λ(t) where y(t) = λ2 (t) = −α(t)n̂, that is, κ(t) := NT λ̇2 (t) = 0.
y(t) = B T λ(t) (37) Since κ is analytic function of time (due to λ being analytic
function of time), either it has a countable number of zeros
where λ(t) ∈ R6 and y(t) ∈ R3 , and A(ω̂) ad B are given
or there exists time intervals on which it is zero [28]. Since
by (2). Then the following conditions hold true; for any finite
we proved that the latter is not possible, then it must have a
interval [0, t f ]:
countable number of zeros. This concludes the proof.
(i) y is an analytic function, and y(t) = 0 on either [0, t f ] The following lemma is instrumental in the use of Pontryagin’s
or at countable number of instances; Maximum Principle to show that the optimal control occur at
(ii) there is a countable number of instances in [0, t f ] such extreme points of a convex feasible set of controls.
that y(t) = α(t)n̂ for some α(t) > 0. Lemma 3: An optimal solution of the following optimiza-
Proof: λ, hence y, are analytic functions of t from [28, tion problem is also an extreme point of the feasible set of
Th. 3, p. 213]. Hence y(t) = 0, ∀t ∈ [0, t ∗f ], or y(t) = 0 solutions U(
): maxTc yT Tc subject to Tc ∈ U(
) where
occurs at a countable number of points in [0, t ∗f ] (Proposition U(
) := {Tc : Tc  ≤
, n̂T Tc ≥ cos θ
}, and y = 0 and
4.1 on [28, p. 41]). This proves (i). y = −α n̂ for any α > 0. Consequently an optimal solution
To prove condition (ii), suppose that there exists an interval Tc∗ satisfies that Tc∗  =
.
[t1 , t2 ] ⊆ [0, t ∗f ], such that y(t) = −α(t)n̂ ∀ t ∈ [t1 , t2 ], Proof: Since the cost function of the optimization problem
for α(t) > 0. Letting λ = (λ1 , λ2 ) where λ1,2 ∈ R3 , is linear, hence convex, the maximization problem leads to an
v 1 := S(ω̂)T n̂ and v 2 := S(ω̂)T v 1 , this assumption implies optimal solution on the boundary of U, i.e., on ∂U [29]. ∂U
that the following dynamics hold for t ∈ [t1 , t2 ] : λ̇1 (t) = has a portion of the sphere with radius
and a portion of
−α(t)v 2 and λ̇2 (t) = −λ1 (t) − 2α(t)v 1 . a hyperplane with unit normal n̂. All extremal points of the
Due to the construction of ω̂ in Problem 4, we have v 1 = 0 boundary are on the sphere: if any point of the boundary is not
and NT v 2 = 0. The first inequality is straightforward to show. on the sphere then it is on the hyperplane, on a line segment
To see the second one, we need to consider the three cases in between two other points of the boundary. This implies that
(21). First consider the case when S(ω)n̂ = 0, NT S(ω)2 n̂ = 0. an extremal point of U satisfies that Tc  =
.
Then ω = ω̂ and NT v 2 = NT S(ω)2 n̂ = 0. Second, consider Note that y = c1 n̂ + c2 n̂⊥ for some c2 = 0 and unit vector
the case when S(ω)n̂ = 0, that is, ω = a n̂ for some a ∈ R. n̂ that is orthogonal to n̂. This implies that Tc∗ = a1 n̂ +a2 n̂⊥
⊥
2
Then ω̂ = a n̂ +  n̂⊥ . Then v 2 = S(ω̂)T n̂ = S(ω̂)2 n̂ = where a2 = 0 and a12 + a22 ≤
2 . Suppose that n̂T Tc∗ =
 S(a n̂ +  n̂⊥ )S(n̂⊥ )n̂. Since S(n̂⊥ )n̂ is orthogonal to both a n̂ cos θ
, which implies that a1 = cos θ
. Since yT Tc∗ = c1 a1 +
and  n̂⊥ , it is orthogonal to their sum, which is nonzero, and c2 a2 = c1 cos θ
+ c2 a2 , the cost is the maximum if c2 and a2
hence v 2 = 0 and it is in the null space of n̂T . Consequently has the same sign and the a12 + a22 =
2 , that is, Tc∗  =
.
2
NT S(ω̂)T n̂ = 0. Finally, consider the case when S(ω)n̂ = 0 Consequently an optimal solution must be an extreme
2
but NT S(ω)T n̂ = 0. Then we have S(ω̂)2 n̂ = S(ω)2 n̂ + point of U.
AÇIKMEŞE et al.: CONTROL BOUND AND CONSTRAINTS OF OPTIMAL CONTROL 2113

ACKNOWLEDGMENT [22] Y. Wang and S. Boyd, “Fast model predictive control using online
optimization,” IEEE Trans. Control Syst. Technol., vol. 18, no. 2,
The authors would like to thank M. Ivanov and J. Casoliva pp. 267–278, Mar. 2010.
of the Jet Propulsion Laboratory for their valuable comments, [23] Y. Kim and M. Mesbahi, “Quadratically constrained attitude control via
and S. Boyd, Y. Wang, and J. Mattingley of Stanford semidefinite programming,” IEEE Trans. Autom. Control, vol. 49, no. 5,
pp. 731–735, May 2004.
University for their insights on convexification and real-time [24] G. P. Sutton and O. Biblarz, Rocket Propulsion Elements. New York:
convex optimization. This research was partially performed Wiley, 2010.
at the Jet Propulsion Laboratory, California Institute of [25] M. Tamiz, D. Jones, and C. Romero, “Goal programming for decision
making: An overview of the current state-of-the-art,” Eur. J. Oper. Res.,
Technology, under a contract with the National Aeronautics vol. 111, no. 3, pp. 569–581, Dec. 1998.
and Space Administration. [26] L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze, and E. F.
Mischenko, The Mathematical Theory of Optimal Processes. New York:
R EFERENCES Pergamon, 1964.
[27] B. Açıkmeşe and L. Blackmore, “Lossless convexification of a class of
[1] B. Açıkmeşe and S. R. Ploen, “Convex programming approach to optimal control problems with non-convex control constraints,” Auto-
powered descent guidance for mars landing,” Amer. Inst. Astron. Aeron. matica, vol. 47, no. 2, pp. 341–347, 2011.
J. Guid., Control Dyn., vol. 30, no. 5, pp. 1353–1366, 2007. [28] H. Cartan, Elementary Theory of Analytic Functions of One or Several
[2] L. Blackmore, B. Açıkmeşe, and D. P. Scharf, “Minimum landing error Complex Variables. New York: Dover, 1995.
powered descent guidance for mars landing using convex optimization,” [29] L. D. Berkovitz, Convexity and Optimization in R n . New York: Wiley,
Amer. Inst. Astron. Aeron. J. Guid., Control Dyn., vol. 33, no. 4, pp. 2002.
1161–1171, 2010.
[3] J. S. Meditch, “On the problem of optimal thrust programming for Behçet Açıkmeşe received the Ph.D. degree in
a lunar soft landing,” IEEE Trans. Autom. Control, vol. 9, no. 4, aerospace engineering from Purdue University, West
pp. 477–484, Oct. 1964. Lafayette, IN, in 2002.
[4] A. Miele, The Calculus of Variations in Applied Aerodynamics and He joined The University of Texas at Austin,
Flight Mechanics in Optimization Techniques, G. Leitmann, Ed. Austin, in 2012, where he is currently involved in
New York: Academic, 1962. research on robust and nonlinear control, optimal
[5] A. R. Klumpp, “Apollo lunar descent guidance,” Automatica, vol. 10, control, model predictive control, convex optimiza-
no. 2, pp. 133–146, Mar. 1974. tion, and linear matrix inequalities in guidance,
[6] U. Topcu, J. Casoliva, and K. D. Mease, “Minimum-fuel powered control, and estimation. He was a Visiting Assis-
descent for mars pinpoint landing,” J. Spacecraft Rockets, vol. 44, no. 2, tant Professor of aerospace engineering with Purdue
pp. 324–331, 2007. University. He joined the Jet Propulsion Laboratory
[7] R. Sostaric and J. Rea, “Powered descent guidance methods for (JPL), NASA, Pasadena, CA, in 2003, where he was a Technologist and a
the moon and mars,” in Proc. AIAA Guid., Navigat., Control Conf., Senior Member of the Guidance and Control Analysis Group and developed
Aug. 2005, no. AIAA 2005-6287. onboard control algorithms for planetary landing, formation flying, and
[8] F. Najson and K. D. Mease, “Computationally inexpensive guidance asteroid and comet sample return.
algorithm for fuel-efficient terminal descent,” J. Guid., Control, Dyn., Dr. Açkmese was a member of the Mars Science Laboratory Controls Team
vol. 29, no. 4, pp. 955–964, Jul. 2006. and developed the flight algorithms for the “fly-away” phase of the landing,
[9] B. Steinfeldt, M. Grant, D. Matz, and R. Braun, “Guidance, navigation, which were successfully executed during the Mars landing on 5 August 2012.
and control technology system trades for mars pinpoint landing,” in John M. Carson III received the B.S. degree
Proc. AIAA Atmosph. Flight Mech. Conf. Exhibit, Aug. 2008, no. AIAA in aerospace engineering and the M.S. degree in
2008-6216. engineering mechanics from the University of Texas
[10] S. Boyd and L. Vandenberghe, Convex Optimal. Cambridge, U.K.: at Austin, Austin, in 1997 and 1999, respectively,
Cambridge Univ. Press, 2004. and the Ph.D. degree in mechanical engineering from
[11] Y. Nesterov and A. Nemirovsky, Interior-Point Polynomial Methods in the California Institute of Technology, Pasadena, in
Convex Programming. Philadelphia, PA: SIAM, 1994. 2008.
[12] J. Peng, C. Roos, and T. Terlaky, Self-Regularity: A New Paradigm for He was with the Lockheed Martin Skunk Works
Primal-Dual Interior-Point Algorithms. Princeton, NJ: Princeton Univ. from 1999 to 2000, where he was involved in
Press, 2001. research on conducting flight and structural dynam-
[13] J. F. Sturm, “Using sedumi 1.02, a MATLAB toolbox for optimiza- ics tests and analyses on the Joint Strike Fighter
tion over symmetric cones,” Optim. Methods Softw., vol. 17, no. 6, prototype. Since 2001, he has been with the Jet Propulsion Laboratory,
pp. 1105–1154, 2002. NASA, Pasadena, CA, where he is involved on GN&C design, analysis,
[14] C. D’Souza, “An optimal guidance law for planetary landing,” in Proc. and hardware development for several missions, including Cassini/Huygens
AIAA Guid., Navigat. Control Conf., 1997, pp. 1376–1381. and Mars Exploration Rover, and in research and technology development
[15] S. Ploen, B. Açıkmeşe, and A. Wolf, “A comparison of powered programs for small-body and planetary landing systems. He currently plays a
descent guidance laws for mars pinpoint landing,” in Proc. AIAA/AAS part in the project management and integration and testing activities with the
Astrodynamics Special. Conf. Exhibit, Aug. 2006, no. AIAA 2006-6676. NASA Autonomous Landing and Hazard Avoidance Technology project that
[16] B. A. Steinfeld, M. J. Grant, D. A. Matz, R. D. Braun, and G. H. is developing safe precision landing sensors and algorithms for robotic and
Barton, “Guid., navigation and control system performance trades crewed lander vehicles.
for mars pinpoint landing,” J. Spacecraft Rockets, vol. 47, no. 1,
pp. 188–198, 2010. Lars Blackmore received the M.Eng. (Hons.)
[17] J. M. Carson, B. Açıkmeşe, and L. Blackmore, “Lossless convexification degree from Cambridge University, Cambridge,
of powered-descent guidance with non-convex thrust bound and pointing U.K., and the Ph.D. degree in aeronautics and astro-
constraints,” in Proc. Amer. Control Conf., Jun. 2011, pp. 2651–2656. nautics from the Massachusetts Institute of Technol-
[18] L. D. Berkovitz, Optimal Control Theory. New York: Springer-Verlag, ogy, Cambridge.
1975. He is currently a Senior Guidance, Navigation,
[19] K. Toh, M. Todd, and R. Tutuncu, “SDPT3—a Matlab software package and Control Engineer with Space Exploration Tech-
for semidefinite programming,” Optim. Methods Softw., vol. 11, no. 1, nologies (SpaceX), Hawthorne, CA, where he leads
pp. 545–581, 1999. the propulsive landing effort, aimed at creating fully
[20] J. Mattingley and S. Boyd, “Real-time convex optimization in signal reusable spacecraft and he is involved in the devel-
processing,” IEEE Signal Process. Mag., vol. 27, no. 3, pp. 50–61, opment of systems, operations, and algorithms for
May 2010. precision propulsive landing of booster stages and capsules, including the
[21] J. Mattingley and S. Boyd, Automatic Code Generation for Real-Time full-scale “Grasshopper” experimental vehicle. He was with the Jet Propulsion
Convex Optimization. Convex Optimization in Signal Processing and Laboratory, NASA, where he was involved in the development of provably
Communications, Y. Eldar and D. Palomar, Eds. Cambridge, U.K.: optimal algorithms for Mars precision landing, and was a member of the GNC
Cambridge Univ. Press, 2010. Team for the Soil Moisture Active Passive climate science mission.