Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
5
Minneapolis, Minnesota, USA, June 14-16, 2006
Abstract— Development of feasible G&C (Guidance and Con- to make use of current state and parameter knowledge.
trol) methods for precision atmospheric re-entry has remained a Additionally, these techniques provide a systematic method
challenge since pre-Apollo-era space exploration. The inherent for simultaneously satisfying state and control constraints in
difficulty arises from the governing hypersonic dynamics being
significantly nonlinear, subject to parametric uncertainty, and the real-time guidance generation.
limited with control authority. Vehicle safety requirements im- The scope of this paper is application-oriented, presenting
pose further constraints, and desired cost objectives complicate a G&C algorithm for reentry based on outer-loop, nonlinear
an already difficult G&C problem. The scope of this paper is to trajectory generation for optimal guidance trajectories and
present a guidance algorithm for optimal trajectory generation inner-loop feedback to handle disturbances, and parametric
based on a reduced-dimension reentry formulation. Preliminary
simulations demonstrate the algorithm with feedback used to and dynamic uncertainty. The guidance trajectory is deter-
track the guidance trajectory in the presence of initial state mined by using a reduced-dimension reentry model to mini-
uncertainty. The objective is to further this approach toward mize the number of equality constraints in the optimization.
an onboard receding-horizon implementation. Preliminary simulations utilize a guidance trajectory and
I. I NTRODUCTION LQR feedback as a proof-of-concept for the algorithm. This
research focuses on reduced-dimension models as a potential
Precision atmospheric re-entry has been a significant enabler for onboard RHC during hypersonic re-entry.
and ongoing research focus for spacecraft G&C (Guidance The remainder of the paper is organized as follows: Sec-
and Control) since the early years of the space program tion II presents the governing reentry dynamics and the basic
[1]−[6]. Significant challenges with re-entry G&C include G&C approach, Section III details the primary contribution
dynamics with significant nonlinearities, state and trajectory of the paper, the reduced-dimension formulation that is
constraints, limits on control authority, and parametric uncer- used for the outer-loop guidance determination, Section IV
tainty. Apollo and Shuttle-era drag-based tracking techniques discusses the inner-loop feedback design, Section V provides
([3]) remain the state-of-the-art for re-entry G&C and are preliminary simulations with the inner-loop/outer-loop algo-
proposed for forthcoming Mars entry vehicles [7], [4], [5]. rithm developed in the paper, and Section VI summarizes
Drag-based tracking, in essence, fits off-line a quadratic or developments in the paper, along with future directions of
low-order polynomial in drag-velocity space to lie inside a the research.
safe re-entry corridor. Onboard control then attempts to track
this curve and provide some degree of guidance updating. II. H YPERSONIC DYNAMICS AND AN A PPROACH TO
The approach is very intuitive; if measured drag is below G UIDANCE AND C ONTROL
desired, drop down in the atmosphere (higher density) to A. Governing Dynamics
increase drag, otherwise, rise up in the atmosphere (lower The kinematics and dynamics governing the descent of
density) to decrease drag. While very effective for re-entry a point-mass vehicle through the atmosphere of a non-
when atmospheric and aerodynamic parameters are well rotating planet are given by Vinh’s equations [2]; the non-
known, the drag-based technique requires careful off-line rotating assumption is a valid approximation for short-
construction of guidance profiles to ensure state and control duration flights [2]:
constraints remain satisfied in the presence of uncertainty.
V cos γ sin χ
With advances in computation power and real-time con- θ̇ = (1)
r cos φ
trol, constrained nonlinear trajectory generation problems
V
are being pushed on-line and computed in real-time [8]− φ̇ = cos γ cos χ (2)
[10]. Rather than using a guidance trajectory computed off- r
ṙ = V sin γ (3)
line, computing power and numerical algorithms now allow
admissible guidance trajectories and control to be computed V̇ = −D − g sin γ (4)
online in a RHC (receding-horizon control) fashion [10] 1 V2
γ̇ = L cos σ − g − cos γ (5)
V r
This work was supported by and conducted at the California Institute of
L sin σ V
Technology, Pasadena, California, USA. χ̇ = + cos γ sin χ tan φ, (6)
J. Carson and M. Epstein are graduate students in Mechanical V cos γ r
Engineering, D. MacMynowski is a Senior Research Fellow and
R. Murray is a Professor in Control & Dynamical Systems. where vehicle parameters (relative to the planet) are as
{carson,epstein,macmardg,murray}@cds.caltech.edu follows: r is radius; V is velocity; γ is flight path angle;
inputs are σ, the bank angle (counter-clockwise about V from Feedback policies are designed to track the guidance
local vertical to L), and α, the angle of attack, which enters trajectories z(t) and handle parametric uncertainty in the
through L and D. nominal parameters pol . The origin is not an equilibrium
Models for specific lift and drag for the vehicle are point of system (12), so we make it one by translating the
nonlinear dynamics about the nominal solution. Let η =
L = κCL V 2 and D = κCD V 2 , (7)
x − z, w = u − v, and δ = p − pol , then
where κ = 12 ρ m S
(S is surface area, m is mass, and ρ η̇ = ẋ − ż = f (x, u; p) − ż
is atmospheric density), CD and CL (vehicle drag and lift
coefficients, respectively) are functions of α, taken as in [6]: = f (η + z, w + v; δ + pol ) − ż = g(η, w; δ) (13)
5783
system dynamics. Although the hypersonic re-entry dynam- The states and controls are now defined in terms of “flat”
ics in equations (1)-(6) are not differentially flat, as shown variables, and thus, the cost function and constraints can be
by [6], concepts from differential flatness significantly reduce recast as functions of the yi ’s and their time derivatives.
the dimensionality of the problem. For the reformulation, we The derivation above utilizes all but equation (4) of
choose three of the states as “flat” variables to parameterize Vinh’s equations, and thus all but equation (4) are implicitly
as splines, and then, define the other states and controls in enforced in the optimization. For y1 (t), y2 (t) and y3 (t) to
terms of the “flat” variables and their derivatives. represent valid reentry trajectories, equation (4) must be
Similar to the formulation in [6], the “flat” variables explicitly enforced:
utilized are
F1 = V̇ol + g sin γol + Dol = 0 (21)
y1 = rol
where the variables V̇ol , g, γol and Dol are replaced by
y2 = φol (14) their dependencies on the yi ’s. Thus, the constraint F1 is
y3 = θol . a function of the “flat” variables as well.
This re-parameterization converts Problem 1 into the fol-
By differentiating the “flat” variables and manipulating lowing, reduced-dimension optimization problem, now with
Vinh’s equations (1)–(3) the remaining states can be deter- only one equality constraint from the dynamics:
mined according to Problem 2: Guidance Trajectory - “Flat” Representation
Vol = ẏ12 + y1 2 (ẏ22 + ẏ32 cos2 y2 ) min J(z(y), v(y), t; pol , Tk )
v(y)
⎧ tk
+Tk
ẏ1 ⎪
γol = arctan
(15) ⎪
⎪ J = q(z(y), v(y), τ )dτ + ν(z(y), tk + Tk )
y1 ẏ22 + ẏ32 cos2 y2 ⎪
⎪
⎪
⎨ tk
ẏ3 subject to F1 (y) = 0
χol = arctan cos y2 . ⎪ − π2 < γol (y), χol (y), αol (y), σol (y) < π2
ẏ2 ⎪
⎪
⎪
⎪ z(y) ∈ Xol
⎪
⎩
To determine the controls we first differentiate the states v(y) ∈ Uol
above to obtain
∀t ∈ [tk , tk + Tk ]
V̇ol = ÿ1 sin γol + cos γol cos χol (ÿ2 y1 + ẏ2 ẏ1 )
where y = {yi (t), ẏi (t), ÿi (t), ∀i = 1, 2, 3}, tk ≥ 0 is the
+ H cos γol sin χol initial time, and Tk > 0 is the horizon.
1 1
γ̇ol = ÿ1 cos γol − sin γol cos χol (ÿ2 y1 + ẏ2 ẏ1 ) IV. D EVELOPMENT OF I NNER -L OOP F EEDBACK
Vol Vol
1 C ONTROL
− H sin γol sin χol (16) Feedback laws are developed for cases with small per-
Vol
−1 turbations from the guidance trajectory. This assumes the
ẏ 2
χ̇ol = 1 + 32 cos2 y2 linear terms in an expansion of system (13) dominate in some
ẏ
2 region about the nominal trajectory, so system (13) becomes
ÿ3 ẏ3 ÿ2
× cos y2 − ẏ3 sin y2 − 2 cos y2 η̇ = Aη + Bw (22)
ẏ2 ẏ2
with H = (ÿ3 y1 cos y2 + ẏ3 ẏ1 cos y2 − ẏ2 ẏ3 y1 sin y2 ). The where
bank angle can then be found from equations (5) and (6):
A = ∇x f |(z,v;pol ) and B = ∇u f |(z,v;pol ) .
⎛ ⎞
χ̇ol cos γol − Vyol1 cos2 γol sin χol tan y2 The matrices A and B are functions of the nominal trajectory,
σol = arctan⎝ ⎠.
so system (22) is LTV (Linear Time Varying). The parametric
γ˙ol − cos γol Vyol1 − Vgol
perturbations δ in system (13) are neglected here. The
(17) linearization of Vinh’s equations (3)-(6) about a nominal
The specific lift is then given by trajectory gives the following B and A matrices (the A
Vol Vol g matrix is on the top of the next page):
Lol = γ̇ol − cos γol − . (18) ⎡ ⎤
cos σol y1 Vol 0 0
⎢ 0 0 ⎥
The angle of attack is given by ⎢ ⎥
⎢ 0 0 ⎥
1 2mLol B(z, v; pol ) = κV ⎢
⎢ −CDα V
⎥
⎥ (23)
αol = − a0 , (19) ⎢ 0 ⎥
a1 ρVol 2 S ⎣ CLα cos σ −CL sin σ ⎦
and finally the specific drag is determined by making the CLα sin σ CL cos σ ol
appropriate substitutions in equation (7)
where all state variables and parameters are nominal values,
Dol = κCD Vol 2 . (20) as implied by the subscript ol. Note, this formulation makes
5784
A(z, v; pol ) =
2 3˛
0 V sin χ tan φ
− rV2 sin χ sin χ
− Vrγcos
sin χ V cos χ ˛
r cos φ cos φ r cos φ φ r cos φ ˛
6 0 0 V cos χ
− r2 cos χ V γ cos χ
− r − V sin χ 7˛
6 r r 7˛
6 0 0 0 γ V 0 7˛
6 7˛
6 κCD V 2 7˛
6 0 0 −2κCD V −g 0 7˛
6 “ href ” ` ´ `g ´ 7˛
6 7˛
6 0 0 −V κChLref cos σ
+ r12 κCL cos σ + Vg2 + r1 γ V − Vr 0 7˛
4 “ ” ` ´ 5˛
˛
0 V sin χ
r
(1+tan2 φ) −V κChLref
sin σ
+ sin χr2tan φ κCL sin σ+ sin χrtan φ Vγ κCL sin σ− tan φrsin χ V tan φ cos χ
r ˛
ol
(24)
two further assumptions: the perturbation in gravity between sufficient constraints exist that are much larger algebraic
actual and nominal systems is negligible (which affects the expressions.
A matrix formulation), and the flight-path angle is small. The
variables CLα and CDα are the derivatives of CL and CD V. C ASE E XAMPLE
with respect to α, respectively. This section provides a proof-of-concept simulation for
Assuming system (22) is stabilizable, then LQR (Lin- the G&C concept developed in this paper and is a key
ear Quadratic Regulation) feedback provides a convenient step toward the ultimate interest of implementing RHC in
method for demonstrating the re-entry algorithm and key the outer-loop guidance. A guidance trajectory is computed
components of a potential RHC implementation. The feed- through the method of Section III, and inner-loop feedback
back is given by w = −Kη where gain matrix K = is computed as outlined in Section IV to deal with small
1 −1
2R B(t)T P (t) comes from the LQR problem [13] uncertainties in the initial conditions. These uncertainties can
∞
∗
T represent either the effect of computation delay in the outer-
w = min η Qη + wT Rw dt (25) loop trajectory generation, or they can represent errors in the
w 0
estimated entry conditions.
Q = QT >= 0 and R = RT > 0. (26)
Past studies in [14] and [6] used the maximization of
and solving a Riccati equation for P . Since system (22) is cross-range as the cost function in designing optimal entry
LTV, an approximation is made that A, B and P are constant trajectories
over small time increments along the guidance trajectory, J = −y2 (tf ). (30)
which allows P to be determined by solving an Algebraic
Riccati Equation As a baseline we use the same cost function and parameters
(See Table I) because they provide a means for comparison
0 = A(ti )T P + P A(ti ) − P B(ti )R−1 B(ti )T P + Q of the guidance trajectory and a way of validating the overall
at set time indices ti along the nominal trajectory. The G&C approach developed in this paper.
resultant feedback gain Ki is then fixed until the next time
TABLE I
index. The stability of this switching scheme is not addressed
S IMULATION PARAMETERS
in this paper. Robustness to parametric uncertainty is ignored
for now too. Parameter Value
Sufficient constraints can be imposed on the guidance r(0) 260000 + Re (feet)
φ(0) 0◦
Problem 2 (or used as verification checks for a computed θ(0) 0◦
trajectory) to ensure the linearized system is controllable v(0) 25600 (ft/sec)
along the nominal trajectories, which will ensure a solution γ(0) −1◦
to the LQR problem. Some of these constraints are intuitive, ψ(0) 90◦
tf 2000 (sec)
like non-zero velocity (V > 0), already a constraint in h(tf ) 80000 (ft)
Problem 2, and non-zero lift, which puts an additional v(tf ) 2500 (ft/sec)
constraint on α; the α constraint, plus some of the other γ(tf ) −5◦
μ 0.14976539 × 1017
sufficient controllability constraints include Re 20902900 (feet)
S 2690 (ft2 )
CL > 0 ⇒ α > − aa01 (27)
m 6309.44 (slugs)
CDα > 0 ⇒ α> − 2b
b1
2
(28) href 23800 (feet)
ρ0 0.002378
CLα tan γ + CDα cos σ = 0 (29) a0 −0.20704
a1 0.029244 (deg.−1 )
Equation (29) indicates a coupling between angle-of-attack b0 0.07854
and flight-path angle where ratios of lift and drag properties b1 −0.61592 × 10−2 (deg.−1 )
can cause a loss of control at certain bank angles. Other b2 0.621408 × 10−3 (deg.−2 )
5785
Note, these simulations (as in [14], [6]) do not incorporate
namic pressure, max load capacity, etc.) or control rate limits,
but the approach is capable of handling such constraints. In
addition, the controllability constraints (27)-(29) were not
◦
θ(0)actual = 0.1
φ(0)actual = 0.1◦
γ(0)actual = −0.95◦ ,
with the LQR design matrices chosen as Fig. 3. Dynamic States of Nominal Trajectory
0.25 0 30
Constraints 1 & 2
C >0 CD > 0
Q = diag(1, 1, 4, 4, 5, 1) and R = () L
0 0.5 20
10
with higher penalties on radius, velocity, and flight path angle
0
in order to keep the state trajectories close to the guidance 0 200 400 600 800 1000 1200 1400 1600 1800 2000
latitude φ) is 34.0011◦ , which is comparable to the results of 0 200 400 600 800 1000 1200 1400 1600 1800 2000
6
2 with 10 intervals of 7th -order B-splines having 4th -order
smoothness between intervals. 5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (sec)
100
Fig. 4. Constraints that Impose Controllability on the Feedback System
()
50
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 Figure 5 depicts the outer-loop along with the inner-loop
40 feedback control generated with LQR. Notice in the plot that
the feedback control magnitude is small.
()
20
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
7
x 10
2.12
2.11
r (ft)
2.1
2.09
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Fig. 2. Kinematic States of Nominal Trajectory
5786
loop optimal trajectory) in the simulations when the inner- upon the guidance trajectory. The continuation of this work
loop feedback is applied as well a simulation where the will explore different nonlinear optimizers for use in gener-
inner-loop feedback is turned off and the only control signal ating the guidance trajectories, with an intent to implement
applied is that from the outer-loop. The benefit of feedback RHC in the outer loop. This will enable resolving for the
is apparent, the small perturbation in the initial conditions guidance trajectory, thus allowing an assessment of how well
has a large affect on the system when inner-loop feedback is the algorithm handles sources of uncertainty and external
not included, this is particularly profound in the final radius, disturbances.
off by 10000 ft at the end time without feedback. With even The cost function used in the case example was chosen
the small amount of feedback as shown in Figure 5, the state for comparison with previous studies. Studies with other
remains close to the desired, optimal guidance trajectory. cost functions may be more relevant to actual missions
and should be investigated. Additionally, the complicated
algebraic constraints associated with the controllability of the
feedback system need to be further refined or a new feedback
methodology chosen. Overall stability of the inner/outer-loop
architecture should be explored as well as the ensuring a
method to handle parametric uncertainty.
ACKNOWLEDGEMENTS
The authors would like to thank Mark Milam and Bill
Dunbar, alumni of California Institute of Technology, and
Raktim Bhattacharya, former post-doc under Richard Mur-
ray, for their assistance with nonlinear trajectory generation.
R EFERENCES
[1] R.C. Wingrove. survey of atmosphere re-entry guidance and control
methods. AIAA Journal, 1(9):2019–2029, September 1963.
[2] N.X. Vinh et. al. Hypersonic and Planetary Entry Flight Mechanics.
Fig. 6. Error from Nominal: Feedback ON (solid), Feedback OFF (dashed) University of Michigan Press, 1980.
[3] J.C. Harpold and C.A. Graves Jr. Shuttle entry guidance. The Journal
of Astronautical Sciences, Vol. XXVII, No. 3, July-September, pages
239–268, 1979.
[4] K-Y Tu, M.S. Munir, K.D. Mease, and D.S. Bayard. Drag-based
predictive tracking guidance for mars precision landing. Journal of
Guidance, Control, and Dynamics, Vol. 23, No. 4, July-August, pages
620–628, 2000.
[5] S.A. Striepe, D.W. Way, A.M. Dwyer, and J. Balaram. Mars smart lan-
der simulations for entry, descent, and landing. In AIAA Atmospheric
Flight Mechanics Conference and Exhibit, August 5–8 2002.
[6] T. Neckel, C. Talbot, and N. Petit. Collocation and inversion for a
reentry optimal control problem. In Proceedings of the 5th Interna-
tional Conference on Launcher Technology, 2003.
[7] G.F. Mendeck and G.L. Carman. Guidance design for mars smart
landers using the entry terminal point controller. In AIAA Atmospheric
Flight Mechanics Conference and Exhibit, August 5–8 2002.
[8] R. Bhattacharya, G. J. Balas, M. Alpay Kaya, and A. Packard.
5787