Modern Homing Missile Guidance Theory

and Techniques

Neil F. Palumbo, Ross A. Blauwkamp, and Justin M. Lloyd

lassically derived homing guidance laws, such as proportional

navigation, can be highly effective when the homing mis-
sile has significantly more maneuver capability than the threat.
As threats become more capable, however, higher performance is required from the
missile guidance law to achieve intercept. To address this challenge, most modern guid-
ance laws are derived using linear-quadratic optimal control theory to obtain analytic
feedback solutions. Generally, optimal control strategies use a cost function to explicitly
optimize the missile performance criteria. In addition, it is typical for these guidance
laws to employ more sophisticated models of the target and missile maneuver capabil-
ity in an effort to improve overall performance. In this article, we will present a review
of optimal control theory and derive a number of optimal guidance laws of increasing
complexity. We also will explore the advantages that such guidance laws have over
proportional navigation when engaging more stressing threats

INTRODUCTION
Classical guidance laws, with proportional navigation optimal control theory to missile guidance problems had
(PN) being the most prominent example, had proven to sufficiently matured, offering new and potentially prom-
be effective homing guidance strategies up through the ising alternative guidance law designs.1–3 Around this
1960s and early 1970s. By the mid-1970s, however, the time, the computer power required to mechanize such
predicted capability of future airborne threats (highly advanced algorithms also was sufficient to make their
maneuverable aircraft, supersonic cruise missiles, tacti- application practical.
cal and strategic ballistic missile reentry vehicles, etc.) Most modern guidance laws are derived using linear-
indicated that PN-guided weapons might be ineffective quadratic (LQ) optimal control theory to obtain
against them. However, by that time, the application of analytic feedback solutions.1, 4, 5 Many of the modern

42 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

formulations take target maneuver into account to deal
$
x (t) = f (x (t), u (t), t) . (1)
with highly maneuvering target scenarios (particularly
true for terminal homing guidance). The availability In Eq. 1, x is the n-dimensional state vector of real
of target acceleration information for the guidance law elements (x ! R n), u ! R m is the control vector, and
varies, depending on targeting sensor capability and t represents time (later we will provide more detail as
type and the specific guidance law formulation. Typi- to the structure of the state and control vectors for the
cally, there also is an explicit assumption made about homing guidance problem). With this general system, we
the missile airframe/autopilot dynamic response charac- associate the following scalar performance index:
teristics in the modern formulations. We will show later
that PN is an optimal guidance law in the absence of J (x (t 0), u (t 0), t 0) =  (x (t f), t f)
airframe/autopilot lag (and under certain other assumed tf
(2)
conditions). To some extent, the feedback nature of the + #t L (x (t), u (t), t) dt .
homing guidance law allows the missile to correct for 0

inaccurate predictions of target maneuver and other

unmodeled dynamics (see Fig. 1). However, the require- In this equation, [t0,  tf] is the time interval of inter-
ment for better performance continues to push optimal est. The performance index comprises two parts: (i)
guidance law development, in part by forcing the con- a scalar algebraic function of the final state and time
sideration (inclusion) of more detailed dynamics of the (final state penalty),  (x (t f), t f) , and (ii) a scalar inte-
interceptor and its target. It is interesting that most if gral function of the state and control (Lagrangian),
not all modern guidance laws derived using optimal con- #t0tf L (xt (t), u (t), t) dt . The choice of  (x (t f), t f)
trol theory can be shown to be supersets of PN. and #t f L (x (t), u (t), t) dt (this choice is a significant part
In the following sections, we first provide a cursory 0
review of dynamic optimization techniques with a focus of the design problem) will dictate the nature of the
on LQ optimal control theory. Using this as background, optimizing solution. Thus, the performance index
we then develop a number of terminal homing guid- is selected to make the plant in Eq.  1 exhibit desired
ance laws based on various assumptions, in the order of characteristics and behavior (transient response, stabil-
assumption complexity. In our companion article in this ity, etc.).
issue, “Guidance Filter Fundamentals,” we introduce For our purposes, the optimal control problem is
guidance filtering, with a focus on Kalman guidance to find a control, u *(t) , on the time interval [t0, tf ] that
filter techniques. drives the plant in Eq.  1 along a trajectory, x *(t) , such
that the scalar performance index in Eq. 2 is minimized.
It is difficult to find analytic guidance law expressions
REVIEW OF LQ OPTIMAL CONTROL for such general nonlinear systems. Therefore, we will
Here, we start by considering a general nonlinear turn our attention to a subset of optimal control that
dynamics model of the system to be controlled (i.e., the can yield tractable analytic solutions, known as the LQ
“plant”). This dynamics model can be expressed as optimal control problem.

Kalman Filter Optimal Control Servomechanism

Theory Theory Theory
(1960s) (1960s) (1930s–1950s)

Target Target state Acceleration Actuation

direction estimates commands commands
Target
motion Target Guidance Guidance Airframe/
Autopilot
sensors filter law propulsion

Inertial
navigation

Vehicle
motion

Figure 1.  The traditional guidance, navigation, and control topology for a guided missile comprises guidance filter, guidance law,
autopilot, and inertial navigation components. Each component may be synthesized by using a variety of techniques, the most
popular of which are indicated here in blue text.

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

43­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

The LQ Optimal Control Problem Then, from the calculus of variations, four necessary
Here, we assume that the nonlinear model in Eq. 1 conditions for optimality must be satisfied5 to solve our
can be linearized about an equilibrium point (x 0, u 0) problem: state, costate, boundary, and stationarity con-
and represented by the time-varying linear dynamic ditions must all hold. These four conditions are listed in
Table 1.
system described in Eq. 31, 4:
From Table 1 (and based on our previous description
$ of the dimension of the plant state vector), it can be seen
x (t) = A (t) x (t) + B (t) u (t) . (3)
that there is a system of 2n dynamic equations that must
be solved; n equations must be solved forward in time
In this model (as in the nonlinear case), x ! R n is the
over [t0,  tf] (state equations), and n equations must be
state vector, u ! R m is the control vector, and t repre-
solved backward in time over [tf, t0] (costate equations).
sents time. Here, however, A(t) and B(t) are the time-
We further note that the equations are coupled. Apply-
varying Jacobian matrices ing the stationarity condition to Eqs. 3–6 yields the fol-
lowing result for the control:
(2 f (x, u, t)/2 x x=x and 2 f (x, u, t)/2 u x=x , respectively)
0 0
u=u u=u
0 0
u (t) = –R –1 (t) B T (t)  (t) . (7)
of appropriate dimension.
Using the “shorthand” notation z (t) 2 _ z T (t) Sz (t) , Using Eq. 7 in the state and costate necessary conditions,
S
we next define the quadratic performance index (a and taking into account the boundary condition, leads
special case of the performance index shown in Eq. 2) to the following two-point boundary value problem:
shown in Eq. 4:
$
A (t) –B (t) R –1 B T x (t)
=$ G== G;
x (t)
J(x (t 0), u (t 0), t 0) = 1
x (t f) 2 E,
2 Qf  (t) –Q (t) –A T (t)  (t)
(8)
(4) x (t 0) given
)  (t ) = Q x (t ) 3 .
tf
2 t 8
# Bdt .
1 2 2
+ x (t) + u (t)
Q (t) R (t) f f f
0

In Eq.  4, the following assumptions are made: the ter- Here, we are concerned with solution methods that
minal penalty weighting matrix, Q f, is positive semi- can yield analytic (closed-form) solutions rather than
definite (the eigenvalues of Q f are 0, expressed as iterative numerical or gain scheduling techniques. We
Q f    0); the state penalty weighting matrix, Q(t), is note, however, that the sparseness/structure of the con-
positive semi-definite (Q(t)  0); and the control pen- stituent plant and cost matrices—e.g., A(t), B(t), Q(t),
alty matrix, R(t), is positive definite (R(t) > 0). Thus, the and R(t)—will dictate the ease by which this can be
LQ optimal control problem is to find a control, u * (t) , accomplished. Qualitatively speaking, the level of dif-
such that the quadratic cost in Eq. 4 is minimized sub- ficulty involved in obtaining analytic solutions is related
ject to the constraint imposed by the linear dynamic primarily to the complexity of state equation coupling
system in Eq. 3. into the costate equations, most notably by the structure
To solve this continuous-time optimal control prob- of Q(t) and, to a lesser extent, by the structure of A(t). As
lem, one can use Lagrange multipliers,  (t) , to adjoin we will see later, however, this fact does not negate our
the (dynamic) constraints (Eq.  3) to the performance ability to derive effective guidance laws.
index (Eq. 4).1, 5 Consequently, an augmented cost func- Given a suitable system structure in Eqs. 7 and 8 (as
tion can be written as discussed above), one conceptually straightforward way
to solve this problem is to directly integrate the costate
1 2 equations backward in time from tf to t  t0 using the
Jl = 2 x (t f)
Qf
terminal costate conditions and then integrate the state
R1 2 2 V
S 2 ( x (t) Q (t) + u (t) )
R (t) W (5)
tf S W
+ # S +  T (t)(A (t) x (t) W dt . Table 1.  Necessary conditions for optimality.
t0
S W
+ B (t) u (t) – xo (t))   Condition Expression
T X
State
2 H (x (t), u (t), t)/2 (t) = xo (t), for t $ t 0
Referring to Eq. 5, we define the Hamiltonian function Equation
H as shown: Costate
2 H (x (t), u (t), t)/2 x (t) = –o (t), for t # t f
Equation
1 2 2
H (x (t), u (t), t) = 2 x (t) +1 u (t) Stationarity 2 H (x (t), u (t), t)/2 u (t) = 0, for t $ t 0
Q (t) 2 R (t) (6)
+  T (t)(A (t) x (t) + B (t) u (t)) . Boundary 2  (x (t f), t f)/2 x (t f) =  (t f), x (t 0) given

44 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

equations forward in time using the costate solutions From linear system theory, we know that the solution
and initial conditions on the states. This process can be to Eq.  13 can be expressed by the following matrix
done by hand or by using any available symbolic solu- exponential:
tion software (Maple, Mathematica, etc.). Another way W (t go) W (t go = 0)
to solve the two-point boundary value problem speci- = G = exp (t go)= G,
fied in Eq. 8 employs the sweep method.1 This technique Y (t go) Y (t go = 0)
(14)
assumes that the state x (t) and costate  (t) satisfy the –A BR –1 B T
linear relation shown below over the interval t ∈ [t0, tf ]  _= G.
Q AT
and given an (as yet) unknown matrix function Pc(t):

 (t) = Pc (t) x (t) . (9) In Eq. 14,  is the Hamiltonian matrix. With regard to
ease of determining analytic solutions, in an analogous
Using the assumed relation in Eq. 9, the control in Eq. 7 way to our previous discussion, the complexity of the
can be written as system structure (and the system order) will dictate how
difficult it may be to obtain an analytic solution to the
u (t) = –R –1 (t) B T (t) Pc (t) x (t) . (10) matrix exponential exp(tgo). Once an analytic solution
is found, the exponential matrix solution is partitioned
To find Pc(t) such that the control (Eq.  10) is com- as shown:
pletely defined, we differentiate Eq. 9 and make use of  11 (t go)  12 (t go)
the dynamic equations in Eq.  8. Doing so leads to a exp (t go) / = G. (15)
requirement to solve the following matrix Riccati dif-  21 (t go)  22 (t go)
ferential equation:
Using Eq.  15, the relation Pc(tgo) = Y(tgo)W–1(tgo),
–Po c (t) = Pc (t) A (t) + A T (t) Pc (t) Y(0)  Q f, W(0) = I, and the initial condition from
– Pc (t) B (t) R –1 (t) B T (t) Pc (t) (11) Eq. 12 (Pc(tgo = 0) = Q f), the solution to the matrix Ric-
cati differential equation Pc(tgo) can be expressed as
+ Q (t), Pc (t f) = Q f .

The optimal control is determined by first solving the Pc (t go) = 6 21 (t go) +  22 (t go) Q f@
(16)
6 11 (t go) +  12 (t go) Q f@ .
matrix Riccati differential equation backward in time –1
from tf to t and then using this solution in Eq. 10. There
are many ways to solve this equation. For completeness,
From Eq. 16, it becomes clear that the existence of Pc(tgo)
we present a matrix exponential method of solution.
is equivalent to having a nonsingular 11(tgo) + 12(tgo)Q f.
(We do not explore this important issue here, but the
Solving the Matrix Riccati Differential Equation via interested reader is referred to Refs. 1 and 4–11 for fur-
Matrix Exponential Method ther treatment of the subject.) Using Eq.  16 in Eq. 10,
the optimal control then is fully specified. Hand calcu-
We first want to rewrite Eq. 11 in terms of the time-
to-go variable defined as t go _ t f – t. We use the fact that lations (for systems of low order) or symbolic solution
dtgo/dt  =  –1 (for fixed tf) to express Eq.  11 in terms of software can be employed to mechanize this technique.
tgo as shown (note that plant matrices must be time-
independent for this technique):
THE PLANAR INTERCEPT PROBLEM
Po c (t go) = Pc (t go) A + A T Pc (t go) In general, the guidance process takes place in three-
dimensional space. However, such analysis can be com-
– Pc (t go) BR –1 B T Pc (t) + Q , (12)
plex and is beyond the scope of this article. Thus, here
Pc (t go = 0) = Q f . we will consider the formulation of the planar intercept
(pursuit-evasion) problem that we will use, subsequently,
Clearly, this matrix differential equation is quadratic in in the derivation of a number of different modern guid-
Pc(tgo) and is of dimension n. Note, however, that if we ance laws. This approach is not overly restrictive or
assume that Pc(tgo) takes the form Pc(tgo) = Y(tgo)W–1(tgo), unrealistic in that many (if not most) guidance law
then its solution may (instead) be found by solving the implementations, including modern ones, use the same
homogeneous linear matrix differential equation of approach, i.e., planar guidance laws are devised and
dimension 2n shown in Eq. 13: implemented in each of the maneuver planes. Figure 2
illustrates the planar (two-dimensional) engagement
d W (t go) –A BR –1 B T W (t go)
= G =
= G= G. (13) geometry and defines the angular and Cartesian quan-
dt go Y (t go) Q AT Y (t go) tities depicted therein. In Fig.  2, the x axis represents

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

45­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

M = Missile flight path angle aM = Missile acceleration, The homing kinematics can be expressed in vari-
normal to LOS
T = Target flight path angle
aT = Target acceleration, ous ways. Here, we concentrate on Cartesian forms
 = LOS angle normal to VT of expression. From a previous section, the target-to-
rM = Missile inertial position vector L = Lead angle
missile range was defined as R = r , and target–mis-
rT = Target inertial position vector rx = Relative position x (rTx – rMx)
vM = Missile velocity vector ry = Relative position y (rTy – rMy)
sile closing velocity was defined as Vc _ –Ro / –v : 1 r ,
vT = Target velocity vector R = Range to target where the LOS unit vector is 1 r = r/R . Thus, refer-
ring to Fig.  2, expressions for target–missile relative
yI /zI
position, relative velocity, and relative acceleration
T
vT
aT are given below, where we have defined the quantities
v M _ v M , a M _ a M , v T _ v T , and a T _ a T .
vM T
aM Target

r = r x 1 x + r y 1 y = 6R cos ()@ 1 x + 6R sin ()@ 1 y

(T)
R rT
L
v = v x 1 x + v y 1 y = 6" –v T cos (T) – v M cos (L + ) ,@ 1 x
 ry
M

+ 6" v T sin (T) – v M sin (L + ) ,@ 1 y (17)


Missile
(M) rM a = a x 1 x + a y 1 y = 6" a T sin (T) + a M sin () ,@ 1 x
xI + 6" a T cos (T) – a M cos () ,@ 1 y
Origin rx
(O)
Referring again to Fig.  2, we note that if the clos-
ing velocity is positive (Vc > 0), then we only need to
Figure  2.  Planar engagement geometry. The planar intercept
actively control the kinematics in the y/z coordinate to
problem is illustrated along with most of the angular and Carte-
achieve an intercept. That is, if Vc > 0 and the missile
sian quantities necessary to derive modern guidance laws. The
x  axis represents downrange while the y/z axis can represent
actively reduces and holds ry to zero by appropriately
either crossrange or altitude. A flat-Earth model is assumed with
accelerating normal to the LOS, then rx will continue
an inertial coordinate system that is fixed to the surface of the
to decrease until collision occurs. We will assume this
Earth. The positions of the missile (M) and target (T) are shown
condition holds and disregard the x components in the
with respect to the origin (O) of the coordinate system. Differen-
following analysis.
tiation of the target–missile relative position vector yields relative
The homing kinematics shown in Eqs. 17 are clearly
velocity; double differentiation yields relative acceleration.
nonlinear. In order to develop guidance laws using linear
optimal control theory, the equations of motion must
be linear. Referring to the expression for relative posi-
downrange, for example, while the y/z axis can repre- tion in Eqs.  17, note that for l very small, the y-axis
sent either crossrange or altitude, respectively (we will component of relative position is approximately given
use y below). For simplicity, we assume a flat-Earth by ry  Rl. Moreover, for very small T and l (e.g.,
model with an inertial coordinate system that is fixed near-collision course conditions), the y-axis compo-
to the surface of the Earth. Furthermore, we will assume nent of relative acceleration is approximately given by
that the missile and target speeds, v M and v T , ay  aT – aM. Hence, given the near-collision course
respectively, are constant. conditions, we can draw the linearized block diagram
In Fig. 2, the positions of the missile (pursuer) M and of the engagement kinematics as shown in Fig. 3. Cor-
target (evader) T are shown with respect to the origin (O) respondingly, we will express the kinematic equations
of the coordinate system, rM and rT , respectively. Thus,
the relative position vector [or line-of-sight (LOS) vector,
as it was previously defined] is given by r = rT – rM, and Relative Relative
the relative velocity vector is given by ro _ v = v T – v M . velocity position
+ vy ry 1
From Fig. 2, note that an intercept condition is satisfied aT ∫ ∫ 
R
if yt  = 0 and Vc > 0 (i.e., a collision triangle condition). Target –
acceleration
As illustrated, for this condition to be satisfied, the mis-
sile velocity vector must lead the LOS by the lead angle aM
Missile acceleration
L. Determining the lead angle necessary to establish a
collision triangle is, implicitly, a key purpose of the guid-
ance law. How this is done is a factor of many things, Figure  3. Linear engagement kinematics. Planar linear homing
including what measurements are available to the guid- loop kinematics are illustrated here. The integral of target–missile
ance system and what assumptions were made during relative acceleration yields relative velocity; double integration
formulation of the guidance law (e.g., a non-maneuver- yields relative position. The LOS angle, λ, is obtained by dividing
ing target was assumed). relative position by target–missile range.

46 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

of motion in state-space form. To this end, we define With these assumptions in mind, and referring to Eq. 18,
the state vector x _ 6x 1 x 2@T , the control u, the plant the LQ optimization problem is stated as
disturbance w, and the pseudo-measurement y where tf
x 1 _ ry, x 2 _ v y, u _ a M, w _ a T (any residual target min J ^x (t 0), u (t 0), t 0h = 1
2 x (t f) 2
+ 12 # u 2 (t) dt
acceleration is treated as a disturbance), and y _ Rl u (t) Qf t0
0 1 0
Subject to: xo (t) = ; E x (t) + ; E u (t),
(the linearized Cartesian pseudo-measurement is com-
(19)
posed of range times the LOS angle). Given these defini- 0 0 –1
tions, the kinematic equations of motion in state-space y (t) = 61 0@ x (t) .
form are written as
In words, find a minimum-energy control u(t) on the
xo (t) = Ax (t) + Bu (t) + Dw (t), x 0 = x (0)
time interval [t0,  tf] that minimizes a quadratic func-
y (t) = Cx (t) (18) tion of the final (terminal) relative position and relative
0 1 0 0
A=; E, B = ; E, C = 61 0@, D = ; E . velocity and subject to the specified dynamic con-
0 0 –1 1 straints. In Eq. 19, the terminal performance weighting
matrix, Q f, is yet to be specified. We define the scalar
These equations will form the basis for developing a b > 0 as the penalty on relative position at time tf (i.e.,
number of terminal homing guidance laws in the fol- final miss distance) and scalar c  0 as the penalty on
lowing subsections. In some cases, the equations are relative velocity at tf (c specified as a positive nonzero
modified or expanded as needed to reflect additional value reflects some desire to control or minimize the
assumptions or special conditions. terminal lateral relative velocity as is the case for a ren-
dezvous maneuver). Given the penalty variables, Q f is
the diagonal matrix given in Eq. 20:
DERIVATION OF PN GUIDANCE LAW VIA
b 0
LQ OPTIMIZATION Qf = ; E. (20)
0 c
In the previous article in this issue, “Basic Principles
of Homing Guidance,” a classical development of PN The choice of b and c is problem-specific, but for inter-
guidance was given based on that found in Ref. 12. In cept problems we typically let b → , c = 0, and for ren-
the present article, we will leverage off the discussion of dezvous problems we have b → , c → .
the previous subsection and develop a planar version of
the PN guidance law using LQ optimization techniques.
General Solution for Nonmaneuvering Target
To start, we will state the key assumptions used to
develop the guidance law: We can solve the design problem posed in Eq. 19 by
using the LQ solution method discussed previously. We
• We use the linear engagement kinematics model define t go _ t f – t and initially assume the scalar vari-
discussed previously and, hence, the state vector is ables b and c are finite and nonzero. Using Maple, we
x (t) = 6x 1 (t) x 2 (t)@T (the state vector comprises symbolically solve Eqs. 15 and 16 to obtain an analytic
the components of relative position and relative expression for Pc(t). We then use Pc(t) in Eq. 10 to obtain
velocity perpendicular to the reference x axis shown the following general guidance law solution:
in Fig. 2).
• All the states are available for feedback. u 1 (t) =
R V
S ` 1 + 2 ct go j x 1(t) + c 1 + 3 ct go + bt 2 m x 2(t) t go W
1 1 c
• The missile and target speeds, v M = v M and
3 S go W . (21)
v T = v T , respectively, are constant. 2 S ct go
1 + 3 ^1 + ct go h + 4
t go 3 W
• The missile control variable is commanded acceler- S bt go W
ation normal to the reference x axis (u = ac), which, T X
for very small LOS angles (l), is approximately per-
pendicular to the instantaneous LOS. We emphasize the fact that this general solution is
relevant given the assumptions enumerated above,
• The target is assumed to be nonmaneuvering particularly the fact that a nonmaneuvering target and
(aT = 0) and, therefore, the linear plant disturbance perfect interceptor response to commanded acceleration
is given to be w = 0. are assumed.
• The missile responds perfectly (instantaneously) to The structure of the guidance law u1(t), shown above,
an acceleration command from the guidance law can be expressed as
(aM  ac).The system pseudo-measurement is rela-
u (Q , t )/t 2 B z (x, Q , t ) ,
u 1 (t) = 8N
tive position y = x1. f go go f go

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

47­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

Target 250
direction
20.0 of travel Terminal penalties
b, c
Terminal penalties 200
b, c 0.1, 100
19.8
0.1, 100 100, 100

Acceleration (m/s2)
100, 100 100, 1
19.6 100, 1 150 100, 0.1
–0.2 0 0.2 0.4 0.6 100, 0
100, 0.1
100, 0
20 100

50
15
Downrange (km)

0
0 5 10 15 20
10 t go (s)

Figure  5.  Rendezvous acceleration. A plot of called-for accel-

erations for intercept trajectories with differing terminal relative
5
position and relative velocity penalties is shown here. The engage-
Start (missile trajectories) ment is the same as in Fig 4. As can be seen, substantially more
acceleration is required as the penalty on the terminal relative
0 velocity is increased.
–0.5 0 0.5 1.0 1.5
Crossrange (km)
able states for feedback. From above, we also see that
Figure 4.  Rendezvous trajectory. A plot of intercept trajectories time-to-go (tgo) is needed in the guidance law. Later, we
(in downrange and crossrange) are shown with different terminal also discuss how one can estimate time-to-go for use in
penalties for relative position and relative velocity, respectively. the control solution.
The target trajectory is traveling from right to left (shown as the Employing Eq.  21 in a simple planar engagement
dashed black line) at the top. The target velocity is a constant simulation, we can show what effect a terminal rela-
500 m/s. The initial missile velocity is 1000 m/s with no heading tive velocity penalty will have (in combination with a
error. The final time is 20  s. The inset highlights the endgame terminal relative position penalty) on the shape of the
geometries in each case. Notice that as the terminal penalty missile trajectory. In Fig. 4, a missile is on course to inter-
on relative velocity is increased, the missile trajectory tends to cept a constant-velocity target; it can be seen that the
“bow out” such that the final missile velocity can be aligned with missile trajectory shape changes for different terminal
the target velocity. Similarly, as the terminal penalty on relative relative position and velocity penalties (i.e., variations
position increases, the final miss distance is reduced. in the penalty weights b and c). Note that, in this simple
example, there is no assumption that controlled missile
where the quantity acceleration is restricted to the lateral direction. Figure 5
shows the resulting acceleration history for each case.
3 ^ 1 + ct go h + 4
ct go
u (Q , t ) _ 3/ c 1 +
N 3
m
Note that with a nonzero terminal penalty on relative
f go bt go position in combination with no penalty on relative
velocity (the blue curve in Fig. 5) the missile does not
is the effective navigation ratio and z^x, Q f, t go h com- maneuver at all; this is because the target maneuver
prises the remainder of the numerator in Eq. 21. Thus, assumption implicit in the guidance law derivation (i.e.,
for this general case, the effective navigation ratio is not the target will not maneuver) happens to be correct in
a constant value. this case, and the missile is already on a collision course
It is clear that mechanization of this guidance law at the start of the engagement.
requires feedback of the relative position and relative
velocity states to compute the control solution. Typically, Special Case 1: PN Guidance Law
only the relative position pseudo-measurement is avail- If we assume that c = 0 and we evaluate lim u 1 (t) =
able, formed using measurements of LOS angle and b"3
uPN(t), then Eq. 21 collapses to the well- known Carte-
range (assuming a range measurement is available). In sian form of the PN guidance law:
our companion article in this issue, “Guidance Filter
Fundamentals,” we discuss filtering techniques that 6x (t) + x 2 (t) t go@ .
3
u PN (t) = (22)
t 2go 1
enable us to estimate any unmeasured but observ-

48 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

Leveraging the discussion above for Eq. 21, we can see homing portion of the engagement. A planar (down-
that the effective navigation ratio in Eq. 21 has now col- range and crossrange) Simulink terminal homing sim-
lapsed to become N u = 3. ulation was used to examine the PN performance. To
PN
In Eq. 22, the quantity in square brackets now repre- avoid layering of complexity in the results, all simula-
sents the miss distance that would result if the missile tion noise sources (seeker noise, gyro noise, etc.) were
and target did not maneuver over the time period [t, tf], turned off for this study. However, a simplified guidance
which often is referred to as the zero-effort-miss filter still was used in the loop to provide estimates of
(ZEM). We want to emphasize that the guidance law relative position and velocity to the PN guidance law,
ZEM is an estimate of future miss that is parameterized thereby having some effect on overall guidance per-
on the assumptions upon which the guidance law was formance. Figure  6a illustrates nominal missile and
derived (linear engagement, constant velocity, non- target trajectories in a plot of downrange versus cross-
maneuvering target, etc.). Hence, the PN ZEM (in the range. The total terminal homing time is about 3 s. As
y axis, for example) is given by ZEMPN = ry(t) + vy(t)tgo. is evident, at the start of terminal homing, the missile
Note that the accuracy of the guidance law ZEM esti- is traveling in from the left (increasing downrange)
mate is directly related to how well the guidance law will and the target is traveling in from the right (decreas-
perform in any particular engagement (actual final miss ing downrange). The missile, under PN guidance, is
distance, maximum commanded acceleration, amount initially on a collision course with the target, and the
of fuel used, etc.). target is, initially, non-maneuvering. At 2 s time-to-go,
Under the current stated assumptions, we can show the target pulls a hard turn in the increasing crossrange
that Eq.  22 is equivalent to the traditional LOS rate direction.
expression for PN, which was shown in the previous Recall that PN assumes that the missile acceleration
article in this issue, “Basic Principles of Homing Guid- response to guidance commands is perfect or instan-
ance,” to be a Mc = NVc lo . We first differentiate ry  Rl taneous (i.e., no-lag), and that the target does not
to obtain v y . Ro l + Rlo . Next, recalling that Ro = maneuver. Thus, we will examine the sensitivity of PN
v : 1 r = –Vc and noting that we can express range-to- to these assumptions by simulating a non-ideal mis-
go as R = Vctgo, we have the following relationship for sile acceleration response in addition to target maneu-
LOS rate: ver. We will parametrically scale the nominal missile
ry + v y t go autopilot time constant (100 ms for our nominal case),
lo = . (23) starting with a scale factor of 1 (the nominal or near-
Vc t 2go perfect response) up to 5 (a very “sluggish” response
of approximately 500 ms) and examine the effect on
Examining the traditional LOS rate expression for PN, guidance performance. Figure  6b illustrates the accel-
as well as Eqs. 22 and 23, it can be seen that if we set eration step response of the interceptor for 100-, 300-,
N = 3, then the traditional LOS rate expression for and 500-ms time constants. Figures 6c and 6d pres-
PN and Eq.  22 are equivalent expressions. Hence, the ent the simulation results for levels of target hard-turn
“optimal” PN navigation gain is N = 3. acceleration from 0 to 5  g. Figure  6c shows final miss
distance versus target hard-turn acceleration level for
Special Case 2: Rendezvous Guidance Law the three different interceptor time constants. As can
be seen, as the missile time response deviates from
For the classic rendezvous problem, we desire to con-
the ideal case, guidance performance (miss distance)
verge on and match the velocity vector of the target at
degrades as target maneuver levels increase. Figure  6d
the final time tf. If we evaluate lim u 1 (t) = u REN (t), illustrates the magnitude of missile total achieved accel-
b, c " 3
then Eq. 21 collapses to the Cartesian form of the ren- eration for variation of target maneuver g levels shown
dezvous (REN) guidance law and for the three missile time constants considered.
As can be seen, when the missile autopilot response
u REN (t) = 62 8x 1 (t) + 23 x 2 (t) t goB . (24) time deviates further from that which PN assumes,
t go
respectively higher acceleration is required from the
missile; this is further exacerbated by higher target
Example Results g levels.
The example in this section is illustrated by Fig.  6,
where we demonstrate how PN performance degrades if EXTENSIONS TO PN: OTHER OPTIMAL HOMING
(i) the interceptor response is not ideal (i.e., the actual
interceptor response deviates from that assumed in the GUIDANCE LAWS
derivation of PN) and (ii) the target performs an unan- In Eq. 21, a general guidance law that is based on
ticipated maneuver some time during the terminal some specific assumptions regarding the (linearized)

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

49­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

engagement kinematics, 2

Crossrange (km)
(a)
target maneuver (actually,
a lack of it), and intercep- 1
tor response characteristics Missile Target
(the perfect response to an
acceleration command) is 0
0 500 1000 1500 2000 2500 3000 3500 4000
shown. We also showed that Downrange (m)
the general optimal guid-

Step response (g)

1.5
ance law derived there col- (b) Step input command
lapses to the well-known PN 1.0
guidance law if we assign
0.5
cost function components as
c = 0 and b →  (see Eqs. 19 0
and 20). For completeness’ –0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
sake, here we will derive a Time (s)
number of related optimal Autopilot time constant
0.1 s 0.3 s 0.5 s
guidance laws under differ- 3.0 40
ing assumptions regarding (c) (d)
target maneuver and inter- 2.5
Final miss distance (m)

ceptor response models. 30

For certain scenarios, these

Acceleration (g)
2.0
assumptions can have a
significant impact on guid- 1.5 20
ance law performance. For
example, in Example Results, 1.0
it was observed that, during 10
the endgame, PN (which is 0.5
derived assuming no target
maneuver) can call for sig- 0 0
0 1 2 3 4 5 0 1 2 3 4 5
nificant acceleration from Target maneuver (g) Target maneuver (g)
the interceptor when pitted
against a target pulling a Figure 6.  PN performance versus time constant. (a) A planar missile–target engagement with a
hard-turn maneuver. Hence, plot of downrange versus crossrange. The total terminal homing time is approximately 3 s. At the
one can develop a guidance start of terminal homing, the missile is traveling in from the left (increasing downrange) and the
law that specifically takes target is traveling in from the right (decreasing downrange). (b) The acceleration step response of
target acceleration into the missile interceptor for 100-, 300-, and 500-ms time constants. (c and d) Simulation results for
account. Of course, mecha- a PN-guided missile versus varying levels of target hard-turn acceleration from 0 to 5 g. (c) Final
nization of such a guidance miss distance versus target hard-turn acceleration level for the three different interceptor time
law will be more complex constants. The graph illustrates the fact that, for non-ideal interceptor response, PN-homing guid-
and will require the feed- ance performance degrades with increasing target maneuver levels. This degradation worsens as
back of additional informa- the autopilot response becomes more sluggish. Notice that as the autopilot response approaches
tion (e.g., target maneuver). the ideal case (blue curve), the miss distance becomes nearly insensitive to target maneuver. For
In addition, if the derivation an ideal autopilot response, PN-homing would result in acceleration requirements of three times
assumptions become too spe- the target maneuver. (d) The magnitude of total achieved missile acceleration for the same varia-
cific, the resulting guidance tion of target maneuver g levels and for the three missile time constants considered. As the missile
law may work well if those autopilot response time deviates further from that which PN assumes, increasingly higher accel-
assumptions actually hold, eration levels are required from the missile.
but performance might rap-
idly degrade as reality devi-
ates from the assumptions.
exception of target maneuver; we will assume that the target is pulling a hard-turn
maneuver (i.e., constant acceleration in a particular direction). Therefore, we aug-
Constant Target ment the previous (PN) state vector (where x 1 _ ry, x 2 _ v y ) to include a target
Acceleration Assumption acceleration state x 3 _ a Ty, leading to x _ 6x 1 x 2 x 3@T . As before, the control u is
Here, all of the previous missile acceleration ^u _ a Mh ; the plant (process) disturbance is not considered when
assumptions hold with the deriving the guidance law (w = 0), but it will come into play when developing a target

50 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

acceleration estimator; and the pseudo-measurement, again, is relative position

^y _ x 1h . With these modeling assumptions in mind (particularly that x 3 _ a Ty
constant), the LQ optimization problem is stated as
tf
min J ^x (t 0), u (t 0), t 0h = 1
2 x (t f) 2
+ 12 # u 2 (t) dt
u (t) Qf t0
0 1 0 0
Subject to: x (t) = >0
o 0 1H x (t) + >–1H u (t) , (25)
0 0 0 0
y (t) = 61 0 0@ x (t) .

As before (Eq.  20), the terminal penalty matrix is defined as Q f  =  diag{b,  c,  0}.
Following a solution procedure identical to that outlined previously, we obtain the
following general solution:

R 2 V
S ` 1 + 2 ct go j x 1 (t) + c 1 + 3 ct go + bt 2 m x 2 (t) t go + 2 c 1 + 6 ct go + m x 3 (t) t go W
1 1 c 1 1 2c
2
bt go
3
u 2 (t) = 2 S
go W. (26)
1 + 33 ^1 + ct go h + 4
ct go
t go S W
S bt go W
T X

If we compare the guidance law in Eq.  26 to our previous result (Eq.  21), it is
clear that the only difference is in the numerator; Eq. 26 includes the addition of a
(time-varying) gain multiplying the target acceleration state. Therefore, in addition to
requiring estimates of relative position, relative velocity, and time-to-go, the new law
also requires an estimate of target acceleration. This requirement has implications on
the guidance filter structure, as we will see later. More important, estimating target
acceleration given a relative position (pseudo-)measurement can be very noisy unless
the measurement quality is very good. Hence, if sensor quality is not sufficient, lead-
ing to a target acceleration estimate that is too noisy, then guidance law performance
could actually be worse than if we simply used PN to begin with. These factors must
be carefully considered during the missile design phase.

Special Case 1: Augmented PN

If we assume that c = 0 and we evaluate lim u 2 (t) = u APN (t), then Eq. 26 col-
b"3
lapses to the well-known augmented PN (APN) guidance law:

u APN (t) = 32 8x 1 (t) + x 2 (t) t go + 12 x 3 (t) t 2goB . (27)

t go

Leveraging the discussion above, if we compare Eq. 27 to the expression for PN given
in Eq. 22, the only difference is the addition of the 12 a T t 2go term in the numerator
of the APN guidance law. Thus, for APN, the effective navigation ratio is the same
as in PN guidance ( N u = 3 ), but the ZEM estimate is now given by ZEM APN =
APN
1
ry (t) + v y (t) t go + 2 a T (t) t go
2 .

Special Case 2: Augmented REN Guidance Law

If we evaluate b,lim u (t) = u AREN (t), then Eq.  26 collapses to the augmented
c"3 2
REN (AREN) guidance law:

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

51­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

u AREN (t) = 62 8x 1 (t) + 23 x 2 (t) t goB + x 3 (t) . (28)

Guidance gain Guidance law

t go N=3 N=5 PN

Normalized missile acceleration (aC /aT)

N=4 N=6 APN
3.0
Notice that, unlike the APN law given in Eq. 27, the
AREN guidance law calls for a direct cancellation of the 2.5
target maneuver term in the acceleration command. 2.0

Example Results: Maneuver Requirements for 1.5

PN Versus APN 1.0
Here, in a similar approach to that found in Ref. 13,
we compare the maneuver requirements for PN and 0.5
APN guidance laws versus a target that pulls a hard-turn
0
maneuver under ideal conditions (e.g., no sensor mea- 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
surement noise or interceptor lag). Recall that the PN Normalized time (t/tf )
guidance law is derived assuming that the target does
not maneuver, whereas the key APN assumption is that Figure  7.  A comparison is made of PN and APN accelera-
the target pulls a hard turn. tion requirements for various guidance gain values versus a
Assuming that the interceptor responds perfectly to target pulling a hard turn. PN results (solid lines) indicate that
PN acceleration commands, using Eq. 22 we can write PN demands an increasing level of missile acceleration as flight
the following second-order differential equation: time increases. Notice that for a guidance gain of 3, PN requires
three times the acceleration of the target to effect an intercept;
d 2 r (t) = a (t) – N [r (t) + v (t)(t – t)] . (29) hence, the well-known 3-to-1 ratio rule of thumb. Increasing the
y T y y f guidance gain can theoretically relax the 3-to-1 rule of thumb,
dt 2 (t f – t) 2
but higher gains may lead to excessive noise throughput. The
graph illustrates that APN (dashed lines) is more anticipatory in
Note that we have left the navigation gain as the that it demands maximum acceleration at the beginning of the
variable N. Next, we assume zero initial conditions— engagement and less as engagement time increases. Note that for
ry(0) = 0, vy(0) = 0—and use Maple to solve Eq. 29, thus a guidance gain of 3, APN (theoretically) requires half the accel-
giving analytic expressions for ry(t) and vy(t). We then eration that PN does when engaging a target that is pulling a
take ry(t) and vy(t) and reinsert them into Eq.  22 to hard turn.
obtain the following expression for missile acceleration
caused by a hard-turn target maneuver:
3-to-1 rule of thumb can be relaxed. In practice, how-
a M (t) = N =1 – c 1 – t m G a T .
N–2
(30) ever, higher gains may lead to excessive noise through-
PN N–2 tf put, perhaps negating any theoretically perceived
benefits.
For the APN case, we employ an analogous proce- Unlike PN, APN (dashed lines in Fig.  7) is more
dure, using Eq. 27 as the starting point. For this case, we anticipatory in that it demands maximum acceleration
obtain the following expression for missile acceleration, at the beginning of the engagement and less accelera-
given that the target pulls a hard-turn maneuver: tion as engagement time increases. Moreover, note that
for a guidance gain of 3, APN (theoretically) requires
N–2 half the acceleration that PN does when engaging
aM (t) = N ;1 – t E aT . (31) a target that is pulling a hard turn. Note that, for a
APN 2 tf
guidance gain of 4, the theoretical acceleration require-
ments for PN and APN are the same and, for gains
Figure  7 illustrates a comparison of PN and APN above  4, APN demands more acceleration than PN
acceleration requirements for various guidance gain (although saturating the acceleration command early is
values versus a target pulling a hard turn, via Eqs.  30 not nearly the problem that saturating late is).
and 31, respectively. Referring to Fig.  7, we see that
PN (solid lines) demands an increasing level of mis-
sile acceleration as flight time increases. In fact, for a Constant Target Jerk Assumption
guidance gain of 3, PN requires three times the accel- The assumptions stated during the deriva-
eration of the target to effect an intercept; hence, the tion of APN are still valid with the exception
well-known 3-to-1 ratio rule of thumb. If we increase of target maneuver; here, we will assume that
the guidance gain, theory says that the theoretical the target acceleration is linearly increasing (i.e.,

52 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

constant jerk in a particular direction). This may be a reasonable assumption,

for example, in order to develop a terminal homing guidance strategy for use
during boost-phase ballistic missile defense, where it is necessary to engage and
destroy the enemy missile while it is still boosting. In such a context, it is pos-
sible that a linearly increasing acceleration model (assumption) better reflects actual
target maneuver (acceleration) as compared to the APN assumption. Therefore,
we augment the previous (PN) state vector (where x 1 _ ry, x 2 _ v y) to include a
target acceleration state x 3 _ a Ty and a target jerk state x 4 _ xo 3 = j Ty, leading to
x _ 6x 1 x 2 x 3 x 4@T . As before, the control u is missile acceleration (u _ a M), the
plant (process) disturbance is not considered (w = 0), but it will come into play when
developing a target state estimator, and the pseudo-measurement, again, is rela-
tive position (y _ x 1). With these modeling assumptions in mind (particularly that
x 4 _ j Ty / constant), the LQ optimization problem is stated as
tf
min J ^x (t 0), u (t 0), t 0h = 1
2 x (t f) 2
+ 12 #
u 2 (t) dt
u (t) Qf t0
R0 1 0 0V R0V
S W S W
Subject to: xo (t) = SS
0 0 1 0W
x (t ) + S–1W u (t), (32)
0 0 0 1W S0W
S W S W
T0 0 0 0X T0X
y (t) = [1 0 0 0] x (t).

Defining the terminal penalty matrix to be Q f  =  diag{b,  c,  0,  0}, and following
a solution procedure identical to that outlined previously, we obtain the following
general solution:

R V
S `1 + j x 1 (t) + c 1 +
ct go ct go ct go
m x 2 (t) t go + 2 c 1 + m x 3 (t) t go + 6 c 1 + m 4
c 1 2c 2 1 3c 3 W
2 3 + 2 6 + 2 2 x (t ) t go
3 S bt go bt go bt go W
u 3 (t) = 2 S ct go W . (33)
^1 + ct go h + 4
t go S 1+ 3 W
S bt 3go W
T X

If we compare the guidance law in Eq.  33 to our previous result (Eq.  26), we see,
again, that the only difference is in the numerator; Eq. 33 includes the addition of a
(time-varying) gain multiplying the target jerk state. Analogous to the previous case,
this has (additional) implications on the guidance filter structure. More important,
estimating target jerk given a relative position (pseudo-)measurement can be very
noisy unless the measurement quality is excellent. If sensor quality is not sufficient,
then guidance law performance could be significantly worse than if we simply used
PN or APN to begin with.

Special Case 1: Extended PN

If we assume that c = 0 and we evaluate lim u 3 (t) = u EPN (t) , then Eq. 33 collapses
b "3
to the extended PN (EPN) guidance law

u EPN (t) = 32 8x 1 (t) + x 2 (t) t go + 12 x 3 (t) t 2go + 16 x 4 (t) t 3goB . (34)

t go

By this time, a pattern should be emerging regarding the current line of guid-
ance laws. For example, if we compare PN (Eq. 22), APN (Eq. 27), and EPN (Eq. 34),
we see that the effective navigation ratios for these three cases are all the same
constant:

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

53­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

Truth Model Estimate Figure  8.  Comparison of ZEM

4
removal against a boosting
PN threat. Plots of the predicted
2
ZEM for PN, APN, and EPN guid-
ance laws are illustrated. Here,
“truth” is computed by integrat-
0 ing forward a duplicate of the
ZEM (km)

threat truth model and propa-

2
gating the kill vehicle ballisti-
APN cally. The “model” curve reflects
the assumed ZEM calculation
0
for the particular guidance
2 law, using truth data. The “esti-
EPN mate” curve is this same calcu-
lation with estimated data (via
0
0 5 10 15 noisy measurements), as must
t go(s) be used in practice. Intercept
is achieved at tgo  =  0 in each
case, since the ZEM goes nearly to zero. The magnitude of the ZEM is plotted; in each plot, the portion of the model-based ZEM and
estimated ZEM to the right of 12 s is in the opposite direction from the true ZEM.

u
N u
=N u
=N = 3. estimate quality (noise) as more derivatives are used
PN APN EPN
in the calculation. Figure 9 demonstrates the resulting
It is the ZEM estimates that evolve from acceleration commands and fuel usage for the different
guidance laws via logging of the resultant DV, which is
ZEM PN = ry (t) + v y (t) t go , to defined as
tf
ZEM APN = ZEMPN + 21 a T (t) t go
2 , and now to #t a M () d ,
0

3
ZEMEPN = ZEMAPN + 16 j T (t) t go . where a M is the achieved acceleration vector. The
required DV (translating to fuel usage) for PN, at
1356  m/s, is substantially more than APN, at 309  m/s,
With regard to EPN, the addition of target accelera-
or EPN, at 236 m/s. The required acceleration capabil-
tion and target jerk states will dictate a more complex
ity of the kill vehicle also is substantially different, with
guidance filter structure, and it may be very sensitive
PN requiring 27 g capability, APN requiring 3.7 g, and
to sensor noise and actual target maneuver modali-
EPN requiring 3.2 g.
ties as compared with PN or APN. We also note that,
if we evaluate b,lim u (t), then Eq.  33 collapses to
c"3 3 PN APN EPN
the AREN guidance law previously given in Eq. 28.
A simulation study of PN, APN, and EPN guid- 200
ance laws against a boosting threat illustrates the
|am| (m/s)

benefits of better matching the target assumptions to 100

the intended target. In this engagement, an exoatmo-
spheric kill vehicle intercepts a threat that accelerates 0
(boosts) according to the ideal rocket equation, with a 1000
maximum 8.5-g threat acceleration occurring at inter-
V used (m/s)

cept. Figure 8 shows how the ZEM prediction for each

500
guidance law compares with the true ZEM as each
evolves over time and under the control of the relevant
guidance law. Each of the models of ZEM has signifi- 0
cant error initially (the direction is wrong, causing the 0 5 10 15
t go (s)
truth to increase while the assumed ZEM decreases),
but for an assumed target maneuver model that more Figure  9.  Acceleration command history and cumulative ∆V
closely matches the actual target maneuver (i.e., EPN), used for PN, APN, and EPN guidance laws versus a boosting
this error is much less and the sign of the ZEM is cor- threat. The progression is from right to left as time-to-go (tgo)
rect sooner. The curves also show some trade-off in decreases toward intercept in this single-simulation run.

54 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

Non-Ideal Missile Response Assumption

Here, the assumptions stated during the derivation of APN are still valid, save for
that pertaining to a perfect missile response. Instead, we will add the more realistic
assumption that the missile responds to an acceleration command via the first-order
lag transfer function
a M (s)/a c (s) = 1 ,
Ts + 1
hence the reference to a non-ideal missile response. The time constant, T, is a compos-
ite (roll-up) function of the missile response at a specific flight condition and depends
largely on the missile aerodynamic characteristics and flight control system design.
We augment the previous (APN) state vector (where x 1 _ ry, x 2 _ v y, and x 3 _ a Ty
) to include a missile acceleration state x 4 _ a My , leading to x _ 6x 1 x 2 x 3 x 4@T .
(Note that the fourth state here is missile acceleration, not target jerk as was the case
when deriving EPN guidance law.) As before, the control u is missile acceleration
(u _ a M), the plant (process) disturbance is not considered (w = 0), but it will come into
play when developing a target acceleration estimator, and the pseudo-measurement,
again, is relative position (y _ x 1). In addition, we add a missile accelerometer mea-
surement. With these modeling assumptions in mind, the LQ optimization problem is
stated as
tf
min J ^x (t 0), u (t 0), t 0h = 1
2 x (t f) 2
+ 12 #
u 2 (t) dt
u (t) Qf t0
R0 1 0 0 V R0V
S W S W
S0 0 1 –1 W S0W
Subject to: xo (t) = S W xr (t) + S W u (t), (35)
0 0 0 0 0
S W S1W
S0 0 0 – T1 W STW
T X T X
1 0 0 0
y (t) = ; E xr (t) .
0 0 0 1

Here, we do not consider a terminal velocity penalty in order to reduce over-

all guidance law complexity, which leads to the terminal penalty matrix given by
Q f = diag{b, 0, 0, 0}. Thus, following an identical solution procedure to that outlined
previously, we obtain the following general guidance law solution:

6` j ` + e –t go /T – 1 j8x 1 (t) + x 2 (t) t go + 12 t go x 3 (t) – T 2 ` + e –t go/T – 1 j x 4 (t)B

t go 2 t go t go
2
T T T
u4 = 2
. (36)
+ c3 + 6 –2t go /T m
t go t go t 2go t 3go t go
6 –t go /T
T –6 +2 – 12 T e – 3e
bT 3 T2 T3

Upon examination of Eq. 36, it becomes clear that the non-ideal missile response
assumption adds additional complexity to the guidance law structure (remember that
we have not considered a terminal penalty on relative velocity, i.e., c = 0). To better
visualize this complexity, consider the constant target acceleration guidance law
given in Eq. 26. If we take lim u 2 (t) , we obtain the following result:
c"0

x 1 (t) + x 2 (t) t go + 12 x 3 (t) t 2go

u 2 (t) = 32 > H. (37)
c=0 t go 1 + 33
bt go

The structure of Eq. 37 is significantly less complex than that given in Eq. 36 despite
the fact that the cost function for both is identical (i.e., b is finite and c = 0).
If we take blim
"3 4
u (t) , we obtain the well-known “optimal” guidance law (OGL)
referred to in many texts (see Refs. 7, 13, and 14 for examples):

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

55­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

Guidance law
Guidance law Gain
PN APN OGL
Target acceleration
KaT 3 3 Nu
Kr
t 2go t 2go t 2go
Relative velocity 3 3 Nu
Kv Kv
vy t go t go t go

aT + ry + + Missile 3 u
N
–
 ∫ ∫ Relative
Kr +  response
KaT 0
2 2
–
position u 2 t go
KaM 0 0
–NT c + e –t go /T – 1 m
KaM t 2go T
t go 2 t go
m c 6c
+ e –t go /T – 1 m
u
N _ T T
OGL
t t2 t3 t
e 3 + 6 go – 6 go2 + 2 go3 – 12 go e –t go /T – 3e –2t go /T o
Missile acceleration
T T T T

Figure  10.  The feedback structure of the PN, APN, and OGL guidance laws is depicted here. The relative complexity of the different
guidance laws is established as we add additional assumptions regarding the engagement and missile response characteristics. The
diagram emphasizes the fact that a substantial increase in complexity arises when the assumptions move from an ideal to non-ideal
interceptor response assumption.

6` j ` + e –t go /T – 1 j8x 1 (t) + x 2 (t) t go + 12 t 2go x 3 (t) – T 2 ` + e –t go /T – 1 j x 4 (t)B

t go 2 t go t go
T T T
u OGL = 2
. (38)
c3 + 6 –2t go /T m
t go t go t 2go t 3go t go
–t go /T
T –6 +2 – 12 T e – 3e
T2 T3

Referring back to the APN law presented in Eq. 27, a couple of important points are
noted. First, we compare the APN ZEM estimate given by

ZEM APN = x 1 (t) + x 2 (t) t go + 12 t 2go x 3 (t)

with that in Eq. 38 and see that the OGL ZEM estimate is

ZEM OGL = ZEM APN – T 2 ` + e –t go /T – 1 j x 4 (t) .

t go
T
u
In addition, the effective navigation ratio for APN is given by N = 3. In contrast,
APN
u
from Eq. 38, N OGL is time-varying and can be expressed as shown below:

6` j ` + e –t go /T – 1 j
t go 2 t go
u T T
N _ . (39)
OGL
c3 + 6 –2t go /T m
t go t 2go t 3go t go
–t go /T
T –6 +2 – 12 T e – 3e
T2 T3

For illustrative purposes, Fig. 10 depicts the feedback structure of the PN, APN,
and OGL guidance laws discussed this far and thus helps to establish the relative
complexity of the different guidance laws as we add additional assumptions regarding
the engagement and missile response. From Fig. 10, it is obvious that a substantial
increase in complexity arises when the assumptions move from an ideal to non-ideal
interceptor response assumption.

Example Comparisons of PN, APN, and OGL

In this example, Fig.  11 illustrates the miss distance and called-for acceleration
statistics for PN, APN, and OGL guidance laws versus a target pulling a 5-g hard turn.
The Monte Carlo data are displayed in cumulative probability form. From Fig. 11, we

56 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

APN OGL PN Figure  11.  The cumulative prob-

100 ability performance of PN, APN,
and OGL guidance laws versus a 5-g
80
Probability (%)

hard-turn target is shown. Both the

60 cumulative probability of the maxi-
40 mum interceptor acceleration and
CPA for 100-run Monte Carlo sets are
20 plotted. With these graphs, it is easy
0 to quickly ascertain the probability
0 0.33 0.66 0.98 1.31 1.64 1.97 2.30 2.62 2.95 of achieving the x-axis parameter
Closest point of approach (ft) value (e.g., maximum acceleration).
100 In both graphs, lines that are more
vertical and farther to the left are
80
Probability (%)

considered more desirable. All noise

60 and error sources are turned off, but
the target maneuver start time is ran-
40
domized over the last second of ter-
20 minal homing. Note that the actual
missile response model is non-ideal,
0
0 5 10 15 20 25 30 35 40 45 and it varies with the flight condi-
Maximum interceptor acceleration (g) tions of the missile trajectory.

see a distinct trend considering PN versus APN versus OGL guidance laws. Given values of terminal miss pen-
that PN is derived assuming the target is not maneuvering, we expect the perfor- alty. The curves are normal-
mance to degrade if the target does maneuver. It also is not surprising to see that, ized with respect to the missile
statistically, more acceleration is required versus APN or OGL. The improvement time constant T. Referring to
from APN to OGL is explained by the fact that APN assumes a perfect missile Fig.  12, consider the b  =  1000
response to acceleration commands and OGL assumes that the missile responds as curve. As t go /T " 2 , Ñ achieves
a first-order lag (refer back to Eq. 27 compared with Eq. 38). its maximum value and then
reduces as t go /T " 0 . Clearly,
this guidance law gain curve
TIME-TO-GO ESTIMATION evolves in a way that places
As shown in the previous section, Extensions to PN: Other Optimal Homing much greater emphasis on ZEM
Guidance Laws, many modern guidance laws require an estimate of time-to-go (tgo), at certain times near intercept.
which is the time it will take the missile to intercept the target or to arrive at the Imagine that the actual (true)
closest point of approach (CPA). The tgo estimate also is a critical quantity for mis- tgo is 2  s but that the estimate
siles that carry a warhead that must detonate when the missile is close to the target. of tgo is in error and biased
For example, recall the general optimal guidance law shown in Eq. 36. This guid- toward the positive direction
ance law can be expressed as Ñ 3 ZEM, where by four missile time constants
(4T). Then, from Fig.  12 we
ZEM = 8x 1 (t) + x 2 (t) t go + 12 t 2go x 3 (t) – T 2 ` + e –t go /T – 1 j x 4 (t)B
t go can see that the guidance gain
T would be about one-seventh of
what it should be at t go /T = 2 ,
and the effective navigation ratio, Ñ, is shown in Eq. 40, where b and T represent thereby not placing optimal
the terminal miss penalty and assumed missile time constant, respectively (see the emphasis on ZEM at that time,
section Non-Ideal Missile Response Assumption, wherein Eq.  36 was derived, for a and degrading overall guidance
further description): performance.
The simplest time-to-go
6` j ` + e –t go /T – 1 j
t go 2 t go estimation scheme uses mea-
u = T T surements (or estimates) of
N . (40)
+ c3 + 6 e –t go /T – 3e –2t go /T m
t go t 2go 3
t go t go
6
–6 +2 – 12 range and range rate. Consider
bT 3 T T2 T3 T
the engagement geometry of
Fig.  13, where v M   =  missile
It is clear from Eq.  40 that Ñ is a function of tgo. Figure  12 illustrates the tgo velocity, v T   = target veloc-
dependence of the general optimal guidance law effective navigation ratio for three ity, r  =  target–missile relative

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

57­­
N.  F.  PALUMBO,  R.  A.  BLAUWKAMP,  and  J.  M.  LLOYD

6 c t go + e –t go /T – 1 m P
u _ T2 T
N 2 3 CPA
6 + e 3 + 6 go – 6 t go + 2 t go – 12 t go e –t go /T – 3e –2t go /T o
t
bT 3 T T 2
T3 T
vM
v

Terminal penalty
80 b =10 b = 100 b = 1000 vT
70  R = || r||
60 Missile LOS Target
Guidance gain

50

40 Figure  13.  The missile–target (planar) engagement geometry

30
is shown here. This depiction places the LOS to the target along
the x axis of the coordinate system. Missile and target velocity
20 – and v
– , respectively. The relative veloc-
vectors are indicated as v M T
10 –
ity, v, makes an angle  with the LOS and passes through the
0 points indicated by CPA and P.
0 1 2 3 4 5 6 7 8 9 10
t go/T (s)
r (t) : v (t)
t go = – . (44)
Figure  12.  The normalized tgo dependence of the effective v (t) : v (t)
navigation ratio Ñ for the OGL is shown here for three terminal
penalty values. Normalization is with respect to the missile time Using this expression for tgo in Eq.  42, we obtain the
constant T. We note that, as tgo approaches infinity, the effective target–missile relative separation at the CPA:
navigation ratio always approaches 3.
r (t)[ v (t) : v (t)] – v (t)[ r (t)(t) : v (t)]
rCPA =
v (t) : v (t)
position, and v  = target–missile relative velocity. Refer- (45)
ring to Fig. 13, we assume that the missile can measure [ v (t) # r (t)] # v (t)
= .
or estimate relative range (R = r ) to the target and v (t) : v (t)
range rate (Ro ) along the LOS to the target. If we assume
that the missile and target speeds are constant, then one Conceptually, the differences between time-to-
can estimate time-to-go as go estimation using Eq.  41 rather than Eq.  44 can be
explained by using Fig. 13. For this discussion, and with-
tgo = – R . (41) out loss of generality, we assume that the missile velocity
Ro is constant and that the target is stationary. Referring to
Another common approach to estimating time-to-go Fig. 13, we see that Eq. 41 estimates the flight time for
also assumes that the missile and target speeds are con- the missile to reach point P. However, Eq. 44 estimates
stant. Define D t _ t ) – t , where t is the current time and the time it takes for the missile to reach the CPA. If
t ) is a future time. Thus, given estimates of target–missile the missile and target have no acceleration (the up-front
relative position and relative velocity at the current time assumption during the derivation), then Eq. 44 is exact.
t, the future target–missile relative position at time t ) is
given as
CLOSING REMARKS
r (t )) = r (t) + v (t) D t . (42) In this article, we have focused on developing
homing guidance laws by using optimal control tech-
At the CPA, the following condition holds: niques. To this end, a number of modern guidance laws
(PN and beyond) were derived using LQ optimal con-
r (t )) : v (t )) = 0 . (43) trol methods. We note that, regardless of the specific
structure of the guidance law (e.g., PN versus OGL), we
This condition is illustrated in Fig. 13 by the perpendic- developed the relevant guidance law assuming that all
ular line from the target to the relative velocity. Based of the states necessary to mechanize the implementa-
on our assumptions, using Eq. 42 in Eq. 43, and recog- tion were (directly) available for feedback and uncor-
nizing the constant velocity assumption, we obtain the rupted by noise (recall we referred to this as the “perfect
following expression for D t / t go : state information problem”). In a companion article in

58 JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

 MODERN HOMING MISSILE GUIDANCE THEORY AND TECHNIQUES

  5Jackson,
this issue, “Guidance Filter Fundamentals,” we point P. B., TOMAHAWK 1090 Autopilot Improvements: Pitch-
Yaw-Roll Autopilot Design, Technical Memorandum F1E(90)U-1-305,
to the separation theorem, which states that the opti- JHU/APL, Laurel, MD (1 Aug 1990).
mal solution to this problem separates into the optimal   6 Basar, T., and Bernhard, P., H-Infinity Optimal Control and Related
deterministic controller (i.e., the “perfect state informa- Minimax Design Problems, Birkhäuser, Boston (1995).
tion solution”) driven by the output of an optimal state   7Ben-Asher, J. Z., and Yaesh, I., Advances in Missile Guidance Theory,

estimator. Thus, in that article, we discuss guidance American Institute of Aeronautics and Astronautics, Reston, VA
(1998).
filtering, which is the process of taking raw (targeting,   8Grewal, M. S., and Andrews, A. P., Kalman Filtering Theory and Prac-
inertial, and possibly other) sensor data as inputs and tice, Prentice Hall, Englewood Cliffs, NJ (1993).
estimating the relevant signals (estimates of relative   9Pearson, J. D., “Approximation Methods in Optimal Control,” J. Elec-

position, relative velocity, target acceleration, etc.) upon tron. Control 13, 453–469 (1962).
10Vaughan, D. R., “A Negative Exponential Solution for the Matrix
which the guidance law operates.
Riccati Equation,” IEEE Trans. Autom. Control 14(2), 72–75 (Feb
1969).
11Vaughan, D. R., “A Nonrecursive Algebraic Solution for the Discrete
REFERENCES
Riccati Equation,” IEEE Trans. Autom. Control 15(10), 597–599 (Oct
  1Bryson, A. E., and Ho, Y.-C., Applied Optimal Control, Hemisphere 1970).
Publishing Corp., Washington, DC (1975). 12Pue, A. J., Proportional Navigation and an Optimal-Aim Guidance Tech-
  2Cottrell, R. G., “Optimal Intercept Guidance for Short-Range Tacti- nique, Technical Memorandum F1C (2)-80-U-024, JHU/APL, Laurel,
cal Missiles,” AIAA J. 9(7), 1414–1415 (1971). MD (7 May 1980).
  3Ho, Y. C., Bryson, A. E., and Baron, S., “Differential Games and 13Zames, G., “Feedback and Optimal Sensitivity: Model Reference
Optimal Pursuit-Evasion Strategies,” IEEE Trans. Automatic Control Transformations, Multiplicative Seminorms, and Approximate
AC-10(4), 385–389 (1965). Inverses,” IEEE Trans. Autom. Control 26, 301–320 (1981).
  4Athans, M., and Falb, P. L., Optimal Control: An Introduction to the 14Shneydor, N. A., Missile Guidance and Pursuit: Kinematics, Dynamics
Theory and Its Applications, McGraw-Hill, New York (1966). and Control, Horwood Publishing, Chichester, England (1998).

The Authors
Neil  F.  Palumbo is a member of APL’s Principal Professional Staff and is the Group Supervisor of the Guidance,
Navigation, and Control Group within the Air and Missile Defense Department (AMDD). He joined APL in 1993
after having received a Ph.D. in electrical engineering from Temple University that same year. His interests include
control and estimation theory, fault-tolerant restructurable control systems, and neuro-fuzzy inference systems.
Dr. Palumbo also is a lecturer for the JHU Whiting School’s Engineering for Professionals program. He is a member
of the Institute of Electrical and Electronics Engineers and the American Institute of Aeronautics and Astronautics.
Ross A. Blauwkamp received a B.S.E. degree from Calvin College in 1991 and an M.S.E. degree from the University
of Illinois in 1996; both degrees are in electrical engineering. He is pursuing a Ph.D. from the University of Illinois.
Mr. Blauwkamp joined APL in May 2000 and currently is the supervisor of the Advanced Concepts and Simulation
Techniques Section in the Guidance, Navigation, and Control Group of AMDD. His interests include dynamic games,
nonlinear control, and numerical methods for control. He is a member of the Institute of Electrical and Electronics
Engineers and the American Institute of Aeronautics and Astronautics. Justin M. Lloyd is a member of the APL Senior
Professional Staff in the Guidance, Navigation, and Control Group of AMDD. He holds a B.S. in mechanical engineer-
ing from North Carolina State University and an M.S. in mechanical engineering from Virginia Polytechnic Insti-
tute and State University. Currently, Mr.  Lloyd
is pursuing his Ph.D. in electrical engineering
at The Johns Hopkins University. He joined
APL in 2004 and conducts work in optimization;
advanced missile guidance, navigation, and con-
trol; and integrated controller design. For further
information on the work reported here, contact
Neil Palumbo. His email address is neil.palumbo@
Neil F. Palumbo Ross A. Blauwkamp Justin M. Lloyd jhuapl.edu.

The Johns Hopkins APL Technical Digest can be accessed electronically at www.jhuapl.edu/techdigest.

JOHNS HOPKINS APL TECHNICAL DIGEST, VOLUME 29, NUMBER 1 (2010)

59­­