JOURNAL OF GUIDANCE, CONTROL, AND DYNAMICS

Vol. 33, No. 5, September–October 2010

On the Benefits of In-Flight System Identification

for Autonomous Airdrop Systems

Michael Ward∗ and Mark Costello†

Georgia Institute of Technology, Atlanta, Georgia 30332
and
Nathan Slegers‡
University of Alabama in Huntsville, Huntsville, Alabama 35899
DOI: 10.2514/1.49030
A unique feature of airdrop systems is the inherent and large variability in flight dynamic characteristics. The
same physical article dropped on two different occasions will exhibit significantly different dynamic response. The
problem only becomes worse for different test articles. Control systems for autonomous airdrop systems explicitly or
implicitly assume knowledge of the flight dynamic characteristics in some way, shape, or form. A question facing
autonomous airdrop designers is whether to use precomputed dynamic characteristics inside the control law, or to
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

compute the needed flight dynamic characteristics in-flight and subsequently employ them in the control law. This
paper establishes conditions when in-flight identified characteristics, with a focus on the turn rate dynamics, should
be used, and when it is better to use precomputed results. It is shown that with expected levels of system variability,
sensor noise, and atmospheric wind, in-flight identification generally produces significantly more accurate dynamic
behavior of the lateral dynamics than a precomputed model of the nominal system, even when the in-flight
identification is performed with highly inaccurate sensor data. The only exception to this rule observed in this work is
the situation where atmospheric winds are high and a direct heading measurement is not available. In this situation, a
precomputed estimate of the time constant of the lateral dynamics is more accurate than an in-flight estimate. These
conclusions are reached though a comprehensive simulation study using a validated airdrop flight dynamic model
applied to both a small and large parafoil.

Nomenclature
SBcg;P  = skew symmetric cross product matrix for

A, B, C, = apparent mass and inertia coefficients distance vector from system center of gravity to
P, Q, R payload
b = canopy span SC , SP , Si = canopy, payload, and canopy element reference
CD;P = payload drag coefficient areas
CL;i , CD;i = lift and drag coefficients of the ith canopy u, v, w = velocity components of system mass center
element expressed in body frame
d = canopy arc radius u~ i , v~ i , w~ i = aerodynamic velocity components of ith canopy

I = identity matrix element in ith canopy frame

IT  = total system inertia matrix V = projection of velocity vector on horizontal plane
LAM , MAM , = components of apparent mass moment in the V0 = projection of velocity vector on horizontal plane
NAM body reference frame in wind-fixed frame
m = total system mass Vw = wind vector
p, q, r = angular velocity components expressed in the XA;i , YA;i , = components of aerodynamic force on ith canopy
body reference frame ZA;i element in body frame

SBx  = skew symmetric cross product matrix for vector XA;P , YA;P , = components of aerodynamic force on payload in
x in body frame ZA;P body frame

SB!  = skew symmetric cross product matrix for system XAM , YAM , = components of apparent mass force in body
angular velocity ZAM frame

SBcg;i  = skew symmetric cross product matrix for XW , YW , ZW = components of weight vector in body frame
distance vector from system center of gravity to x, y = north and east position
ith canopy element
= sideslip angle

SBcg;M  = skew symmetric cross product matrix for
= canopy arc angular span
distance vector from system center of gravity to L , R = normalized left and right brake deflections
apparent mass reference point  = azimuthal rate time constant
e = correlated measurement noise time constant
, , = Euler roll, pitch, and yaw angles
Received 22 January 2010; revision received 31 March 2010; accepted for
publication 31 March 2010. Copyright © 2010 by Michael Ward. Published
 = azimuth angle (course over ground)
by the American Institute of Aeronautics and Astronautics, Inc., with 0 = wind-relative azimuth angle (course in wind-
permission. Copies of this paper may be made for personal or internal use, on fixed frame)
condition that the copier pay the $10.00 per-copy fee to the Copyright _ 0;ss = steady-state wind-relative azimuthal rate
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923; include w = wind vector heading angle
the code 0731-5090/10 and $10.00 in correspondence with the CCC.
∗
Graduate Research Assistant, School of Aerospace Engineering. Member
AIAA. I. Introduction
†
Sikorsky Associate Professor, School of Aerospace Engineering.
Associate Fellow AIAA.
‡
Assistant Professor, Department of Mechanical and Aerospace
Engineering. Member AIAA.
A UTONOMOUS guided parafoils are an attractive option for
payload delivery due to their low weight and packing volume
and their ability to penetrate winds and maneuver to achieve high
1313
 1314 WARD, COSTELLO, AND SLEGERS

placement accuracy. These systems require an onboard guidance, analysis and prior flight testing do not have sensor requirements, but
navigation, and control system to generate and track a desired path to are not capable of adapting to the peculiar dynamics of each dropped
the target. Either directly or indirectly, current autonomous airdrop item. The purpose of the current article is to explore the tradeoff
system control laws contain an embedded model of the flight between precomputed and preloaded dynamic characteristics versus
dynamics that is loaded into the automatic guidance unit (AGU) in-flight estimated characteristics. The approach presented here is
before each flight [1–8]. In the case of model predictive control, the meant to be generally applicable in that it does not depend on specific
controller contains the actual dynamic model, but in the more guidance, navigation, and control strategies.
traditional control schemes the model is embedded in the form of the Toward this end, a straightforward and robust prototype procedure
preselected control gains and logic gates. For control system design to perform the entire system identification task in-flight is defined.
purposes, a flight dynamic model is typically created from analytical The procedure is based on using an extended Kalman filter observer
modeling or a best fit to flight data [9–12]. A unique feature of to estimate the wind-relative azimuth angle and rate in-flight. The
parafoil and payload aircraft is the high degree of variability of the process is demonstrated with simulated measurements for a GPS unit
basic flight dynamics from drop to drop, particularly the control operating alone and a GPS unit and heading measurement combina-
response. Parafoil canopies made to the same specifications will tion. Windows of estimated azimuthal rate data from a few key flight
exhibit some variation in canopy and rigging geometry, and the segments are then used to generate a model of the lateral dynamics.
weight and geometry of the payload will change from drop to drop. To compare the differences between use of precomputed models and
Furthermore, airdrop systems are susceptible to preflight program- in-flight estimated models, a 6 degrees of freedom, nonlinear parafoil
ming errors (such as specifying an incorrect payload weight or and payload flight dynamic model is employed. Appropriate levels of
canopy), rigging errors (such as control line or riser adjustment), and uncertainty are injected into the dynamic model to represent drop to
deployment malfunctions (such as partially inflated or torn canopies drop variability in the basic vehicle dynamics and appropriate levels
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

and twisted rigging lines). Even with no apparent malfunctions, the of atmospheric wind are also injected into the simulation. Sensor
same parafoil and payload system will often tend to exhibit a noise and bias are added to the simulation data that is fed into the in-
significantly different control bias with every drop. This means that flight estimation algorithm to simulate low-cost, commercially avail-
the nominal or average model will differ substantially from the actual able sensors. Results are generated for two exemplar systems,
flight dynamics, and this can induce substantial degradations in path namely, a microparafoil flight tested by Slegers et al. [14] and the
tracking ability and subsequently landing accuracy. Airborne Systems 30k Megafly [15]. These systems have a total
A solution to the above problem is to identify the needed dynamic weight of approximately 5 and 30,000 lb, respectively. These
and control characteristics in-flight using a system identification systems were chosen to demonstrate that the in-flight system identifi-
algorithm. A significant literature database on parafoil and payload cation procedure is applicable to airdrop systems of any scale with
aircraft system identification has been reported including estimation minimal modifications. Example results are given to demonstrate the
of flight dynamic models and atmospheric wind velocity. While most ability of the observer to estimate the wind and wind-relative
of this work has focused on batch processing, some work is directed dynamics. A Monte Carlo simulation is performed to evaluate the
toward in-flight estimation [1,2,5,6]. The key quantities to estimate benefits of in-flight system identification with expected levels of
for autonomous parafoil and payload aircraft are the atmospheric sensor noise and variation in the parafoil and payload system. Results
wind velocity vector and the lateral control response. Atmospheric are presented in terms of the quality of the estimated characteristics
wind estimation is obtained by subtracting an estimate of the wind- rather than in terms of overall performance measures such as landing
relative velocity vector from the observed ground track velocity accuracy to preserve the separation of the methods described here
[1–5]. Airspeed of the airdrop system is assumed known from a from any particular guidance, navigation, and control strategies. The
preflight analysis. Preflight models include a constant assumed sensitivity of these results to changes in the assumed levels of sensor
airspeed [5], airspeed determined from a constant reference dynamic noise, system variation, and wind levels is studied to enable state-
pressure [13], and airspeed determined from a constant reference ments to be made regarding when it is most appropriate to use
glide slope and measured sink rate [6]. Jann developed a nonlinear precomputed dynamic characteristics and when it is most appropriate
filter for simultaneous estimation of the wind and airspeed using to use in-flight estimation of dynamic characteristics.
Global Positioning System (GPS) data [5], though the effectiveness
of the filter in estimating airspeed is not mentioned specifically. II. Parafoil Dynamic Simulation Model
Calise and Preston pointed out that it is simple to obtain an airspeed
This section describes the nonlinear simulation model of the
estimate from GPS data by flying a complete circle, though they do
parafoil and payload system used to generate the simulated measure-
not perform this maneuver in every flight [2]. Beyond wind and
ments that serve as inputs to the system identification process. The
airspeed determination, transient and steady-state lateral control
simulation model described here is not part of the linear dynamic
characteristics are estimated. Calise and Preston developed an
model identified in-flight.
adaptive stability augmentation system for guided parafoils and
The combined system of the parafoil canopy and the payload are
demonstrated the effectiveness of the approach in simulation [1,2].
modeled with 6 degrees of freedom, including three inertial position
Their approach identifies a set of reference dynamics for a particular
components of the total system mass center as well as the three Euler
system and then applies an adaptive correction to account for flight-
orientation angles. With the exception of movable parafoil brakes, the
to-flight variability in the system dynamics. This approach requires
parafoil canopy is considered to be a fixed shape. Canopy aerodyn-
high quality measurements and has led them to pursue a control
amics are modeled by splitting the canopy into five discrete elements
scheme based on bank angle feedback using a GPS/inertial
as shown in Fig. 1 and determining the lift and drag on each element
measurement unit combination rather than traditional heading angle/
based on the local angle of attack. This aerodynamic model is a simple
heading rate feedback.
way of simulating a parafoil canopy. The lift and drag coefficients are
If the system identification task is performed in-flight, the
not varied across the span. The lift and drag coefficients are set to
controller can be designed online so that it is always tuned to the
match steady-state longitudinal data from flight tests, and the canopy
actual flight dynamics. While this is advantageous, it is not without
arc is matched to the actual system geometries so that the orientation
problems. Any system identification scheme requires measurement
of the panels produces the aerodynamic moments and side force.
of appropriate control inputs and response outputs to estimate air
The kinematic equations for the parafoil and payload system are
vehicle properties. While autonomous airdrop systems already
provided in Eqs. (1) and (2). A shorthand notation for trigonometric
incorporate a sensor suite for feedback control, the sensor suite
functions is employed where sinx  sx , cosx  cx , and
required for system identification requires more accurate measure-
tanx  tx .
ment of more vehicle states than the sensor suite required for flight ( _) ( )
control. Also, errors associated with sensors can be sufficiently large x u
to render estimation of dynamic characteristics inaccurate. On the y_ 
TIB T v (1)
other hand, dynamic characteristics generated before flight from z_ w
 WARD, COSTELLO, AND SLEGERS 1315

Fig. 1 Canopy schematic.

( _ ) 21 s t 3
c t ( p ) wind velocity components in the inertial frame as fVW;x ; VW;y ;
  
_  4 0 c s 5 q (2) VW;z gT , the aerodynamic velocity of the ith element is given by
_ 0 s =c c =c r Eq. (8). The aerodynamic velocity of the payload is given by the same
equation with the body frame to element transformation equal to the
The matrix
TIB  represents the transformation matrix from the identity matrix
inertial reference frame to the body reference frame 0( ) 8 91
( ~ ) ( )
2 3 ui u p < VW;x =
c c c s s v~ i 
TB;i @ v 
SBcg;i  q 
TIB  VW;y A (8)

TIB   4 s s c  c s s s s  c c s c 5 (3) w~ i w r
:
VW;z
;
c s c  s s c s s  s c c c
The aerodynamic forces on the canopy elements are expressed in
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

The dynamic equations are formed by summing forces and

moments about the system center of gravity, both in the body terms of lift and drag coefficients, which are functions of the angle of
reference frame, and equating to the time derivative of linear and attack of each element i  tan1 w~ i =u~ i  and the brake deflectionsi
angular momentum, respectively if a brake spans the element, as shown in Eqs. (9) and (10).
08 9 Equation (11) defines the canopy aerodynamic forces in the body
8 91 8 9 8 9
>
> u_ >
> > u> >
> XW >
> >
> XAM >> reference frame
B < = < = C < = < =
mB@> > v_ 
S B
!  v C  YW  YAM
A >
> >
: > > > > CL;i  CL0;i  CL;i i  CL3;i 3i  CL;i i (9)
: _> ; ; >
: >
; >: >
;
w w ZW ZAM
8 9 8 9
>
> X > > XA;i >
< A;P >= X> < >
= CD;i  CD0;i  CD2;i 2i  CD;i i (10)
 YA;P  YA;i (4)
>
> > > >
:Z > ; i >:Z > ; 8 9 0
A;P A;i 8 9
>
> XA;i >
> w~
< = 1 B q


















>
< i > =
8 9 8 9 8 9 8 9 YA;i  Si
TB;i T B @ CL;i u~i 2  w~ 2i 0
> p_ > p > LAM > > X > >
> > 2 > >
>
< > = >
< > = > < >
= >
< AM >= :Z > ; : ;
A;i u~ i

IT  q_ 
SB! 
IT  q  MAM 
SBcg;M  YAM 8 91
>
> > >
: > ; > > > >
: > ; >
: >
; >
: >
; q



























>
u~
< i> =C
r_ r NAM ZAM
8 9 8 9  CD;i u~ i  v~i  w~ i v~i C
2 2 2
A (11)
>
> X > > X > >
: > ;
< A;P >
= X >
< A;i >
= w~ i
B B

Scg;P  YA;P 
Scg;i  YA;i (5)
>
> > > >
:Z > ; i >
:Z > ;
A;P A;i
The aerodynamic force on the payload consists entirely of profile
drag and is given by Eq. (12)
S is the skew symmetric operator, used to express the cross product 8 9 8 9
of two vectors as a matrix multiplication of the components of the < XA;P = q





























< u~ P =
1
vectors in a specified frame. For example, if the vectors a, b, and c are Y   SP CD;P u~ 2P  v~ 2P  w~ 2P v~ P (12)
: A;P ; 2 : ;
expressed in terms of their components in frame B ZA;P w~ P
8 9 2 38 9
>
> c > 0 az ay > b >
< x> = >
< x> = Parafoils with small mass to volume ratios can experience large
6 a 7
c  a  b $ cy  4 z 0 ax 5 by forces and moments from accelerating fluid. These are termed
>
> > >
> >
: > ; ay ax 0 : bz ;
> apparent mass effects. A precise accounting of these effects can
cz
8 9 substantially complicate the dynamic equations [14], but it is
>
> b > possible to obtain a good approximation of the effects with only a few
< x> = of the terms. The approximate forms used for the apparent mass
B

Sa  by (6) forces and moments are given in Eqs. (13) and (14). Parametric
>
> >
: > ; approximations given by Lissaman and Brown [16] are used to
bz
determine the apparent mass and inertia coefficients in Eq. (15)
Forces appearing in Eq. (4) have contributions from weight, 8 9 0( )
aerodynamic loads on the canopy and payload, and apparent mass. ( _ )1
< XAM = u_ p
The weight contribution is given in Eq. (7) YAM  
IAM @ v_ 
Scg;M  q_ A
B
(13)
: ;
8 9 8 9 ZAM w_ r_
< XW = < s =
Y  mg s c (7)
: W; : ; 8 9
ZW c c ( _)
< LAM = p
Defining the transformation from the body frame to the frame MAM  
IAI  q_ (14)
: ;
attached to the ith element of the canopy as
TB;i  and defining the NAM r_
 1316 WARD, COSTELLO, AND SLEGERS

2 3 2 3
A 0 0 P 0 0

IAM   4 0 B 0 5;
IAI   4 0 Q 0 5 (15)
0 0 C 0 0 R

The dynamic equations of motion are found by substituting all

forces and moments into Eqs. (4) and (5), resulting in the matrix
solution in Eqs. (16–18)
8 9
>
> u_ >
>
>
> > >
>
> >
>
> v_ >
>
>
>
> >
>
>
> >
> Fig. 2 Decomposing measured velocity vector (top view).
" # >
> w _ >
m
I 
IAM  
IAM 
SBcg;M  < > =


Scg;M 
IAM 
IT  
IAI  
Scg;M 
IAM 
Scg;M  >
B B B
> >
> flight segments can be executed separately at any time during the
>
> p_ >>
>
> >
> flight and in any order. After the flight segments have been
>
> >
> completed, the in-flight system identification process is executed as
>
> >
>
>
> q_ >
> >
> > follows: 1) estimate the forward airspeed and initialize the wind
: r_ >; estimate, 2) generate a stream of wind-relative azimuthal rate data for
2 3 key flight segments, 3) derive a mapping of control input to steady-
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

B1 state turn rate from the azimuthal rate data, and 4) identify the turn
6 7 rate dynamics from azimuthal rate data.
 4 5 (16)
The first two flight segments correspond to the typical initial
B2 segments of a guided parafoil drop (allow the system to reach an
initial equilibrium then point at some waypoint), and they are used to
( ) 8 9 8 9 8 9 estimate the control bias. The circling segment is used to estimate the
u < XW = < XA;P = X< XA;i = forward airspeed, and the circling segment combined with the last
B1  m
SB!  v  YW  YA;P  Y (17) segment constitutes a brake doublet, which is used to estimate the
: ; : ; : A;i ;
w ZW ZA;P i ZA;i turn rate dynamics.

( ) 8 9 8 9 B. Estimating Forward Airspeed

p < XA;P = X < XA;i =
B2  
SB! 
IT  q 
SBcg;P  YA;P 
SBcg;i  YA;i (18) The circling segment of the fight procedure is flown until it is
: ; : ; apparent from the ground track azimuth measured with GPS that at
r ZA;P i ZA;i
least one complete circle has been flown. The ground track azimuth
data is then processed to precisely estimate the beginning and end of a
Equation (16) represents a set of coupled, nonlinear differential single complete circle. The instants marking the beginning and end of
equations. The matrix on the left hand side of Eq. (16) is a function of the complete circle are denoted as i and f, respectively. Equation (19)
the mass and geometry properties of the parafoil. The geometry of the is then used to estimate the forward airspeed, where the mean value is
parafoil is assumed to be fixed, so this matrix is constant and only computed over the entire set of measurements for the complete circle
needs to be inverted once at the beginning of the simulation. With (i  k  j). The drift in the system position over the complete circle
specified initial conditions, the states can be numerically integrated is also used to initialize the wind estimates
forward in time.
s









































































   

xf  xi 2 yf  yi 2
V0  mean x_ k   y_ k  (19)
III. In-Flight System Identification tf  ti tf  ti
A. Overview of Process
With current autonomous parafoil and payload systems, typically
C. State Estimation
only lateral control is used to guide the parafoil to the target [1–8].
This reduces the parafoil trajectory planning problem to generating a The goal of the state estimation process is to produce an estimate of
desired trajectory in the horizontal plane in the form of heading the wind and a stream of wind-relative azimuthal rate data _ 0 t, from
commands, which in turn reduces the control problem to tracking a which the system dynamics can be identified. The observer is an
desired heading with a single lateral control input. In the presence of extended Kalman filter [17]. The system state vector xk and
wind, it is normally desirable to generate the trajectory in a wind- measurement vector zk are considered to be nonlinear functions
relative reference frame so that the commanded heading to be tracked 
is actually the wind-relative azimuth angle 0 . This quantity xk1  fk xk  
Nk xk nk
(20)
corresponds to the Euler yaw angle , when the sideslip is zero, as zk  gk xk   wk
shown in Fig. 2. This means that the autonomous control law is
designed based on an estimate of the wind-relative azimuthal rate where nk is the process noise vector, and wk is the measurement noise
dynamics. Determination of these dynamics relies on the solution of vector, both of which are assumed to be zero-mean Gaussian white
the vector diagram in Fig. 2. The forward airspeed component V0 is noise sequences.
assumed to remain constant during each flight and will be identified The wind vector and the aerodynamic velocity vector are defined
in-flight. If a heading measurement is available, the sideslip is as states in the observer xk, so the state update equations fk xk  are
neglected (
0 ), so that the vector diagram in Fig. 2 is fully defined by the vector geometry that relates these states to the
defined. Without a heading measurement, the wind-relative velocity measured ground track velocity vector (Fig. 2). These equations
vector can be determined over a series of measurements by assuming assume that the wind vector, airspeed, and wind-relative azimuthal
that the wind vector changes slowly, similar to Jann [5], as well as rate are all constants that are perturbed by a process noise vector nk
Carter et al. [6]. consisting of independent perturbations to the north and south wind
The overall flight procedure requires four key flight segments: components, airspeed, and azimuthal rate (Vwx , Vwy , V0 , and _ 0 ,
1) zero control input, 2) track a heading command, 3) complete a respectively). The observer state update equations are given in
circle in one direction, and 4) turn briefly in the other direction. The Eq. (21)
 WARD, COSTELLO, AND SLEGERS 1317

0 1 0 1
xk xk  x_ k t deflection is limited to a region where the steady-state turn rate is
B C B C approximately proportional to control deflection, so the control to
B yk C B yk  y_ k t C turn rate mapping need only account for control bias and asymmetric
B C B C
B C B C
B x_ k C B V0 cos0k   Vwxk C control sensitivity. Four points of steady-state turn rate data
B C B C corresponding to the four key flight segments are computed from the
B C B C
B y_ k C B V0 sin0k   Vwyk C azimuthal rate estimates produced by the observer. These four points
B C B C
B C B C are used to generate the linear mapping from control input to steady-
xk  B Vwxk C; fk xk   B Vwxk C;
Nk xk nk
B C B C state turn rate as shown in Fig. 3.
BV C B Vwyk C
B wyk C B C
B C B C
B  C B 0k  _ 0k t C
B 0k C B C E. Turn Rate Dynamics
B C B C
B _ C B _ 0k C The open-loop wind-relative azimuthal rate dynamics are approxi-
@ 0k A @ A
mated as a first order filter of the steady-state turn rate
zbk zbk
0 1
0 _ 0;k1  _ 0;k  _ 0;ss  _ 0;k = (23)
B C
B 0 C
B C
B C This leaves only the azimuthal rate time constant  to be estimated
B cos0k V0 C
B C from the data stream produced by the observer. The assumption that
B C
B sin0k V0 C the turning dynamics can be approximated as a first order system was
B C made based on the authors’ experience flight testing small parafoils.
B C
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

B Vwx C (21) Nonlinear simulation results for the example systems also appear to
B C
B Vwy C be approximately first order, and there is at least one example in the
B C
B C literature where the turning dynamics from flight-test data of a
B 0 C
B C medium sized (250 lb) system appeared to be reasonably
B C
B
@ _ 0 C
A approximated as first order [19]. However, second-order turning
dynamics have been observed for some systems in-flight test [3,20].
xb The approach described in the present work can be applied to these
systems by modifying Eq. (23) to account for second-order dynamics
so that two dynamic parameters (e.g., natural frequency and damping
There are two measurement cases considered, GPS only and GPS ratio) would need to be identified rather than one. Another possibility
with a heading measurement. The measurement vector from GPS is that the rate limit on the brake deflection may be such that the
(zk ) consists of north and east position and north and east velocity lateral dynamics are never excited, so that the system is always flying
components, and the heading measurement is assumed to represent at the steady-state turn rate [18,21]. In this case, only the mapping of
the wind-relative azimuth (zero sideslip assumption). To make the control input to steady-state turn rate would need to be identified.
process more robust to measurement error, biases in the velocity The time constant is determined from the estimated azimuthal rate
vector and heading angle measurements are also estimated by data by a Gauss–Newton optimizer to minimize the error between the
appending them to the state vector [zb in Eq. (21), where zb  observed azimuthal rate and the azimuthal rate predicted with
fzbx_ ; zby_ gT for GPS only and zb  fzbx_ ; zby_ ; zb gT for GPS with a Eq. (23). The optimization process was found to be robust in the
heading measurement]. The measurement vector is shown in presence of large disturbances during the system identification
Eq. (22); the last entry z is not present for the GPS only case maneuvers. Also, while the use of an optimizer in system identifi-
cation is normally one of the more computationally intensive
0 1 0 10 0 1 approaches, the current problem can be solved very quickly because
zxk xk wx there is only one parameter, the objective function is quickly
B zyk C B yk C B wy C determined by propagating Eq. (23), and the problem is well
B C B C
zk  B z C; gk xk   B
B x_  z C
C w  B wx_ C (22) conditioned so that it normally converges in less than five iterations.
B _
xk C @
k bxk
_
A k B C
@ zyk
_
A y_ k  zbyk_ @ wy_ A
z k 0kzb k w
IV. Simulation Parameters
A. Example Parafoil and Payload System Parameters
D. Steady-State Turn Rate
The mapping of control input to steady-state turn rate is performed To investigate the characteristics of the in-flight estimation
separately from the estimation of the transient characteristics of the procedure, a large and a small parafoil system were considered
lateral dynamics. The idea is that there is a deterministic mapping (Fig. 4). These two systems span the entire range of autonomous
from each left and right brake position to steady-state turn rate, and airdrop systems. The physical and aerodynamic parameters for both
while this mapping is not generally linear [10,18] the dynamics that parafoils can be found in Tables 1 and 2, while Fig. 5 defines some of
describe how the turn rate reaches the given steady-state value may be the geometry parameters used in the tables. Parameters for both
approximated as linear. The current work assumes that the control parafoils were chosen to match flight-test data [4,14]. The values for
these parameters represent a nominal model of each vehicle. A
significant amount of variability in the flight dynamics from flight-to-
flight was observed in the flight-test data. Estimates for these levels of
uncertainty form the basis for the Monte Carlo simulations discussed
below.
The canopy brakes are assumed to extend across both outer panels
for the Megafly (panels 1, 2, 4, and 5 in Fig. 1) and across only the
outermost panels for the microparafoil. The control sensitivities
(CL;i and CD;i ) are normalized so that a control input of 1 produces a
5 deg =s turn rate for the Megafly and a 20 deg =s turn rate for the
microparafoil. The Megafly turn rate limit was chosen based on the
turn rate limit specified in [4]. The microparafoil turn rate limit was
chosen to avoid a nonlinear spiraling behavior at high brake
Fig. 3 Mapping left and right brake to steady-state turn rate. deflections. Rate limits are imposed on the brake deflections so that it
 1318 WARD, COSTELLO, AND SLEGERS
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

Fig. 4 Megafly and microparafoil systems.

8
takes 5 s to reach full brake deflection for the Megafly, and 1 s to reach > zk  yk  nk  B
full brake deflection for the microparafoil. >
> p
























< n  et=e n  1  e2t=e
k k1 k
(24)
> k  N0;
N 
>
B. Simulated Measurement Parameters >
:
Measurements from a GPS unit are assumed to be available for all B  N0;
B 
flights, and a heading measurement is added for some flights. GPS
sensor errors are modeled as exponentially correlated Gaussian noise The observer requires estimates of the variances of the process
and heading sensor error is modeled as Gaussian white noise [22,23]. noise and measurement noise components. The measurement noise
In addition, each sensor is given a bias that is constant over each flight variances are set to the actual values for each measurement. For the
and follows a Gaussian distribution over a series of flights. The form process noise, standard deviations of 1 m=s for the aerodynamic
of the measurement signals is given in Eq. (24), where yk is the actual velocity perturbation V0 and 0:001 rad=s for the azimuthal rate
value, zk is the measured value, nk is the measurement noise, and B is perturbation _ 0 are assumed. If only GPS measurements are
the measurement bias. Sensor parameters were selected to represent available, the wind perturbations Vwx ; Vwy  are assumed to have
low-cost, commercially available sensors. The position measure- standard deviations of 0:0001 m=s, but if a heading measurement is
ments are given a standard deviation of 2 m for the noise, a standard added, this value is increased to 0:01 m=s. The standard deviations
deviation of 1 m for the bias, and a time constant of 20 s. The velocity for the GPS velocity measurement biases (B x, _ B y)
_ are set to
measurements are given a standard deviation of 0:2 m=s the noise, a 1E  5 m=s, and the standard deviation for the heading measure-
standard deviation of 0:1 m=s for the bias, and a time constant of 1 s. ment bias is set to 1E  6 rad.
The heading measurement noise is given a standard deviation of
2 deg and the bias is given a standard deviation of 10 deg. The C. Wind Model
sampling intervalt is set for a 4 Hz update rate for all measurements A simplified Dryden turbulence model [24] is used to simulate
atmospheric winds with the wind magnitude and direction
Table 1 Megafly and microparafoil physical parameters represented as exponentially correlated Gaussian random variables.
The standard deviation of the wind magnitude is assumed to be 10%
Parameter Megafly Microparafoil Units of the mean wind magnitude, and the standard deviation of the wind
d 30 1.2 m direction is assumed to be 10 deg. The time constants for the wind

120 70 deg magnitude and direction were both chosen to be 10 s. Vertical winds
b 51.8 1.37 m are not simulated.
c 16.3 0.64 m
SC 840 0.93 m2
SP 9.0 0.033 m2 Table 2 Megafly and microparafoil
Total mass 13,605 2.37 kg aerodynamic parameters
IXX 361,650 0.42 kg m2
Coefficient Megafly Microparafoil
IYY 328,200 0.40 kg m2
IZZ 131,400 0.05 kg m2 CL0;i 0.45 0.30
IXZ 2; 283 0.03 kg m2 CL;i 2.2 0.723
A 293 0.01 kg CL3;i 1:0 0:35
B 3,133 0.05 kg CL;i 0.125 0.0065
C 9,858 0.40 kg CD0;i 0.15 0.12
P 78,960 0.008 kg m2 CD2;i 1.5 0.90
Q 102,560 0.007 kg m2 CD;i 0.125 0.010
R 84,590 0.002 kg m2 CD;P 0.8 0.3
 WARD, COSTELLO, AND SLEGERS 1319

model of the azimuthal rate dynamics that was tuned to match the
nominal simulation model. This represents the fixed model that
would normally be derived from an initial set of flight tests and
loaded on to the flight computer before each drop. The other two
trajectories were generated using azimuthal rate models derived with
the in-flight system identification procedure with and without the
specified sensor errors.
Figure 6 demonstrates the tradeoff with in-flight identification.
The in-flight identification process uses data that is corrupted by
sensor noise and wind disturbances, so a perfect realization of the
system dynamics cannot be achieved. If there is very little variability
in the parafoil dynamics, then a precomputed model of the nominal
system will always be a good representation, but if the level of
variability is high enough relative to the sensor errors and wind
disturbances, then a better representation will be obtained with a
dynamic model identified in-flight. A challenge for autonomous
airdrop system designers is to understand when it is preferable to
perform system identification in-flight and when a precomputed
Fig. 5 Geometry parameters. model should be used.
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

VI. Example Wind and Wind-Relative

V. Example Trajectory Predictions Dynamics Estimation
Figure 6 shows 10 s of a simulated trajectory of the microparafoil The turn rate dynamics during a 50% left brake step input for the
with a 50% step left brake input. Figure 6a represents the nominal Megafly and microparafoil simulation models are shown in Fig. 7.
simulation model, and Fig. 6b represents a modified model with a There is no wind in these cases, so the azimuthal rate corresponds to
15% right control bias and the left brake effectiveness reduced by the wind-relative azimuthal rate. The azimuthal rate response of the
15%. The fixed linear model trajectory was generated from a linear Megafly and microparafoil simulation models both appear to be

40

35

30 Simulation
25 Fixed Linear Model
In−Flight Est. (No Sensor Error)
20 In−Flight Est. (Sensor Error)
North (m)

15

10

−5

−10
0 10 20 30 40 50 60 70 80 90
East (m)
a) Nominal simulation model

40

35

30 Simulation
Fixed Linear Model
25
In−Flight Est. (No Sensor Error)
20 In−Flight Est. (Sensor Error)
North (m)

15

10

−5

−10
0 10 20 30 40 50 60 70 80 90
East (m)
b) Modified simulation model
Fig. 6 Microparafoil example trajectories with a) nominal simulation model, and b) with 15% right control bias and left brake effectiveness reduced
by 15%.
 1320 WARD, COSTELLO, AND SLEGERS

2
Azimuthal Rate
1 Heading Rate
Side Slip Rate
0

Angular Rate (deg/s)

−1

−2

−3

−4

−5
0 5 10 15 20 25 30
Time (s)
a) Megafly

2
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

0
Azimuthal Rate
−2 Heading Rate
Angular Rate (deg/s)

Side Slip Rate

−4

−6

−8

−10

−12
0 5 10 15 20 25 30
Time (s)
b) Microparafoil
Fig. 7 Megafly (part a) and microparafoil (part b) azimuthal and heading rate response.

approximately first order. However, the simulation models have very selected based on flight-test experience. While every parameter listed
different sideslip dynamics. The microparafoil turns with very little in Tables 1 and 2 has an associated uncertainty, the same overall
sideslip, but the tendency of the Megafly to turn with a significant effects are achieved by varying a few key parameters. The wind
sideslip causes the heading rate to initially lead the azimuthal rate. magnitude was set to vary from zero to the expected airspeed of each
Wind-relative azimuthal rate estimates obtained from the Kalman system to cover the entire operational envelope. All of the random
filter observer during a left brake step input using the Megafly and variables in the Monte Carlo simulation were uniformly distributed
microparafoil simulation models are shown in Fig. 8. In both cases, with the following ranges: 15% asymmetric control bias, 25%
the mean wind magnitude is 10 m=s. The mean squared error for the left and right control sensitivities, 25% payload weight, and 5%
microparafoil is reduced by 37% with the addition of the heading CL0 and CD0 for the entire canopy. The mean wind speed was varied
sensor, but the error for the Megafly is only reduced by 15% with the uniformly from 0–10 m=s for the microparafoil and from 0–20 m=s
addition of the heading sensor. It is expected that the benefit from the for the Megafly. One hundred cases were run for each parafoil
heading sensor is reduced for the Megafly because of the increased system. All of the errors in the quantities estimated in-flight from the
amount of sideslip compared with the microparafoil, as shown in simulated measurement data are compared with the errors that would
Fig. 7. A second factor contributing to the large reduction in results from the use of a fixed model that would normally be derived
estimation error for the microparafoil is that the wind-relative from an initial set of flight tests and loaded on to the flight computer
azimuthal rate estimation with GPS is very sensitive to the wind before each drop. These fixed model errors also represent the
estimate for the microparafoil because the forward flight speed of the perturbations in each estimated quantity that are induced by the
microparafoil (9 m=s) is close to the wind speed in this case, while assumed levels of model uncertainty.
the forward speed of the Megafly (20 m=s) is about double the Figures 11 and 12 show the error distributions for the estimated
wind speed. time constant of the azimuthal rate dynamics for the Megafly and
The associated wind estimates for these simulations are shown in microparafoil. Estimates using GPS only and using GPS in conjunc-
Figs. 9 and 10. Note that with GPS only, the wind estimate is a slowly tion with a heading measurement are shown. The time constant error
varying average magnitude and direction, while the addition of the is the difference between the estimated time constant and an optimal
heading sensor provides a significant improvement in the wind time constant obtained by fitting the actual wind-relative azimuthal
estimation for both the Megafly and the microparafoil. rate data from the nonlinear simulation model. The values for the
time constant error are on the order of 20%. Figure 13 demonstrates
the effect of a 20% change in time constant on the turn rate response
VII. Monte Carlo Simulation Results to a step input. Comparison of this figure to the example azimuthal
A. Results for Assumed Levels of Model Uncertainty, Sensor Noise, rate estimates in Fig. 8, provides some qualitative insight in the
and Wind difficulties of in-flight estimation of dynamic parameters with typical
A Monte Carlo simulation was run to assess the benefits of levels of sensor noise and turbulence. Figures 11 and 12 show that the
performing the system identification in-flight compared with using a standard deviation of the time constant estimation error is reduced by
fixed model of the nominal system. Levels of model uncertainty were the addition of the heading measurement for both parafoils, but the
 WARD, COSTELLO, AND SLEGERS 1321

4
Actual
Estimated (GPS)
2 Estimated (GPS+psi)

Azimuthal Rate (deg/s)

0

−2

−4

−6

−8
0 5 10 15 20 25 30
Time (s)
a) Megafly

5
Actual
Estimated (GPS)
Estimated (GPS+psi)
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

0
Azimuthal Rate (deg/s)

−5

−10

−15
0 5 10 15 20 25 30
Time (s)
b) Microparafoil
Fig. 8 Megafly (part a) and microparafoil (part b) example wind-relative azimuthal rate estimation.

effect on the mean of the time constant error is different. There is a lag for the time constant when the heading sensor is added. The
associated with the filtered azimuthal rate estimate, so it is expected microparafoil turns with very little sideslip, so no shift in the time
that the estimated time constant would be generally higher than the constant error is observed when the heading sensor is added.
optimal value, resulting in a positive shift to the mean error. The Figures 14 and 15 show the error distributions for the in-flight
dramatic shift in the mean time constant error for the Megafly can be control bias and airspeed estimations. The techniques for estimating
explained by the sideslip behavior shown in Fig. 7. The tendency of the control bias and airspeed do not depend on the heading
the heading rate to initially lead the azimuthal rate for the Megafly measurement, so the results are the same for both sensor cases. Note
causes the wind-relative azimuthal rate estimate to appear to respond that the airspeed estimation error is skewed for the Megafly. This is
artificially quickly to control input, which results in lower estimates because the airspeed is estimated during a turn, and the Megafly

12
Wind Magnitude (m/s)

11

10

6
0 10 20 30 40 50 60

40
Actual
Wind Direction (deg)

30 Estimated (GPS)
Estimated (GPS+psi)
20

10

−10
−20
0 10 20 30 40 50 60
Time (s)
Fig. 9 Megafly example wind estimation.
 1322 WARD, COSTELLO, AND SLEGERS

11

Wind Magnitude (m/s)

10

8
Actual
7 Estimated (GPS)
Estimated (GPS+psi)
6
0 10 20 30 40 50 60
30
Wind Direction (deg)

20

10

−10
0 10 20 30 40 50 60
Time (s)
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

Fig. 10 Microparafoil example wind estimation.

50 50

40 40
No. Cases

No. Cases

30 30

20 20

10 10

0 0
−40 −20 0 20 40 60 −40 −20 0 20 40 60
Time Constant Error (%) Time Constant Error (%)
a) GPS b) GPS + ψ
Fig. 11 Monte Carlo results: Megafly time constant error.

50 50

40 40
No. Cases

No. Cases

30 30

20 20

10 10

0 0
−20 0 20 40 −20 0 20 40
Time Constant Error (%) Time Constant Error (%)
a) GPS b) GPS + ψ
Fig. 12 Monte Carlo results: microparafoil time constant error.

system exhibits a tendency to slow down while turning due to the 1.2
large span of the canopy brakes. The microparafoil showed no such
tendency. 1 τopt −20%
Figure 16 shows the steady-state turn rate response estimation
Azimuthal Rate (deg/s)

0.8
error. From each Monte Carlo run, an estimation of the steady-state
turn rate vs control input is obtained. By subtracting the actual 0.6 τopt
steady-state turn rate vs control input produced by the simulation
model for that particular run, a discrete value of turn rate error at each 0.4 τopt +20%
control input is obtained. Compiling the results for all of the Monte
Carlo runs, these discrete values become distributions of turn rate 0.2
estimation error at each control input. By excluding the upper and
lower 5% of these distributions, upper and lower bounds enclosing 0
90% of the cases at each control input are determined. These bounds −0.2
are plotted in Fig. 16. The average widths of each of these 90% turn 0 5 10 15 20 25 30
rate error bounds are reported in the last row of Table 3 as mean 90% Time (s)
turn rate error bound. The results show that the addition of a heading Fig. 13 Microparafoil step response demonstrating time constant
sensor provides a slight improvement in turn rate estimation, but errors.
 WARD, COSTELLO, AND SLEGERS 1323

40 40

30 30

No. Cases

No. Cases
20 20

10 10

0 0
−10 −5 0 5 10 −15 −10 −5 0 5
Control Bias Estimation Error (%) Flight Speed Estimation Error (%)
Fig. 14 Monte Carlo results: Megafly control bias and airspeed estimation error.

50 50

40 40
No. Cases

No. Cases
30 30

20 20
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

10 10

0 0
−10 −5 0 5 10 −10 −5 0 5 10 15
Control Bias Estimation Error (%) Flight Speed Estimation Error (%)
Fig. 15 Monte Carlo results: microparafoil control bias and airspeed estimation error.

overall the estimation of steady-state turn rate is robust for both airspeed error is cut in half. Control bias error is reduced by more than
sensor cases. a factor of 3 for both parafoils. Turn rate error is reduced by a factor of
All of the results from the Monte Carlo simulation are summarized 6 for the microparafoil, and by more than a factor of 4 for the Megafly.
in Table 3. It is clearly beneficial to estimate all of the steady-state The estimation of the azimuthal time constant in-flight with a
characteristics of the system (airspeed, control bias, and turn rate) in- heading measurement produces similar standard deviations to the
flight. In-flight estimation of the airspeed for the microparafoil fixed model, though there is a significant increase in the mean time
significantly reduces the airspeed error, and for the Megafly the constant error. For the case where only GPS measurements are used,

50
Estimated (GPS)
40 Estimated (GPS+psi)
Fixed Model
30
Turn Rate Error (%)

20

10

−10

−20

−30

−40
−100 −80 −60 −40 −20 0 20 40 60 80 100
Asymmetric Brake Deflection (%)
a) Megafly
30

20
Turn Rate Error (%)

10

−10

−20 Estimated (GPS)

Estimated (GPS+psi)
Fixed Model
−30
−100 −80 −60 −40 −20 0 20 40 60 80 100
Asymmetric Brake Deflection (%)
b) Micro parafoil
Fig. 16 Monte Carlo results: Megafly (part a) and microparafoil (part b) steady-state turn rate estimation error.
 1324 WARD, COSTELLO, AND SLEGERS

Table 3 Monte Carlo simulation resultsa

Microparafoil Megafly
Error description Fixed model GPS GPS  Fixed model GPS GPS 
Mean flight-test speed 0:2 1.1 1:9 3:1
Standard deviation flight-test speed 6.3 3.8 7.3 2.9
Mean control bias 0.0 0.0 0.0 0.0
Standard deviation control bias 6.8 2.1 7.1 2.1
Mean time constant 2:1 8.2 8.2 4.1 17.3 6:3
Standard deviation time constant 5.6 9.5 5.0 6.5 13.9 9.8
Mean 90% turn rate error bound 33.1 5.7 5.0 36.4 8.8 7.3
a
Errors are percentages.

in-flight estimation of the time constant results in larger errors than (perfect knowledge of the system) to 1.5 times the given ranges. The
the use of a fixed model. This indicates that in-flight estimation of the standard deviations for all of the sensor errors were scaled from zero
time constant is not generally beneficial when only GPS measure- (perfect sensors) to 2 times the given levels. The microparafoil
ments are available. simulation model was used for these runs. Cases were run with no
wind and with constant mean wind magnitudes of 5 m=s (half
system airspeed) and 10 m=s (equal to system airspeed). From the
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

B. Sensitivity to Levels of Model Uncertainty, Sensor Noise, results, the boundaries where the in-flight system identification and
and Wind the fixed model produce the same errors in the steady-state and
An additional set of Monte Carlo simulations was run to transient turn response were calculated as a function of the model
investigate the sensitivity of the results to changes in the assumed uncertainty and sensor error levels at the three different wind levels.
levels of sensor error, model uncertainty, and wind levels. The model Figure 17 explains how these boundaries are plotted. With model
uncertainty levels used in the first simulation were scaled from zero uncertainty on the y axis and sensor error on the x axis, the region
above and to the left of the boundaries represents the space where the
model uncertainty is large enough and the sensor data is good enough
1.5 that it is better to perform the system identification in-flight as
opposed to using a fixed model of the flight dynamics. The actual
results are shown in Figs. 18 and 19.
In−Flight Figure 18 shows that if there is any flight-to-flight variation in the
system at all, then it is better to estimate the steady-state
Scaling Reference Level

Identification
of Model Uncertainty

1 is Better characteristics of the system in-flight. It also shows that with the
addition of a heading measurement, the in-flight estimation of the
steady-state characteristics is made more robust to large levels of
sensor noise. Figure 19 shows that when only GPS measurements are
0.5 available it is difficult to obtain a good estimation of the transient
Fixed Model
is Better
response with moderate to high wind levels. The boundary for the
10 m=s wind case lies off the chart because it is better to use a fixed
model over the entire range of sensor error and model uncertainty
considered. For the results with the heading sensor included, the
0
0 0.5 1 1.5 2 boundaries run through the assumed levels of model uncertainty and
Scaling Reference Level of Sensor Error sensor errors, implying that it makes little difference if the time
Fig. 17 Explanation of in-flight identification vs fixed model bound- constant is estimated in-flight or if a precomputed time constant is
aries. used.

1
Expected Level of
0.9
Model Uncertainty
and Sensor Error
0.8
Scaling Model Uncertainty

0.7

0.6

0.5
No Wind, GPS only
0.4 10 m/s Wind, GPS only
No Wind, GPS+ψ
0.3
10 m/s Wind, GPS+ψ
0.2

0.1

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Scaling Sensor Error
Fig. 18 In-flight identification vs fixed model boundaries for steady-state response.
 WARD, COSTELLO, AND SLEGERS 1325

1.5

Scaling Model Uncertainty

1
Expected Level of
Model Uncertainty
and Sensor Error

No Wind, GPS only

0.5 5 m/s Wind, GPS only
10 m/s Wind, GPS only
No Wind, GPS+ψ
5 m/s Wind, GPS+ψ
10 m/s Wind, GPS+ψ
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Scaling Sensor Error
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

Fig. 19 In-flight identification vs fixed model boundaries for transient response.

System identification relies on state estimates that are degraded by adaptive control scheme for airdrop systems is critically dependent
atmospheric turbulence and sensor errors. The traditional scenario on the incorporation of high quality heading estimates.
when system identification is performed on the ground takes
advantage of the ability to estimate parameters over multiple
maneuvers and multiple flights so that the effects of sensor errors and References
turbulence are averaged out. The results presented here show that it is [1] Calise, A., and Preston, D., “Design of a Stability Augmentation
always beneficial to estimate steady-state quantities in-flight with System for Airdrop of Autonomous Guided Parafoils,” AIAA
reasonable levels of model uncertainty, sensor noise, and wind Paper 2006-6776, May 2006.
because the estimation is performed by averaging over a series of [2] Calise, A., and Preston, D., “Swarming/Flocking and Collision Avoid-
measurements. On the other hand, transient characteristics estimated ance for Mass Airdrop of Autonomous Guided Parafoils,” Journal of
in-flight over the small windows of data during maneuvers are much Guidance, Control, and Dynamics, Vol. 31, No. 4, 2008, pp. 1123–
more sensitive to sensor noise and turbulence, and the results show 1132.
that it is not always beneficial to estimate these transient doi:10.2514/1.28586
[3] Carter, D., George, S., Hattis, P., Singh, L., and Tavan, S., “Autonomous
characteristics in-flight. In other words, for transient characteristics, Guidance, Navigation, and Control of Large Parafoils,” AIAA
the degradation in the quality of in-flight estimates from sensor noise Paper 2005-1643, May 2005.
and turbulence is comparable to the degradation in the quality of [4] Carter, D., George, S., Hattis, P., McConley, M., Rasmussen, S., Singh,
fixed estimates from model uncertainty. L., and Tavan, S., “Autonomous Large Parafoil Guidance, Navigation,
and Control System Design Status,” AIAA Paper 2007-2514,
May 2007.
VIII. Conclusions [5] Jann, T., “Advanced Features for Autonomous Parafoil Guidance,
Dynamic characteristics of airdrop systems vary substantially Navigation and Control,” AIAA Paper 2005-1642, May 2005.
from drop to drop, more than most other air vehicles. This paper [6] Carter, D., Singh, L., Wholey, L., Rasmussen, S., Barrows, T., George,
addresses the question of whether it is more appropriate to use S., McConley, M., Gibson, C., Tavan, S., and Bagdonovich, B., “Band-
precomputed dynamic characteristics of an airdrop system inside an Limited Guidance and Control of Large Parafoils,” AIAA Paper 2009-
2981, May 2009.
autonomous control law or whether it is better to estimate these
[7] Slegers, N., and Costello, M., “Model Predictive Control of a Parafoil
characteristics in-flight. A simulation study was performed using a 6 and Payload System,” Journal of Guidance, Control, and Dynamics,
degrees of freedom nonlinear parafoil and payload simulation with Vol. 28, No. 4, 2005, pp. 816–821.
an associated estimation algorithm. Simulation results with typical doi:10.2514/1.12251
levels of sensor errors and model uncertainty demonstrate that the [8] Kaminer, I., and Yakimenko, O., “Development of Control Algorithm
system identification procedure provides significantly better for the Autonomous Gliding Delivery System,” AIAA Paper 2003-
estimates of the system dynamics than a precomputed, nominal 2116, May 2003.
model. Errors in predicted steady-state turn rate were reduced by a [9] Calise, A., Preston, D., and Ludwig, G., “Modeling for Guidance and
factor of more than 4 for the Megafly and by a factor of 6 for the Control Design of Autonomous Guided Parafoils,” AIAA Paper 2007-
microparafoil. It was found by varying the assumed levels of sensor 2560, May 2007.
[10] Jann, T., and Strickert, G., “System identification of a Parafoil-Load
errors, model uncertainty, and wind that the in-flight identification of Vehicle–Lessons Learned,” AIAA Paper 2005-1663, May 2005.
the steady-state characteristics of airspeed, control bias, and turn rate [11] Yakimenko, O., and Statnikov, R., “Multicriteria Parametrical identifi-
behavior is always beneficial. This is because steady-state quantities cation of the Parafoil-Load Delivery System,” AIAA Paper 2005-1664,
are estimated by averaging over a series of measurements so that the May 2005.
detrimental effects of sensor noise and turbulence are reduced. In [12] Yakimenko, O., “On the Development of a Scalable 8-DoF Model for a
contrast, the in-flight identification of the azimuthal rate time Generic Parafoil-Payload Delivery Systems,” AIAA Paper 2005-1665,
constant was found to produce results that were comparable to a May 2005.
fixed, nominal model if a heading measurement is available, and the [13] Calise, A., and Preston, D., “Approximate Correction of Guidance
Commands for Winds,” AIAA Paper 2009-2997, May 2007.
in-flight identification results when using only GPS were not as good
[14] Slegers, N., Beyers, E., and Costello, M., “Use of Variable Incidence
as the fixed model for moderate to high wind levels. This result stems Angle for Glide Slope Control of Autonomous Parafoils,” Journal of
from the more general conclusion that with typical levels of sensor Guidance, Control, and Dynamics, Vol. 31, No. 3, 2008, pp. 585–596.
noise and atmospheric turbulence it is very difficult to obtain reliable doi:10.2514/1.32099
azimuthal rate estimates for an airdrop system during a maneuver [15] Berland, J., Barber, J., Gargano, B., and Bagdonovich, B. P.,
without a heading estimate. This suggests that the success of any “Autonomous Precision Delivery of 42,000 Pounds (19,000 kg) Under
 1326 WARD, COSTELLO, AND SLEGERS

One Parachute,” AIAA Paper 2009-2928, May 2009. Landing System Using a Parafoil,” NASA TM-4599, Jul. 1994.
[16] Lissaman, P., and Brown, G., “Apparent Mass Effects on Parafoil [21] Madsen, C. M., and Cerimele, C. J., “Flight Performance, Aero-
Dynamics,” AIAA Paper 93-1236, May 1993. dynamics, and Simulation Development for the X-38 Parafoil Test
[17] Chui, C. K., and Chen, G., Kalman Filtering with Real-Time Program,” AIAA Paper 2003-2108, May 2003.
Applications, 3rd ed., Springer–Verlag, New York, 1999, Chap. 8. [22] “Final Report: Development and Demonstration of a Ram-Air Parafoil
[18] Iacomini, C. S., and Cerimele, C. J., “Lateral-Directional Aerody- Precision Guided Airdrop System, Vol. 3,” Draper Lab., Army
namics from a Large Scale Parafoil Test Program,” AIAA Paper 99- Contract DAAK60-94-C-0041, National Soldier Center, Oct. 1996.
1731, Jun. 1999. [23] Dellicker, S., Benney, R., and Brown, G., “Guidance and Control for
[19] Jann, T., “Aerodynamic Model identification and GNC Design Flat-Circular Parachutes,” Journal of Aircraft, Vol. 38, No. 5,
for the Parafoil-Load System ALEX,” AIAA Paper 2001-2015, October 2001, pp. 809–817.
May 2001. doi:10.2514/2.2865
[20] Murray, J., Sim, A., Neufeld, D., Rennich, P., Norris, S., and Hughes, [24] “Military Standard, Flying Qualities of Piloted Aircraft,” MIL-STD-
W., “Further Development and Flight Test of an Autonomous Precision 1797A, Jan. 1990.
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030
 This article has been cited by:

1. Casper Dek, Jean-Luc Overkamp, Akke Toeter, Tom Hoppenbrouwer, Jasper Slimmens, Job van Zijl, Pietro Areso Rossi,
Ricardo Machado, Sjef Hereijgers, Veli Kilic, Marc Naeije. 2020. A recovery system for the key components of the first
stage of a heavy launch vehicle. Aerospace Science and Technology 100, 105778. [Crossref]
2. Haitao Gao, Jin Tao, Matthias Dehmer, Frank Emmert-Streib, Qinglin Sun, Zengqiang Chen, Guangming Xie, Quan
Zhou. 2020. In-flight Wind Field Identification and Prediction of Parafoil Systems. Applied Sciences 10:6, 1958. [Crossref]
3. Martin R. Cacan, Mark Costello, Michael Ward, Edward Scheuermann, Michael Shurtliff. 2019. Human-in-the-loop
control of guided airdrop systems. Aerospace Science and Technology 84, 1141-1149. [Crossref]
4. R Rakesh, R Harikumar. Autonomous Airdrop System Using Small-Scale Parafoil 1-6. [Crossref]
5. Zikang Su, Chuntao Li, Honglun Wang. 2019. Barrier Lyapunov function-based robust flight control for the ultra-low
altitude airdrop under airflow disturbances. Aerospace Science and Technology 84, 375-386. [Crossref]
6. Martin R. Cacan, Mark Costello. 2018. Adaptive Control of Precision Guided Airdrop Systems with Highly Uncertain
Dynamics. Journal of Guidance, Control, and Dynamics 41:5, 1025-1035. [Abstract] [Full Text] [PDF] [PDF Plus]
7. Shuzhen Luo, Panlong Tan, Qinglin Sun, Wannan Wu, Haowen Luo, Zengqiang Chen. 2018. In-flight wind identification
Downloaded by BEIHANG UNIVERSITY on June 2, 2020 | http://arc.aiaa.org | DOI: 10.2514/1.49030

and soft landing control for autonomous unmanned powered parafoils. International Journal of Systems Science 49:5,
929-946. [Crossref]
8. Jie Chen, Cunbao Ma, Dong Song. 2018. Cargo blocking failure analysis, simulation, and safety control of transport aircraft
with continuous heavy airdrop. International Journal of Advanced Robotic Systems 15:1, 172988141875704. [Crossref]
9. Martin Cacan, Edward Scheuermann, Michael Ward, Mark Costello, Nathan Slegers. 2017. Utilizing ground-based
LIDAR measurements to aid autonomous airdrop systems. Proceedings of the Institution of Mechanical Engineers, Part G:
Journal of Aerospace Engineering 231:10, 1763-1778. [Crossref]
10. Martin R. Cacan, Mark Costello. Hyperadaptive Control of Precision Guided Airdrop Systems Subject to Large Deviations
in Nominal Flight Behavior . [Citation] [PDF] [PDF Plus]
11. Bin Xu. 2017. Disturbance Observer-Based Dynamic Surface Control of Transport Aircraft With Continuous Heavy Cargo
Airdrop. IEEE Transactions on Systems, Man, and Cybernetics: Systems 47:1, 161-170. [Crossref]
12. Bin Xu, Jie Chen. 2016. Review of modeling and control during transport airdrop process. International Journal of Advanced
Robotic Systems 13:6, 172988141667814. [Crossref]
13. Bingbing Li, Liying Yang, Yuqing He, Jianda Han. Energy-based controller decoupling of Powered Parafoil Unmanned
Aerial Vehicle 313-320. [Crossref]
14. Martin R. Cacan, Edward Scheuermann, Michael Ward, Mark Costello, Nathan Slegers. 2015. Autonomous Airdrop
Systems Employing Ground Wind Measurements for Improved Landing Accuracy. IEEE/ASME Transactions on
Mechatronics 20:6, 3060-3070. [Crossref]
15. . Precision Aerial Delivery Systems: Modeling, Dynamics, and Control . [Abstract] [PDF] [PDF Plus] [Supplementary
Material]
16. . References 871-900. [Citation] [PDF] [PDF Plus]
17. Michael Ward, Mark Costello. 2013. Adaptive Glide Slope Control for Parafoil and Payload Aircraft. Journal of Guidance,
Control, and Dynamics 36:4, 1019-1034. [Abstract] [Full Text] [PDF] [PDF Plus]
18. Sean Culpepper, Michael B. Ward, Mark Costello, Keith Bergeron. Adaptive Control of Damaged Parafoils . [Citation]
[PDF] [PDF Plus]
19. Michael B. Ward, Mark Costello. Autonomous Control of Parafoils Using Upper Surface Spoilers . [Citation] [PDF]
[PDF Plus]
20. Thomas A. Herrmann, Michael B. Ward, Mark Costello, Nathan Slegers. Utilizing Ground-Based LIDAR for
Autonomous Airdrop . [Citation] [PDF] [PDF Plus]
21. M. Ward, M. Costello. Adaptive glide slope control for autonomous airdrop systems 2557-2562. [Crossref]
22. 2010. Adaptive control and signal processing literature survey (No. 21). International Journal of Adaptive Control and Signal
Processing 24:12, 1107-1112. [Crossref]