You are on page 1of 8

2013 IEEE International Conference on Robotics and Automation (ICRA)

Karlsruhe, Germany, May 6-10, 2013

Preliminary Experiments in Comparative Experimental Identification of Six


Degree-Of-Freedom Coupled Dynamic Plant Models for Underwater Robot
Vehicles
Stephen C. Martin and Louis L. Whitcomb
Department of Mechanical Engineering
Johns Hopkins University

Abstract— This paper addresses the modeling and experimen- of this type of vehicle is shown in Fig.1. We report a
tal identification of six degree-of-freedom (6-DOF) coupled non- comparative experimental evaluation of six different candidate
linear second order plant models for low-speed, fully-actuated, plant models whose unknown plant parameters are estimated
and neutrally buoyant open-frame underwater vehicles. We
report a comparative experimental evaluation of six different experimentally with data from free-motion vehicle trials.
candidate plant models whose unknown plant parameters are Few previously reported studies of the dynamics of this
estimated from data obtained in free-motion vehicle trials. We class of vehicles have reported experimentally validated
report an experimental evaluation of the performance of each plant models. Most studies that have reported experimentally
of the six different 6-DOF coupled non-linear finite-dimensional identified models have addressed decoupled one-degree-of-
plant models for underwater vehicles estimated by total least
squares (TLS) by comparing the mean absolute error between freedom plant models [6], [21], [5], [18], [25], [7], [2], [3], [1].
the experimentally observed vehicle velocities and the velocities Fewer reported experimental studies have addressed coupled
obtained by a numerical simulation of the experimentally multi-degree-of-freedom plant models [10], [14], [15], [23].
identified plant models. We also report a cross-validation which This study provides the following contributions: First, This
evaluates the ability of a plant model to accurately reproduce study is the first reported experimental 6-DOF fully coupled
observed plant velocities for experimental trials differing from
the trial from which the plant model parameters were estimated. plant model identification of low-speed, fully-actuated, and
We conclude that plant models including fully parametrized neutrally buoyant underwater vehicles. In addition, this study
coupled quadratic drag terms perform best overall in cross- is the first reported experimental plant identification for this
validation. This study has the following contributions: It is the class of underwater vehicles in free-flight experiments. This
first reported experimental 6-DOF fully-coupled plant model result differs from previously reported studies which primarily
identification and cross-validation of low-speed, fully-actuated,
and neutrally buoyant underwater vehicles; it is the first focus on decoupled plants, or 3-DOF coupled dynamical
experimental 6-DOF plant model identification for this class plants. Second, this study reports the first reported use of total
of underwater vehicles during free-flight experiments; and it is least squares (TLS) to perform 6-DOF model identification
the first reported use of TLS to perform 6-DOF experimental of underwater vehicles. Third, this study reports the first
plant model identification of an underwater vehicle. cross-validation of identified 6-DOF plant models.
I. I NTRODUCTION This paper is organized as follows: Section I-A reviews
Underwater vehicle control research is presently limited by previously reported results on finite-dimensional modeling of
the paucity of explicit experimentally validated plant models. underwater vehicles. Section II defines the plant model equa-
Although the generic form of approximate lumped parameter tions of motion employed in this paper. Section IV describes
finite-dimensional plant models for underwater vehicles is the experimental methodology used in these experiments.
known, actual plant parameters (e.g. hydrodynamic added Section V reports a comparative analysis of experimentally
mass, buoyancy, and drag parameters) must be determined identified plant models of the 6-DOF coupled plant of
experimentally. Improved experimentally validated plant the Johns Hopkins University Remotely Operated Vehicle
dynamical models for underwater vehicles will enable the (JHU ROV) using TLS. Section VI reports a comparative
development of model-based controllers and may also enable cross-validation of the experimentally identified plant models
high-fidelity numerical simulations to predict the performance reported in Section V. Section VII discusses and summarizes.
of underwater vehicles under a great variety of disparate A. Related Literature
missions, vehicle control laws, and operating conditions. The most commonly accepted finite-dimensional dynamical
This paper addresses the modeling and experimental identi- models for submarine vehicles trace their lineage to studies
fication of six degree-of-freedom (6-DOF) coupled nonlinear beginning in the 1950s at the U.S. Navy’s David Taylor
second order plant models for low-speed maneuvering of Model Basin [13], [17], [12], [9]. Many finite-dimensional
fully-actuated open-frame underwater vehicles. An example hydrodynamic model identifications are performed using
planar motion mechanism (PMM) systems. These systems
S. C. Martin and L. L. Whitcomb are with the Department of Mechanical require a substantial amount of infrastructure to be constructed
Engineering (ME) and Laboratory for Computational Sensing and Robotics
(LCSR), Johns Hopkins University (JHU), Baltimore, Maryland, USA. Author and maintained [10], [2], [1]. In [10], the author reports
email addresses: smarti32@jhu.edu and llw@jhu.edu. the identification of a high speed nonlinear coupled 6-DOF

978-1-4673-5643-5/13/$31.00 ©2013 IEEE 2962

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
dynamical plant model of a 1/3 scale model of the Deep that the parameter set with the lowest tracking error in self-
Submergence Rescue Vehicle (DSRV Scheme A) using validation may not be the parameter set that performs best
the David Taylor Model Basin PMM System. This report in cross-validation.
identified most of the hydrodynamic mass parameters, all Least squares methods for identifying the dynamical plant
dominant hydrodynamic drag terms, and a few quadratic drag model parameters of underwater vehicles typically require
terms relating the coupling between the different degrees-of- position, velocity, acceleration, and force/moment measure-
freedom. This report does not include a numerical simulation ments. Acceleration is often difficult to measure directly.
using the identified parameters to evaluate the performance of Differentiating velocity or position to estimate acceleration
the identified plant model. In [2], [1], the authors report the may increase the noise in the data, which can degrade the
use of a PMM to identify a decoupled dynamical plant of the quality of the identified model. In [25], the authors report
surge direction of the LAURS vehicle using ordinary least a comparative analysis of the adaptive and OLS techniques
squares and weighted least squares methods. The identified when applied to decoupled 1-DOF underwater robotic vehicle
model was validated by first computing an estimated force plant identification of the surge, sway, heave, and heading
vector computed from the identified model and experimentally DOFs. The reported adaptive identification technique does
measured velocities and then comparing the estimated force not require acceleration state measurements to estimate the
vector to the actual forces measured by the PMM system. unknown plant parameters of an underwater robotic vehicle.
Numerous studies have reported the identification of a The experimentally identified plant models were evaluated
decoupled hydrodynamic dynamical models, e.g. [6], [5], [24], by computing the error between the experimentally observed
[7], [3], of underwater vehicles during free-flight experiments vehicle velocities to the velocities obtained by a numerical
using least squares techniques. The authors employed the simulation of the identified plant models for a identical time-
identified plant model and experimentally commanded control varying applied thruster forces and moments.
force and moment data as inputs to numerical plant simula- II. F INITE - DIMENSIONAL P LANT OF AN U NDERWATER
tions that compute a simulated plant velocity. The quality of V EHICLE
the identified models typically was evaluated by comparing The general form of 6-DOF plant models for underwater
the experimentally measured velocity to the simulated plant vehicles representing the surge, sway, heave, pitch, roll, and
velocity predicated by the identified plant model. In [15], heading DOFs is
the authors report the identification of a surge, heave, and
heading coupled dynamical plant model for the Hugin 4500
Autonomous Underwater Vehicle (AUV), a torpedo-shaped τ6×1 = M v̇ + C(v)v + DQ(v)v + DLv + b( wb R), (1)
forward-flying vehicle, using ordinary least squares (OLS).
This report identified a partial set of hydrodynamic mass where τ6×1 ∈ IR6×1 is the vector of body-frame forces
parameters, five linear drag terms, and a five quadratic drag and moment applied to the vehicle in all six degrees-of-
terms relating the coupling between the lateral and heading freedom and τ6×1 = [f1 ; f2 ; f3 ; t4 ; t5 ; t6 ], v ∈ IR6×1 is the
directions. The authors compared the coupled dynamical plant vehicle body velocity and v = [v1 ; v2 ; v3 ; v4 ; v5 ; v6 ], v̇ ∈
with a 1-DOF decoupled plant model. The authors found that IR6×1 is the time derivative of vehicle body velocity and
the error between the numerical simulation’s velocity data v̇ = [v̇1 ; v̇2 ; v̇3 ; v̇4 ; v̇5 ; v̇6 ], M ∈ IR6×6 is the positive-definite
for the coupled plant model and experimentally observed vehicle mass matrix, C(v) ∈ IR6×6 is the skew-symmetric
velocity data is smaller than error between the simulation’s vehicle Coriolis matrix-valued function, DQ(v) ∈ IR6×6
velocity data and experimentally observed velocity data for is the positive semi-definite quadratic drag matrix-valued
the decoupled plant model. In [23], we report a comparative function, DL ∈ IR6×6 is the positive-definite linear drag
experimental evaluation of five different candidate 3-DOF matrix, and b( wb R) ∈ IR6×1 is the buoyancy force/moment
(surge, sway, and heading) plant models whose unknown
vector where wb R ∈ SO(3) is the rotation matrix from world
plant parameters are estimated experimentally from data
coordinates to vehicle body coordinates [11]. We adopt the
obtained in free-motion vehicle trials. We employed the
following subscript notation
parameter estimation techniques of TLS and OLS to identify
experimentally the unknown plant model parameters. We
found that fully coupled 3-DOF linear and quadratic drag 1 = Surge or X
models perform better than corresponding decoupled drag 2 = Sway or Y
models. In [3], the authors cross-validate the identified 1-DOF 3 = Heave or Z
(2)
yaw direction model by comparing the performance of the 4 = Roll
predicted vehicle velocity with the experimentally measured 5 = P itch
velocity on experimental trial data different from the one used 6 = Y aw or Heading.
to identify the plant model. Cross-validation is a technique
commonly employed which evaluates the performance of a Although the form of equation (1) is known, the actual
plant model to accurately reproduce observed plant velocities plant parameter (e.g. hydrodynamic added mass and drag
for experimental trials differing from the trials from which the parameters) generally must be determined experimentally.
plant model parameters were estimated. The authors report This Section defines an explicit form for (1).

2963

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
A. 6-DOF Coupled Mass and Coriolis Terms A. Definition of Ordinary Least Squares Problem
The coupled symmetric mass matrix is M where M ∈ Ordinary Least Squares (OLS) has been widely employed
IR6×6 . The Coriolis matrix C(v) ∈ IR6×6 is in plant parameter estimation for underwater robotic vehicles.
OLS is an optimal unbiased estimator when no error exists
J T ( M T vt + M ` vr )
 
03×3 in the input data and the observation vector is corrupted by
C(v) = (3)
J( MT vt + M` vr ) J T ( IT vr + M` vt ) additive zero-mean Gaussian noise [8].
Definition 1 (OLS Problem [16]): Given an overdeter-
where MT ,M` , and IT are defined in (4) and C(v) ∈ IR6×6 , mined set of m linear equations, Ax ≈ y, the ordinary least
where vt , vr , v̇t , and v̇t are defined as squares problem minimizes the error
    kŷ − yk2 . (10)
v1 v4
vt =  v2  , vr =  v5  (5) Once a minimizing ŷ where ŷ ∈ Image(A) is found, then
v3 v6 any solution x̂ such that Ax̂ = ŷ solves the OLS problem.
B. Solution to Ordinary Least Squares Using the Moore-
    Penrose Inverse
v̇1 v̇4
Theorem 1 (Solution to the OLS problem [16]): Given
v̇t =  v̇2  , v̇t =  v̇5  (6)
an overdetermined set of m linear equations y = Ax with
v̇3 v̇6
unknown parameter vector x, if A has full column rank,
then a solution to the OLS problem is
and J : IR3 → IR3×3 is the operator given by
x̂ = (AT A)−1 AT y. (11)

v1

0

−v3 v2
 This method is computationally simple but requires that A
J( v2 ) =  v3 0 −v1  . (7) has full column rank and assumes that no error is present in
v3 −v2 v1 0 the input data A.
C. Definition of Total Least Squares Problem
B. Hydrodynamic Drag Terms Definition 2 (TLS Problem [16]): A TLS solution is any
The linear drag matrix is DL where DL ∈ IR6×6 and the solution x̂ such that Âx̂ = ŷ where  and ŷ are determined,
general 6 × 6 quadratic drag matrix valued function of v is such that ŷ ⊂ Image(Â)

P6
DQ(v) = i=1 |vi |DQi (8) mink[A; y] − [Â; ŷ]kF = mink[4A; 4y]kF (12)
where kokF denotes the Frobenius norm.
where DQi ∈ IR6×6 , i = 1 · · · 6, are the set of six quadratic
D. Total Least Squares Solution Using the Singular Value
drag parameter matrices.
Decomposition (SVD)
C. Buoyancy Terms
Definition 3 (Definition of the SVD [16]): Let
The buoyancy force/moment vector b( wb R) ∈ IR6×1 is A ∈ IRm×n and y ∈ IRm×1 then the Singular Value
 b
 Decomposition (SVD) of [A; y] is defined in the following
b w Rẑβ manner, given a matrix [A, y] ∈ Rm×n+1 with rank n then
b( R) = (9)
w η × wb Rẑ the SVD is defined as
Pn
where β ∈ IR is the net buoyancy of the vehicle in Newtons [A; y] = i=1 σi uTi vi (13)
and η ∈ IR3 is the buoyancy-induced moment vector acting which can be written equivalently as
on the vehicle in N-m, and is given by η = rcg ng − rcb nb
where rcg ∈ IR3 and rcb ∈ IR3 are, respectively, the positions A = U SV T (14)
of the center of gravity and center of buoyancy in the vehicle where S is a matrix of the singular values of ordered from
body frame, in units of meters, nb ∈ IR is the total vehicle largest to smallest along the diagonal, and U and V are
buoyancy in Newtons, and ng ∈ IR is the total vehicle dry unitary matrices of size m and n + 1. A partitioned form of
weight in Newtons. Note that the parameters rcg , ng , rcb , these matrices are
and nb cannot be individually identified with submerged free-
S = Diag(σ1 , · · · , σn , 0) (15)
motion experiments, thus the sequel we will experimentally
1
identify the composite paramaters β and η — representing a where σi ∈ IR
total of four model parameters.  
U= u1 ··· um (16)
III. L EAST S QUARES M ETHODS
The following section describes the least squares methods where ui ∈ IRm×1
employed to identify the vector x in Ax ≈ y, where A ∈  
v1,1 ··· v1,n+1
Rm×n is the input data matrix and y ∈ Rm×1 is the output .. ..
V =
 .. 
(17)
data vector are experimental observations, and x ∈ Rn×1 is . . . 
the unknown parameter vector. vn+1,1 ··· vn+1,n+1
2964

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
     
m1,1 m1,2 m1,3 m1,4 m1,5 m1,6 m4,4 m4,5 m4,6
MT =  m1,2 m2,2 m2,3  , M` =  m2,4 m2,5 m2,6  , IT =  m4,5 m5,5 m5,6  (4)
m1,3 m2,3 m3,3 m3,4 m3,5 m3,6 m4,6 m5,6 m6,6

where vi,j ∈ IR1 and V can be equivalently expressed as


  n+1 n+1
V = v1 · · · vn+1 (18) X X
x̂ = −( vn+1,i [v1,i · · · vn,i ]T )/ 2
vn+1,i . (25)
n+1×1
where vi ∈ IR . i=p+1 i=p+1
Theorem 2 (Solution to TLS Problem [16]): Suppose IV. E XPERIMENTAL S ETUP
(13) is the SVD of [A; y]. If σn > σn+1 , then [Âŷ] = U ŜV T , The experiments reported in this paper employed the
and with corresponding TLS correction matrix Johns Hopkins University Remotely Operated Vehicle (JHU
T ROV) — a research testbed vehicle developed at Johns
k[4A; 4y]kF = σn+1 un+1 vn+1 (19) Hopkins University. It measures 1.37 m long, 0.85 m wide,
solves the TLS problem and and 0.61 m high. Fig. 1(a) depicts the JHU ROV and its
sensor suite [20]. Fig. 1(b) depicts the JHU ROV’s vehicle
1
x̂ = − vn+1,n+1 [vn+1,1 , · · · , vn+1,n ] T
(20) coordinate frame axes. The JHU ROV is equipped for full 6-
DOF state measurement. Vehicle roll, pitch, heading, angular
exists and is the unique solution to ŷ = Âx̂. body velocities, and XYZ acceleration are instrumented
1) Definition of the Under-determined of the Total Least with an IXSEA Phins North-seeking three-axis fiber-optic
Squares Problem inertial measurement unit (IMU) (IXSea SAS, Marly-le-
Definition 4 (Under-determined TLS Problem [16]): Roi, France). Depth is instrumented with a Paroscientific
Consider the TLS problem III-C, but σp > σp+1 = · · · = 10 meter range depth sensor Model Number 8CDP010-1
σn+1 such that p < n. Then any [Â; ŷ] = [A; y] − [4A; 4y] (Paroscientific, Inc., Redmond, Washington, USA). A 1,200
with kHz Doppler Sonar Model Number WHN1200/6K (Teledyne
RD Instruments, Poway, California, USA) provides XYZ
  body velocity measurements and, in combination with the
w  T T
 T T
[4A; 4y] = [A; y] w β k[w ; β ]k (21) gyrocompasses and depth sensor, enables Doppler navigation
β via the DVLNAV software program [19] to compute the
where kok is the Euclidean norm and where 6-DOF vehicle state.
  For these experiments, the vehicle was configured with six
w thrusters that produce force and torque to enable active control
∈ Image([vp+1 · · · vn+1 ]) (22)
β authority in all 6-DOF. The vehicle has two longitudinal
thrusters, two lateral thrusters, and two vertical thrusters.
solves the TLS problem provided β 6= 0. Assume that Each thruster is a custom DC brushless direct drive thruster

w Pn+1
= i=p+1 λvi , then that can produce a maximum of a 150 Newtons of thrust. We
β
employ a static thrust model to model the thrust produced by
n+1
X each thruster. Thrust is linearly proportional to motor torque
[4A; 4y] = σi2 λ2i k[wT ; β T ]k = σn+1
2
(23) at steady state and motor torque is linearly proportional to
i=p+1 motor current [26], [4]. We control motor current using cali-
brated current amplifiers. Although the steady-state thruster
is minimal. The corresponding TLS solution x̂ = −w/β model (i.e. thrust linearly proportional to torque) ignores the
satisfies thruster’s transient dynamics, [26], [4], in our experiments we
employed low-frequency to moderate-frequency sinusoidal
x̂ w
h i h i
[Â; b̂] = − β1 {[A; b]−[4A4y]} reference trajectories and open-loop thrust commands to
−1 β
w
h i   w
h i minimize the effects of thruster transient dynamics. In the
= −1 [A; b] 1− wT β T k[wT ; β T ]k2 . sequel we will see that the resulting identified plant models
β β β
(24) agree very will with the experimentally observed results, thus
2) Minimum Norm Solution to the Total Least Squares supporting the validity of this approach. We define the vehicle
Problem body-coordinate frame to be co-located and aligned with the
Theorem 3 (Minimum norm TLS solution [16]): IXSEA Phins coordinate frame.
Consider the TLS problem. Let (13) be the Singular XYZ acceleration was computed by differentiating Doppler
Value Decomposition of [A; y] and assume that sonar velocity data and angular acceleration data was obtained
σp > σp+1 = · · · = σn+1 with p ≤ n. If not all by differentiating the IXSEA Phins angular velocity data.
vn+1,i = 0 i = p + 1, · · · , n + 1, then the TLS solution x̂ is A median filter was applied to the velocity data from the
given by Doppler sonar velocity data to reject outliers resulting from

2965

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
(a) JHU ROV with major sensors identified. (b) JHU ROV with coordinate axes identified.

Fig. 1: Johns Hopkins University Remotely Operated Underwater Vehicle.


TABLE I: Six different plant models evaluated in this study.
experimental data of six-degrees-of-freedom vehicle motion.
The following six plant models were investigated:
ParameterParam- Description
Set Labeleters 1) FULL QD: This model employs a full 216 quadratic
FULL QD 241 Includes full 216 quadratic drag param- drag parameters, a full 21 mass parameters, and 4
eters, a full 21 mass parameters, and 4 buoyancy terms.
buoyancy terms 2) DIAG QD: This model employs a 6 diagonal quadratic
DIAG QD 16 Includes 6 diagonal quadratic drag pa- drag parameters, a full 21 mass parameters, and 4
rameters, a full 21 mass parameters, and
4 buoyancy terms buoyancy terms.
DIAG 226 Includes full 216 quadratic drag parame- 3) DIAG MASS: This model employs a full 216
MASS ters, a 6-parameter diagonal mass matrix, quadratic drag parameters, a 6-parameter diagonal mass
and 4 buoyancy terms matrix, and 4 buoyancy terms.
DIAG ALL 16 Includes 6 diagonal quadratic drag pa- 4) DIAG ALL: This model employs a 6 diagonal
rameters, a 6-parameter diagonal mass
quadratic drag parameters, a 6-parameter diagonal mass
matrix, and 4 buoyancy terms
FULL LD 61 Includes all 36 linear drag parameters, matrix, and 4 buoyancy terms.
all 21 mass parameters, and 4 buoyancy 5) FULL LD: This model employs a all 36 linear drag
terms parameters, all 21 mass parameters, and 4 buoyancy
CUSTOM 241 This model employs 215 quadratic drag terms.
parameters, and one linear drag param- 6) CUSTOM: This model employs 215 quadratic drag
eter for the roll DOF, a full 21 mass
parameters, and 4 buoyancy terms. parameters, and one linear drag parameter for the roll
DOF, a full 21 mass parameters, and 4 buoyancy terms.
our experimental test-tank setup. A median filter was applied In this model, the linear drag term DL4,4 is included
to the velocity data from the IXSEA Phins differentiated and the quadradic drag term DQ44,4 is omitted.
angular velocity data to compensate for outliers generated The six different plant models are summarized in Table I.
in the numerical differentiation of angular velocity data. A Note that we considered the “custom” model because pilot
median filter of length 9 and width 0.03 m/s was applied to experimental studies indicated that drag in the roll degree of
the velocity data from the Doppler sonar velocity data, where freedom was dominated by linear drag rather than quadratic
length defines the number of data points used in the median drag. We do not have a cogent theoretical justification for the
filter and width defines the maximum absolute difference use of a linear drag model over a quadratic drag model for
between the measured value of the sensor and the median the 4,4 drag term; we report it as an empirical observation.
value of the datum points used in the median filter. To reject The experimental methodology employed was as follows:
outliers in the estimated angular acceleration, a median filter First, all three of the JHU ROV’s translational DOFs were
of length 9 and width 0.3 deg/s2 was applied to differentiated commanded to follow a sinusoidal position and velocity
angular velocity data from the IXSEA Phins IMU. These reference trajectory under closed-loop control and all three of
median filter parameter were selected empirically to perform the rotational DOF were given an open-loop sinusoidal force
reasonable outlier rejection. trajectory. Table II reports the magnitude, frequency, and type
V. C OMPARATIVE E XPERIMENTAL I DENTIFICATION OF of reference trajectory sinusoidal profile for each DOF. Fig.
S IX 6-DOF P LANT M ODELS III reports the update rate and precision of the sensors used in
This section reports an analysis of the TLS parameter this identification. Second, the TLS identification algorithm
estimation algorithm; reported in [16], when applied to was employed to identify the plant model parameters from

2966

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
TABLE II: Magnitude and frequency of reference signal for
parameter identification, and whether the trajectory type is closed-
experimental trial, how well do the plant models predict
loop (CL) or open-loop (OL). the actual experimental plant performance from other vehicle
experimental trials that employ a variety of different reference
DOF Traj. Freq. 1 Mag. 1 Freq. 2 Mag. 2
trajectories? We employ the term cross-validation for this
v1 CL 0.17 rad/s 1m 0.0 rad/s 0m
v2 CL 0.2 rad/s 1m 0.0 rad/s 0m comparative analysis. Our experimental methodology for
v3 CL 0.35 rad/s 1m 0.0 rad/s 0m cross-validating a particular form of a plant model (i.e.
v4 OL 0.45 rad/s 40 Nm 0.35 rad/s 40 Nm decoupled, fully-coupled, etc.) is as follows: First, a set
v4 OL 0.5 rad/s 60 Nm 0.3 rad/s 70 Nm of plant model parameters is identified from from a single
v6 OL 0.35 rad/s 70 Nm 0.0 rad/s 0 Nm
vehicle experimental trial as described in Section V. Second,
TABLE III: Navigation sensors used for 6-DOF parameter identifi- several additional experimental trials are performed that
cation experiments. exhibit a variety of different vehicle forces and trajectories.
Third, numerical simulations compute the model’s predicted
Variable Sensor Precision Update Rate
state in response the actual commanded force data obtained
vt 1200 kHz RD σ = 3mm/s 1-10 Hz
Instruments single ping in the multiple additional experimental trials. Finally, we
Acoustic Doppler standard compute the error between the model performance and the
Current Profiler deviation experimental performance as the normalized mean absolute
θ & φ IXSEA Phins Heading Error: 0.02 10 Hz error between the actual experimental plant velocity and
Inertial × secant(latitude) deg the simulated plant velocity. The normalized error, for a
Navigation Roll and Pitch Error:
System 0.01 deg particular experiment, is computed by dividing the mean
vr IXSEA Phins 0.01 deg/s 10 Hz absolute error of each DOF by the mean absolute measured
Inertial velocity for that experiment. For example, the normalized
Navigation error for the v1 DOF, FULL QD parameter set is nei =
System mean(|vmodeli − vmeasuredi |)/mean(|vmeasuredi |).
experimental data. Third, the identified model parameters for Table V reports the normalized average error of the
each plant model and the logged force data from the identifica- results of multiple simulations using the model parameters
tion experiment were employed in a numerical simulation, the reported in Section V. These data show the parameter models
output of which is a simulated velocity vmodeli . The model identifying a fully parameterized drag model have smaller
simulation velocity for each DOF was then compared with the mean absolute error than parameter models that include a
corresponding experimentally observed plant velocity. The decoupled quadratic drag model. These data also show the
mean absolute error between the experimental velocity and CUSTOM parameter model performs as good or better than
the simulated velocity was computed to evaluate how closely every other parameter set except the v3 and v4 DOF for the
each plant model behavior agrees with the experimentally FULL LD parameter model and the v6 DOF for DIAG MASS
observed plant. The mean absolute error for each DOF i was parameter model.
calculated as ei = mean(|vmodeli − vi |). We conclude the following: First, the parameter sets,
The focus of this paper is discussion of the results of the including a fully coupled drag model yield the smallest overall
identification of the 6-DOF model on an underwater vehicle. error for TLS. Second, fully coupled drag models perform
The identified plant parameter numerical values are reported better than a decoupled drag model. Third, the TLS estimate
in [22]. The parameter models are compared by computing of the CUSTOM parameter model estimated the best set of
the mean absolute error between simulation velocities and model parameters.
the experimental velocities. The results are reported in Table VII. C ONCLUSIONS
IV. Table IV show that the TLS algorithm yields model This paper has reported the modeling and experimental
parameters that perform well and that parameter models identification of six different six degree-of-freedom coupled
including a coupled drag model (linear or quadratic) have a nonlinear second order plant models for low-speed maneu-
smaller mean absolute error than parameter models including vering of fully-actuated open-frame underwater vehicles. We
a decoupled quadratic drag model. Finally, these data show reported a comparative experimental evaluation of six different
that the CUSTOM and DIAG MASS parameter models have
the smallest mean absolute error overall. Fig. 2(a)—2(f) are TABLE IV: Comparison of absolute error between simulation
representative plots comparing the simulation velocity of the model velocity and experimentally observed velocity using the TLS
parameter estimates for the JHU ROV. The mean absolute error is
numerical simulation to the experimentally observed velocity.
in units of cm/s for the v1 , v2 , and v3 directions and deg/s for the
We conclude the following: First, TLS yields a set of v4 , v5 , and v6 directions.
parameters that accurately model the dynamics of this under-
Parameter v1 v2 v3 v4 v5 v6
water vehicle. Second, the CUSTOM and DIAG MASS plant
CUSTOM 4.37 4.18 3.09 1.86 1.88 2.90
models perform best overall than plant models employing a FULL QD 4.77 4.29 3.20 2.26 2.16 2.94
decoupled drag model. DIAG QD 4.70 6.48 3.66 2.94 2.58 4.14
VI. C ROSS -VALIDATION I DENTIFIED P LANT M ODELS DIAG ALL 4.76 6.31 3.68 3.06 2.56 4.09
DIAG MASS 4.74 4.10 3.04 1.94 1.95 2.77
This section addresses the following question: when FULL LD 4.61 4.26 2.79 1.81 1.74 3.18
using model parameters obtained from a single vehicle

2967

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
(a) Surge (v1 ) TLS Model and Experiment (b) Sway (v2 ) TLS Model and Experiment

(c) Heave (v3 ) TLS Model and Experiment (d) Roll (v4 ) TLS Model and Experiment

(e) Pitch (v5 ) TLS Model and Experiment (f) Yaw (v6 ) TLS Model and Experiment

Fig. 2: Comparison of the performance of the TLS identified plant model and the experimentally observed velocities. Velocities shown are
2(a) Surge (v1 ), 2(b) Sway (v2 ), 2(c) Heave (v3 ), 2(d) Roll (v4 ), 2(e) Pitch (v5 ), and 2(f) Yaw (v6 ). The plots show numerical simulation
velocity and actual velocity versus time (A) and the velocity error between simulated and experimentally observed velocities (B).

2968

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.
TABLE V: Comparison of the normalized average of the absolute
Directional Intelligent Navigator,” IEEE Robotics and Automation
error between simulation velocity and experimentally observed Magazine, vol. 2, no. 1, pp. 44–53, Mar. 1995.
velocity using the TLS parameter estimates for the JHU ROV on [7] G. Conte, S. M. Zanoli, D. Scaradozzi, and A. Conti, “Evaluation of
trajectories with significant motion in all 6-DOF. hydrodynamics parameters of a UUV: A preliminary study,” in Control,
Communications and Signal Processing, 2004. First International
Parameter v1 v2 v3 v4 v5 v6 Symposium on, 2004, pp. 545–548.
CUSTOM 0.4000 0.4399 0.3362 0.5015 0.3927 0.2046 [8] T. F. Elbert, Estimation and Control of Systems. New York, NY: Van
FULL QD 0.4179 0.4490 0.3548 0.5811 0.4259 0.2034 Nostrand Reinhold Company Inc., 1984.
DIAG QD 0.4220 0.6997 0.3704 0.7215 0.5164 0.2721 [9] J. Feldman, “DTNSDC revised standard submarine equations of motion,”
US Department of Defense, Tech. Rep., 1979.
DIAG ALL 0.4262 0.7343 0.4287 0.7595 0.5554 0.3274
[10] ——, “Model investigation of stability and control characteristics of
DIAG MASS 0.4494 0.4430 0.3371 0.4907 0.3992 0.1910 a preliminary design for the deep submergence rescue vessel (DSRV
FULL LD 0.4220 0.4933 0.2970 0.4618 0.3923 0.2178 scheme A),” Hydromechanics Laboratory, David Taylor Model Basin,
David Taylor Model Basin, Washington D.C., Tech. Rep. AD637884,
candidate plant models whose unknown plant parameters are June 1966.
estimated experimentally from data obtained in free-motion [11] T. I. Fossen, Marine Control Systems: Guidance, Navigation and
Control of Ships and Underwater Vehicles. Trondheim, Noraway:
vehicle trials. Marine Cybernetics, 2002.
We conclude the following: (1) Free-flight experiments [12] M. Gertier and G. Hagen., “Standard equations of motion for submarine
with coupled motion in all six degrees-of-freedom, described simulation,” US Department of Defense, Tech. Rep., 1967.
[13] A. Goodman, “Experimental techniques and methods of analysis used
in Section IV, provide data from which the TLS algorithm in submerged body research,” US Department of Defense, Tech. Rep.,
yield model parameters for which the plant models accurately 1960.
predict the 6-DOF dynamics of an underwater robotic vehicle. [14] A. J. Healey, F. A. Papoulias, and R. Cristi, “Design and experimental
verification of a model based compensator for rapid AUV depth control,”
(2) The CUSTOM and DIAG MASS plant model estimates in Proceedings of the Sixth International Symposium on Unmanned
performs best on the trajectory that the model parameters are Untethered Submersible Technology Technology, Ellicott City, MD,
obtained. (3) Fully coupled linear and quadratic drag plant USA, June 1989, pp. 458–474.
[15] Ø. Hegrenaes, O. Hallingstad, and B. Jalving, “Comparison of
models perform better than corresponding decoupled drag mathematical models for the HUGIN 4500 AUV based on experimental
models. (4) In cross-validation, the CUSTOM plant model data,” in International Symposium on Underwater Technology, UT 2007
performed best overall. - International Workshop on Scientific Use of Submarine Cables and
Related Technologies 2007, Tokyo, Japan, Apr. 2007, pp. 558–567.
Improved experimentally validated plant dynamical models [16] S. V. Huffel and J. Vandewalle, The Total Least Squares Problem:
for underwater vehicles may enable the development of model- Computational Aspects and Analysis. Philadelphia, PA, USA: Society
based controllers and may also enable high-fidelity numerical for Industrial and Applied Mathematics, 1991.
[17] F. Imlay, “The complete expressions for “added mass” of a rigid body
simulations to predict the performance of underwater vehicles moving in an ideal fluid,” US Department of Defense, Tech. Rep.,
under a great variety of disparate missions, vehicle control 1961.
laws, and operating conditions. [18] J. Kim, K. Kim, H. S. Choi, W. Seong, and K.-Y. Lee, “Estimation
of hydrodynamic coefficients for an AUV using nonlinear observers,”
ACKNOWLEDGMENTS IEEE Journal of Oceanic Engineering, vol. 27, no. 4, pp. 830–840,
The authors would like to thank Dr. D. R. Yoerger who gra- Oct. 2002.
[19] J. Kinsey and L. L. Whitcomb, “Preliminary field experience with
ciously provided the use of his 300-kHz SHARPS navigation the DVLNAV integrated navigation system for manned and unmanned
system. The authors gratefully acknowledge the support of the submersibles.” in Proceedings of the 1st IFAC Workshop on Guidance
National Science Foundation under Award #0812138. Stephen and Control of Underwater Vehicles, GCUV ’03, New Port, South
Wales, UK, April 2003, pp. 82–89.
C. Martin was supported by a National Defense Science [20] J. Kinsey, L. L. Whitcomb, and D. Smallwood, “A new hydrodynamics
and Engineering Graduate Fellowship, a Link Foundation test facility for UUV dynamics and control research.” in Proceedings
Doctoral Research Fellowship in Ocean Engineering and of IEEE/MTS Oceans’2003, San Diego, 2003.
[21] D. Marco and A. Healey, “Surge motion parameter identification for the
Instrumentation, and an Achievement Rewards for College NPS Phoenix AUV,” in Proceedings International Advanced Robotics
Scientists Foundation Scholarship. Program IARP 98, University of South Louisiana, February 1998, pp.
197–210.
R EFERENCES [22] S. Martin, “Advances in six-degree-of-freedom dynamics and control of
[1] J. Avila, K. Nishimoto, C. Sampaio, and J. Adamowski, “Experimental underwater vehicle,” Ph.D. dissertation, The Johns Hopkins University,
investigation of the hydrodynamic coefficients of a remotely operated Baltimore, MD, August 2008, https://dscl.lcsr.jhu.edu/wiki/images/d/
vehicle using a planar motion mechanism,” Journal of Offshore d5/Smartin thesis 10 24 08.pdf.
Mechanics and Arctic Engineering, vol. 134, pp. 021 601(1–6), 2012. [23] S. C. Martin and L. L. Whitcomb, “Preliminary results in experimental
[2] J. Avila and J. Adamowski, “Experimental evaluation of the hydro- identification of 3-DOF coupled dynamical plant for underwater
dynamic coefficients of a ROV through Morison’s equation,” Ocean vehicles,” in Proceedings of IEEE/MTS Oceans’2008, Quebec City,
Engineering, vol. 38, no. 17-18, pp. 2162–2170, September 2011. Canada, 2008.
[3] J. Avila, J. Adamowski, N. Maruyama, F. Takase, and M. Saito, [24] D. Smallwood and L. L. Whitcomb, “Preliminary identification of a
“Modeling and identification of an open-frame underwater vehicle: dynamical plant model for the jason 2 underwater robotic vehicle,” in
The yaw motion dynamics,” Journal of Intelligent & Robotic Systems, Proceedings of IEEE/MTS Oceans’2003, San Diego, CA, 2003, pp.
vol. 64, pp. 1–20, 2011. 688–695.
[4] R. Bachmayer, L. L. Whitcomb, and M. Grosenbaugh, “An accurate [25] D. A. Smallwood and L. L. Whitcomb, “Adaptive identification of dy-
finite-dimensional dynamical model for the unsteady dynamics of namically positioned underwater robotic vehicles,” IEEE Transactions
marine thrusters,” IEEE Journal of Oceanic Engineering, vol. 25, no. 1, on Control Systems Technology, vol. 11, no. 4, pp. 505–515, July 2003.
pp. 146–159, January 2000. [26] L. L. Whitcomb and D. Yoerger, “Development, comparison, and
[5] M. Caccia, G. Indiveri, and G. Veruggio, “Modeling and identification preliminary experimental validation of nonlinear dynamic thruster
of open-frame variable configuration underwater vehicles,” IEEE models,” Oceanic Engineering, IEEE Journal of, vol. 24, no. 4, pp.
Journal of Oceanic Engineering, vol. 25, no. 2, pp. 227–240, April 481–494, 1999.
2000.
[6] S. K. Choi, J. Yuh, and G. Y. Takashige, “Development of the Omni-

2969

Authorized licensed use limited to: Central Mechanical Engineering Research Institute. Downloaded on October 21,2022 at 07:29:21 UTC from IEEE Xplore. Restrictions apply.

You might also like