IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO.
4, JULY 2003
Adaptive Identification of Dynamically Positioned
Underwater Robotic Vehicles
David A. Smallwood, Student Member, IEEE, and Louis L. Whitcomb, Senior Member, IEEE
Abstract—This paper reports a stable online adaptive identifica-
tion technique for the identification of finite-dimensional dynam-
ical models of dynamically positioned underwater robotic vehi-
cles. Proofs for the identifier’s global stability, and for the input-to-
state stability of this class of plants are reported. A direct compar-
ison of the adaptive identification method to a conventional, off-
line, least-squares method is reported. Using experimental data ob-
tained with the Johns Hopkins University Remotely Operated un-
derwater robotic Vehicle (JHUROV), both methods are employed
to identify decoupled, single-degree-of-freedom dynamical plant
models. Performance of the resulting identified dynamical plant
models is quantitatively compared to the experimentally observed
motion of the actual vehicle.
Index Terms—Adaptive estimation, dynamics, least-squares
methods, robot dynamics, underwater vehicles.
I. INTRODUCTION
T HIS paper addresses the problem of experimental identifi-
cation of finite-dimensional nonlinear dynamical models
for open-frame remotely operated underwater vehicles. The de-
velopment of model-based control techniques for the dynamic
positioning of underwater robotic vehicles has been limited by
the lack of experimentally validated plant models.
Very few previously reported vehicle plant identification
studies, [4], [18], both report the experimentally determined
model parameters and quantitatively compare experimentally
observed vehicle dynamical response with that predicted by the
proposed analytical models.
This paper first reviews an approximate finite-dimensional
dynamical model for open-frame underwater vehicles. This
class of plants is then shown to possess input-to-state stability
(ISS). Using this property, we report a stable on-line adaptive
identification method for this class of plants. The performance
of the proposed adaptive identification technique is evaluated
in comparison to conventional least-squares parameter identi-
fication with experimental data obtained by the authors. The
quality of the resulting identified plant models is evaluated
by quantitatively comparing the model’s predicted dynamical
behavior to actual experimentally observed vehicle behavior.
To the best of our knowledge, the application of this general
identification approach to this specific class of plants, the ISS
stability property of this class of plants, and the experimental
Manuscript received September 17, 2001; revised August 8, 2002. Manu-
script received in final form February 13, 2003. Recommended by Associate
Editor D. S. Bayard. This work was supported by the Office of Naval Research
and the National Science Foundation under Grant CMS-0100783.
The authors are with The Johns Hopkins University, Baltimore, MD
21218-2686 USA.
Digital Object Identifier 10.1109/TCST.2003.813377
1063-6536/03$17.00 © 2003 IEEE
505
evaluation of this approach are new and have not been previ-
ously reported.
The paper is organized as follows. Section I-A reviews the
finite-dimensional modeling of actuated marine vehicles. Sec-
tion I-B briefly outlines standard techniques for model param-
eter identification. Section II reviews the application of conven-
tional least-squares parameter identification to this plant. Sec-
tion III reports a stable adaptive identifier for this plant, and ex-
amines its performance in numerical simulations. Section IV re-
ports our experimental setup. Section V quantitatively compares
the performance of the two approaches with actual experimental
data, and investigates several critical implementation issues.
A. Finite-Dimensional Dynamical Modeling of Marine
Vehicles
Exact analysis of a rigid-body underwater vehicle’s dy-
namics includes both the finite-dimensional dynamics of the
vehicle body itself and the infinite-dimensional dynamics of
the fluid surrounding the vehicle. The former is a finite-dimen-
sional dynamical system represented by an ordinary differential
equation (ODE), but the latter is a continuous (infinite-dimen-
sional) dynamical system represented by the incompressible
Navier–Stokes equation, a partial differential equation (PDE).
The overwhelming computational complexity of PDE
dynamical models has motivated the widespread use (by
naval architects and others) of finite-dimensional approximate
models for marine vehicles, sometimes termed “lumped pa-
rameter models.” The expedient of experimentally determining
approximate finite-dimensional models for complex fluid
phenomenon is widely employed in naval architecture and the
field of underwater robotic vehicles [6]–[9], [11]–[13].
Most reported finite-dimensional plant models for holonomic
(fully actuated) underwater vehicles take the general form
(1)
where is a vector of the vehicle position and
orientation in world inertial coordinates; and
are, respectively, vectors of the vehicle velocity
and acceleration in vehicle body coordinates; is a
vector of control forces and moments generated by thrusters and
control surfaces in body coordinates; is a mass
matrix representing both rigid-body mass and velocity-depen-
dent “added mass;” is a vector of vehicle drag
and coriolis forces; and is a vector of vehicle buoy-
ancy forces. Note that the velocity in inertial (world) coordinates
, is related to the velocity in body coordinates by a linear
transformation of the form [9].
506 IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO. 4, JULY 2003
At present, there is no uniform consensus within the research
community on the exact analytical form of the terms comprising
, , and , nor for the force
vector .
Although not theoretically justified, in this report we adopt
the common practice of further approximating the 6-degree of
freedom (DOF) equations by neglecting off-diagonal entries and
coupling terms, tether dynamics, as well as assuming a constant
added mass [4], [18]. These simplifications are empirically jus-
tified by the slow velocities and modest attitude changes typical
of this class of underwater vehicles. The resulting decoupled
1-DOF dynamical equations take the form
(2)
where, for each DOF , is the net control force, is the
effective mass, and represent the hy-
drodynamic drag, and is the buoyancy. Using lumped param-
eters, (2) can be written as
(3)
or written in vector form as
(4)
where is the vector of lumped plant
parameters and is a nonlinear
vector of state and the control input . A nomenclature is
given in Table I. The lumped parameters are defined in Table II.
B. Experimental Identification of Plant Parameters
Several methods of determining the parameters of dy-
namical equations of motion for underwater vehicles have
been presented in the literature. These include the method of
least-squares [4], [5], [10] and the use of extended Kalman
filters (EKFs) [1], [5]. A method based on numerical mini-
mization of the error between the trajectories of the vehicle
and several models during free decay, in 1-DOF, was presented
in [18]. In special idealized cases—e.g., symmetric vehicles
in inviscid flow [17]—plant parameters can be determined
analytically, but this approach does not extend to general
vehicle shapes in viscous fluid.
This paper will present two identification methods and
compare the results obtained with data from free running exper-
imental trials. The first is a conventional offline least-squares
scalar identification method. The second is a novel application
to this class of plants of a well known online identification
method utilizing a stable adaptive identifier [19], [20].
II. OFFLINE LEAST-SQUARES PARAMETER IDENTIFICATION
Parameters for plants of the form (4) can be identified ex-
perimentally using the standard least-squares technique. This
method requires values for , , and for each DOF
. purpose of this identifier is to identify dynamical plant model
TABLE I
NOMENCLATURE
TABLE II
DEFINITION OF LUMPED PARAMETERS
Defining and
, where represent
the sample times, then from (4) we have
(5)
Thus if is full rank, then the least-squares solution for
the unknown parameter vector is given by the standard
Moore–Penrose pseudoinverse
(6)
This method is computationally simple but requires that the
velocity and acceleration of the vehicle are instrumented in ad-
dition to the applied force. Instrumentation of acceleration is
not standard on most underwater vehicles. Further, calculation
of acceleration via direct numerical differentiation, an acausal
operation, is easily corrupted by sensor noise.
For completeness, the authors would like to point out that
there are valid alternative approaches to the least-squares
method in the literature [2], [15], [22], that eliminate the need
for acceleration measurement. Essentially these approaches
propose low-pass filtering (5) to eliminate the need for acceler-
ation measurement.
III. ONLINE ADAPTIVE PARAMETER IDENTIFICATION
This section presents a stable adaptive technique for param-
eter identification as an alternative to the offline least-squares
technique. This approach can be implemented online, and
does not require instrumentation of acceleration. The general
approach to plant parameter identification employed herein is
well known [19], [20], the application to this class of plants and
experimental evaluation reported herein is new. This adaptive
approach to plant parameter identification utilizes a) the actual
plant with full state access to state vector, , (4), and b)
an “identification” plant, with state vector , (7), and c) a
parameter update law, (11). The goal of this approach is to
ensure that all signals remain bounded and that, under sufficient
conditions of persistent excitation, the plant parameter esti-
mates converge to the true plant parameter values. Note that the
SMALLWOOD AND WHITCOMB: ADAPTIVE IDENTIFICATION OF DYNAMICALLY POSITIONED UNDERWATER ROBOTIC VEHICLES 507

parameter values, not the state of the plant. The application 4) In consequence, , , and are bounded.
of this general identification approach to this specific class of 5) From (12) and that and are bounded, we can
plants, the ISS stability property of this class of plants, and the then conclude that . Then using a corollary to
experimental evaluation of this approach are new and have not Barbalat’s lemma from [19], since is bounded, we
been previously reported. can conclude .
Given the scalar plant (4), the chosen adaptive estimator is 6) Given that , and is bounded, then
from (11) we conclude .
(7) In summary, all signals remain bounded, the identifier
state error, , converges asymptotically to zero, and the
where and
time-derivative of the parameter error, , converges
. was used in the adaptive iden-
asymptotically to zero. Although the parameter estimation
tifier instead of because of the typical noise associated with
error remains bounded, absent additional arguments we cannot
the measurement of actual plant velocity, . is an esti-
conclude that the parameter error converges asymptotically
mate of the unknown parameter vector , and is a scalar
to zero. It is well known, however, that the parameter error
gain. The identifier error coordinates are defined as
for this general type of system will converge to zero given a
sufficiently rich input signal [19], [20], [21]. Numerical
simulation of the full system, reported in the next section,
indicates that asymptotic convergence of the plant parameters
occurs for most time varying inputs.
(8)
A. Performance of Adaptive Identifier on a Simulated
Consider the Lyapunov function candidate of the form Analytical Plant
How does this scalar adaptive identifier perform on an “ideal”
(9) simulated analytical plant? Consider the scalar single DOF an-
alytical plant model of the form
where , for i , with , whose time
derivative is (13)

where , , subject
to a propulsion thrust profile of
[N].
(10)
This model represents a vehicle of 250 kg total mass (140 kg
If we choose the parameter update law to be dry mass, 110 kg added mass) that is experiencing both inertial
forces and quadratic drag forces, and is undergoing a 1/4 Hz sin
(11) wave propulsion thrust profile of peak magnitude 100 N. The
velocity error feedback gain was set as , and the
then (10) becomes adaptation gains were set to and .
A fourth order Runge–Kutta ODE integrator was used to con-
duct this simulation and identification. The performance of the
adaptive identifier on the analytical plant is shown in Fig. 1. As
predicted by theory, , , and are bounded and
goes to zero. In addition we see that goes to zero,
(12)
and the estimated parameters converge to the actual simulated
Examining the term we plant values, .
see that when we have which The adaptive identifier offers two advantages in comparison
results in . Likewise to conventional off-line least-squares techniques. First, it does
if then we have , resulting in not require acceleration instrumentation. Second, the adaptive
. Since , , and identifier can be implemented on-line and suffers from no rank
are negative scalar constants, it is clear that (12) is negative deficiency singularity problems.
semidefinite.
We conclude the following points. IV. EXPERIMENTAL SETUP
1) From (9) and (12), we conclude that and The Johns Hopkins University Remotely Operated under-
are bounded. water robotic Vehicle (JHUROV), Fig. 2, is a tethered, remotely
2) Given that is bounded and is a constant, then operated underwater robot. The JHUROV is powered by an
is bounded as well. isolated 10 kW dc power supply. The mass of the vehicle is
3) Given that (4) is bounded-input bounded-state (BIBS) 140 kg and it’s dimensions are 1.5 m long 1 m wide 0.6 m
stable (see Appendix A), and that the input is high. The XYZ position and heading of the vehicle are actively
bounded, then and are bounded. controlled and the vehicle is passively stable in roll and pitch.
508 IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO. 4, JULY 2003

(a)

(b)
Fig. 1. Numerical simulation showing the performance of the adaptive identifier, (7) and (11), on an “ideal” simulated plant, (13). (a) The simulated plant velocity
1 () 1 () 1 ()
(top) and the adaptive identifier estimated plant velocity (bottom). (b) Identifier velocity error v t versus time (top); Parameter error  t and  t versus
()
time (middle); and Lyapunov function W t versus time (bottom).
Actuation is provided by five dc brushless electric thrusters. A
complete description of the JHUROV is reported in [23].
The ROV is equipped for full 6-DOF position measurement.
For this set of experiments, XYZ position was measured
using a 300 kHz Sonic High Accuracy Ranging and Posi-
tioning System (SHARPS) time-of-flight acoustic hardwired
transponder system. Depth was instrumented via a Foxboro/ICT
analog pressure transducer, sampled at 20 Hz. A KVH Azimuth
Digital Gyro Compass (ADGC) measured the vehicle’s roll
and pitch. Heading was measured using a prototype Litton
LN200 IMU. Thruster specifications are given in Table III. The
vehicle’s entire sensor suite is listed in Table 4.
While not directly instrumented, velocity was obtained by
directly numerically differentiating the position and running it
SMALLWOOD AND WHITCOMB: ADAPTIVE IDENTIFICATION OF DYNAMICALLY POSITIONED UNDERWATER ROBOTIC VEHICLES
Fig. 2. Physical layout of the JHUROV.
TABLE III
THRUSTER SPECIFICATIONS
TABLE IV
JHUROV EXPERIMENTAL INSTRUMENTATION
through a low-pass filter. Acceleration was obtained by directly
numerically differentiating the velocity. At the time we per-
formed the experiments reported herein, our vehicle did not have
on-board accelerometers. However, in subsequent experiments,
we have found that the vehicle acceleration measurements ob-
tained as reported in this paper (by differentiating the highly ac-
curate SHARPS LBL positions) compare favorably with accel-
eration measurements obtained directly with the accelerometers
in an IXSEA OCTANS north-seeking fiber-optic gyrocompass
unit mounted on the vehicle.
For these experiments, the vehicle was actuated by two
longitudinal thrusters and one vertical thruster. There were
no thrusters arranged to provide pure lateral movement. Free
running experiments were conducted in the , , and
DOF. Multiple trials in each DOF, one DOF at a time, were
conducted with sinusoidal thrust profiles of varying magnitude
and frequency. The thrust profiles employed in the single DOF
experiments were chosen to minimize the effects of cross
coupling in the dynamics. General maneuvers involving cross
coupled dynamics were not performed.
A static thrust control law was used for these experiments to
model the thrust produced by each thruster; i.e., the thrust pro-
duced is assumed to be proportional to the current commanded,
where is the motor torque constant with units of
[N-m/A] and is the current commanded in amperes. Current
was controlled using current mode amplifiers. The authors are
509
aware of more advanced thruster models [3], [14] that are more
precise, however they require a dynamical thruster characteri-
zation not available at the time of the experiments.
The experiments reported in this paper were conducted using
the JHUROV at the United States Naval Academy’s Hydrome-
chanics Laboratory, Annapolis, MD. The work was done in col-
laboration with Dr. D. Stilwell of the Virginia Polytechnic In-
stitute and State University, who deployed his inertial measure-
ment unit (IMU) on the vehicle during these experiments to ex-
amine the issue of IMU calibration.
V. EXPERIMENTAL RESULTS
This section reports a comparative experimental evaluation of
the following two techniques for determining plant-model pa-
rameters from free-motion experimental data: 1) conventional
least-squares parameter identification and 2) the proposed adap-
tive parameter identification technique. For each 1) and 2), we
report the following steps: First, we state the procedure followed
to execute the proposed identification technique. Second, we re-
port the resulting identified plant parameters. Third, we evaluate
the “quality” of the identified plant parameter values by quan-
titatively comparing experimentally observed plant dynamical
performance to that predicted by the identified plant model.
In addition, this section also reports the effect of different
thrust profiles on the dynamic performance of the identified dy-
namical plant models.
A. Least-Squares Identified Parameters
We employed the conventional least-squares parameter iden-
tification outlined in Section II, to experimentally determine
plant parameter values for the JHUROV. First, dynamic vehicle
trials were conducted with motion in the , , and DOF.
For each experiment, estimates for the plant parameters were de-
termined using the least-squares method. Next numerical sim-
ulations were run on the identified parameters. The simulation
was run using the actual logged force profile from the exper-
iment. The output of the simulation was a velocity, ,
which was compared to the actual logged experimental plant
velocity, . The error, for each DOF , was calculated as
mean . Plots of , , and are
shown in Fig. 3. The parameters that performed best (i.e., had
the lowest “error,” ) of all the experiments in numerical simu-
lations are listed in Table V. Fig. 3 shows that the simulated plant
model performance agrees closely with actual experimental data
when the model employs the least-squares identified parameters
listed in Table V.
B. Adaptively Identified Parameters
We employed the stable adaptive scalar identification method
outlined in Section III to experimentally determine plant pa-
rameters for the JHUROV. Using the same experimental data
as used in the least-squares method in Section V-A, the scalar
adaptive identifier was run on each experimental data set using
a fourth-order Runge–Kutta ODE integrator. The output of this
process is an estimated plant velocity, , as well as estimates
for the unknown plant parameters, . A plot of the output of
the adaptive identifier for one experimental trial in the DOF
510 IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO. 4, JULY 2003

(a)

(b)
Fig. 3. x DOF: Comparison of the simulated performance of (a) the least-squares and (b) adaptively identified plant models, (4), and actual experimental data.
The plots show (top) model-simulated and actual velocity versus time and (bottom) the velocity error between simulated and experimentally observed velocities.
Parameter values used in the simulation are listed in Tables V and VI.

can be seen in Fig. 4. Next a simulation was run using the adap- plant parameters and the “true” plant parameters. These plots
tively identified parameters. The output was a velocity, . correspond to the same adaptive identification seen in Fig. 4. In
The logged plant velocity, , was then compared to . As plotting the plant parameter error, the parameters identified by
done in Section V-A, the error for each DOF , is calculated as the least-squares method were considered to be the “true” plant
mean . The best identified parameters were parameters. The data indicates that the simulated performance
those with the lowest “error,” , and are listed in Table VI. Fig. 3 of the plant model employing the adaptively identified parame-
contains plots of and for the DOF. Fig. 5 shows ters, listed in Table VI, equals or exceeds that of the plant model
the velocity error ( ), and the error between the estimated employing the least-squares identified parameters.
SMALLWOOD AND WHITCOMB: ADAPTIVE IDENTIFICATION OF DYNAMICALLY POSITIONED UNDERWATER ROBOTIC VEHICLES 511

TABLE V
LEAST-SQUARES LUMPED PARAMETERS

Fig. 4. Performance of adaptive identifier on experimental data for a trial in the x DOF. Plot shows identifier velocity v^(t) and the plant parameter estimates
^ (t), ^ (t), ^ (t), and ^ (t). Adaptive identifier is attempting to identify a model of the form reported for the least-squares technique in Table V. 8 are the
least-squares identified parameters.
Note that while care was taken to be methodical and precise
in examining the data and using the adaptive identifier, choosing
of the velocity error feedback gain, , and the adaptation gains
remains a trial and error process. At present, there is no re-
ported analytical technique for “optimally” tuning these adap-
tation gains.
In this section, the adaptive identifier was run on the data
using a fourth-order Runge–Kutta ODE integrator and logged
experimental plant state values. For online implementation,
however, fourth-order Runge–Kutta ODE is not feasible. The
online implementation of the adaptive identifier was explored
using Euler integration and was found to perform comparably
to the Runge–Kutta ODE integrator implementation.
C. Effect of Different Thrust Profiles
When model parameters are obtained from a particular set of
experimental data, it is reasonable to expect that the dynamic
performance of the resulting identified model will agree, to
some degree, to that observed in the original experimental data.
An important question arising in experimental model identifi-
cation is whether or not the a plant model identified from one
set of experimental data accurately predicts the experimental
dynamic performance observed under different experimental
conditions. For example, in the case of linear plant models for
nonlinear systems it is well known that identified linear plant
models are typically only valid in a neighborhood around the
operating point (an equilibrium point) at which the model was
identified. In this paper we have adopted a nonlinear plant
model, (3), which should be valid for a wide range of operating
conditions.
This section examines how an identified plant model’s per-
formance compares to experimentally observed plant behavior
under thrust profiles differing from those employed to obtain the
model parameters. We performed this comparison as follows:
First, we identified a set of plant parameters , from experi-
mental data, for example “data set A,” corresponding to a partic-
ular thrust profile . The response of the identified model was
then simulated both using original thrust profile , and also
a substantially different experimental thrust profile, , from
“data set B.” In each case, the simulated model response was
compared to the actual experimentally observed response. This
was done for each thrust profile.
512 IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO. 4, JULY 2003

1 ()
Fig. 5. x DOF: Adaptive identifier performance on actual experimental data using Runge–Kutta ODE integration. The plot shows identifier velocity error v t
1 () 1 () 1 () 1 ()
versus time (top graph); Parameter error  t ,  t ,  t , and  t versus time (four middle graphs); and Lyapunov function W t versus time ()
(bottom graph). The parameter errors and Lyapunov function value are computed with respect to the least-squares identified plant parameters in Table V.

TABLE VI
ADAPTIVE ID LUMPED PARAMETERS

We observed that a model employing parameters identified
using high-magnitude thrust profiles performed uniformly
well—the single model accurately predicts experimentally
observed response at all thrust levels. In contrast, a model
employing parameters identified using low-magnitude thrust
profiles performed poorly—the single model predicts experi-
mentally observed response only for thrust profiles similar to
its model-identification data set.
In summary, both identification methods (least-squares and
adaptive) were observed to perform best when employing data
from trials containing thrust profiles of higher magnitude and
lower frequency. Under these conditions, the vehicle experi-
enced the largest variation in position, velocity, and accelera-
tion. We surmise that this improves the signal-to-noise ratio of
the experimental data and, in consequence, provides for more
accurate model parameter identification.
D. Discussion of Results
This section first discusses the physical interpretation of the
results and, second, articulates a comparison of the least-squares
and adaptive identification techniques.
1) Physical Interpretation of Parameters: Do the identified
plant parameters seem reasonable? An examination of the
parameters obtained by the two (very different) identification
techniques reveals that the two methods result in nearly iden-
tical plant parameter values. Moreover, the identified parameter
values are consistent with our physical understanding of the
vehicle’s design.
The JHUROV, Fig. 2, has a total dry mass of 140 kg. In the
(surge) direction, in which the vehicle presents a minimal cross-
section, it exhibits an added mass of about 150 kg. As expected,
in the (heave) direction, in which the vehicle presents a large
flat 1.7 m cross section, the vehicle exhibits an added mass
about ten to 20 times larger than in the (surge) direction.
The identified parameters indicate that drag is highest in the
DOF. This is physically consistent with the ROV’s layout, as
seen in Fig. 2. Once again this is due to the fact that the vehicle
is more or less a flat plate moving in the DOF, while the
vehicle’s layout has a smaller cross-sectional area in the and
DOF.
The buoyancy term represents a combination of the actual
buoyant force and any possible thruster zero-calibration offset
error. Only in the DOF did we identify a significant amount
SMALLWOOD AND WHITCOMB: ADAPTIVE IDENTIFICATION OF DYNAMICALLY POSITIONED UNDERWATER ROBOTIC VEHICLES
TABLE VII
DIRECT COMPARISON OF ADAPTIVE AND LEAST-SQUARES IDENTIFIED LUMPED PARAMETERS
of buoyant force. This is consistent with our ballasting of the
vehicle, the ROV was deliberately trimmed to be slightly nega-
tively buoyant. The values listed for the buoyancy term in the
and DOF are much smaller in magnitude than the value for
the DOF. We surmise that these small fictitious “buoyancy”
forces are the result of a small zero-offset in the thruster current
amplifier calibration.
2) Comparison of Least-Squares and Adaptive Identification
Techniques: How does the performance of the two identifica-
tion (adaptive versus least-squares) methods compare? Com-
paring the “error,” as defined in Sections V-A and V-B, of
each method in predicting the dynamical response of the ROV
listed in Tables V and VI, it is clear that the plant models iden-
tified by the adaptive method do a better job predicting the
ROV dynamical motion. Note that the models identified by each
method do not necessarily contain the same lumped parameters.
How does the performance of the two identification methods
compare when they are both used to identify plant models
consisting of the same lumped parameters? Table VII directly
compares the dynamic response of plant models consisting of
the same lumped parameters (with parameters identified by
the adaptive and least-squares approaches, respectively) to the
actual observed experimental dynamic response. The “error,”
, is the mean absolute difference between the identified
model simulated response and the actual logged experimental
response. The data shows the adaptively identified parameters
to yield a plant model whose behavior corresponds more
exactly to the experimental data than do the least-squares iden-
tified parameters. Although there is no theoretical algorithmic
justification for the adaptive method’s apparently superior
performance, we surmise that the difference is principally
due to their respective instrumentation requirements. In our
experiments, vehicle position was measured directly with a
300 kHz long-baseline tracking system at 10 Hz to an accuracy
of several millimeters. The vehicle velocity and acceleration
signals were obtained by direct differentiation of the position
signals. This is theoretically objectionable, but in practice it
is a necessity. Given that the least-squares method requires
both velocity and acceleration signals, while the adaptive
technique only requires velocity signals, we surmise that the
apparently poorer performance of the least-squares approach is
a consequence of numerical differentiation “noise.”
In short, both methods perform well. It should be noted here
that both methods identified simple, nonlinear dynamical plant
models that were validated over a range of operating conditions,
something that would not be expected with the use of a lin-
earized plant model. The adaptive technique performs slightly
better and has the advantage of not requiring acceleration in-
strumentation, whereas the least-square method does require
513
TABLE VIII
MAP OF NOMENCLATURE BETWEEN THIS PAPER AND [28] AND [29]
acceleration instrumentation. One disadvantage of the adaptive
method is that it possesses several gains that must be empirically
tuned.
VI. CONCLUSION
This paper has presented to the best of our knowledge, the first
experimental evaluation of a provably stable, on-line adaptive
identification technique for an underwater robotic vehicle and
a direct comparison to a conventional, off-line, least-squares
identification technique. Decoupled, single DOF dynamical
plant models identified by both techniques were presented and
evaluated. The adaptive method was shown to produce superior
dynamical plant models to the least-squares technique. The effect
of different thrust profiles on model parameter identification
was also investigated, confirming that both the least-squares and
the adaptive identification techniques performed better when
used on thrust profiles of higher magnitude. When nonlinear
plant models of the form (3) are properly identified, they exhibit
both steady-state and dynamic response that closely agrees
with the experimentally observed response over a wide range of
operating conditions. Such performance would not be expected
of a linearized plant model.
As mentioned in Section I, the development of model-based
control techniques for the dynamic positioning of underwater
robotic vehicles has been limited by the lack of experimentally
validated plant models. This paper has reported experimentally
validated plant models for the JHUROV. These reported models
served as the foundation for the design of a family of model-based
controllers for the JHUROV. Extensive experimental and nu-
merical simulation of the performance of these model-based
controllers has been reported by the authors in [24]–[26]. The
accurate trajectory tracking performance of the model-based
controllers reported in these papers serves as an additional tes-
tament to the validity of their ability to represent the dynamical
motion of this class of underwater robotic vehicles.
APPENDIX
The decoupled dynamical system (3) has the property that
bounded control inputs gives rise to bounded state. Note
514 IEEE TRANSACTIONS ON CONSTROL SYSTEMS TECHNOLOGY, VOL. 11, NO. 4, JULY 2003

that the state of the system in this formulation is the plant ve- then
locity . This property can be established directly, by a Lya-
punov-like analysis, or indirectly, by employing recent general
results on ISS. The latter approach is reported here. First, in Ap-
pendix A, the vehicle plant with zero-buoyancy is shown to be and
ISS. Second, in Appendix B, the general vehicle plant with ar-
bitrary buoyancy is shown to be BIBS but not ISS.
which results in
A. Input to State Boundedness: Case With Zero Buoyancy
Recently reported advances on ISS provide a well-defined (19)
general definition of input-output stability for general nonlinear
systems, [27] as well as a family of general Lyapunov-like nec-
Applying (19)
essary and sufficient conditions for ISS systems, e.g., [29]. A
map of the nomenclature used here and that used by Sontag [28],
[29], is provided in Table VIII for clarification.
First consider the case where in which the plant buoyancy
term is zero, i.e., . The decoupled dynamical system (3)
then takes on the form (20)

(14) Thus we conclude the system (14) is ISS.

where , and as defined in (2) and Table II, B. Input to State Boundedness: Case With Nonzero Buoyancy
. In this case, the function The nonzero buoyancy system (3), of the form (16) is not ISS
stable because, among other reasons, it no longer satisfies the
(15) equilibrium condition . The system (3) can, how-
ever, be shown to possess the weaker property of BIBS stability
is an ISS Lyapunov function in the sense reported in [28] and in the usual sense, as follows: Consider the system (3) rewritten
[29]. This is established by showing that the system (14) satis- with a new control input, , defined as
fies the following conditions.
1) The system (14) has the form (21)

(16) The resulting system takes the form

where is continuously differentiable and (22)

.
2) The function is a smooth map from to . identical to the original no-buoyancy case (14) which, from Ap-
3) The functions and pendix-A, is ISS and, in consequence, BIBS. Given that (22) is
are, by inspection, radially unbounded compar- BIBS, and (21) ensures that, since is finite by assumption,
ison functions, , [16], and trivially satisfy then if is finite, is also finite. In summary, finite
the relation implies finite , and finite implies finite , thus (3)
is BIBS.
(17)
ACKNOWLEDGMENT
4) The functions and The authors gratefully acknowledge our ongoing collabora-
are both radially unbounded com- tion with Prof. D. Stilwell of the Virginia Polytechnic Institute
parison functions, , satisfying the relation and State University, who tested his inertial navigation system
that aboard the JHUROV during these experiments. The authors are
indebted to the USNA Hydromechanics Laboratory for their
(18) world-class testing facilities, and to Dr. D. Yoerger of the Woods
Hole Oceanographic Institution, who graciously provided use of
for any and any such that his 300 kHZ SHARPS acoustic transponder system.
. That is, if
REFERENCES
[1] A. Alessandri, M. Caccia, G. Indiveri, and G. Veruggio, “Application
of LS and EKF techniques to the identification of underwater vehicles,”
in Proc. IEEE Int. Conf. Control Applications, Trieste, Italy, 1998, pp.
1084–1088.
SMALLWOOD AND WHITCOMB: ADAPTIVE IDENTIFICATION OF DYNAMICALLY POSITIONED UNDERWATER ROBOTIC VEHICLES
[2] M. Amestegui, R. Ortega, and J. Ibarra, “Adaptive computed torque con-
trol for rigid link manipulators,” in Proc. 25th Conf. Decision Control,
Athens, Greece, 1986, pp. 68–73.
[3] R. Bachmayer, L. L. Whitcomb, and M. Grosenbaugh, “An accurate
four-quadrant nonlinear dynamical model for marine thrusters: Theory
and experimental validation,” IEEE J. Oceanic Eng., vol. 25, pp.
146–159, Jan. 2000.
[4] M. Caccia, G. Indiveri, and G. Veruggio, “Modeling and identification
of open-frame variable configuration underwater vehicles,” IEEE J.
Oceanic Eng., vol. 25, pp. 227–240, Apr. 2000.
[5] S. Eiani-Cherif, G. Lebret, and M. Perrier, “Identification and control
of a submarine vehicle,” in Proc. 5th IFAC Symp. Robot Contr., Nantes,
France, 1997, pp. 307–312.
[6] J. P. Feldman, “State-of-the-art for predicting the hydrodynamic char-
acteristics of submarines,” in Proc. Symp. Control Theory Naval Appli-
cations, Monterey, CA, June 1975, Defense Tech. Inform. Center Doc.
A071 804, pp. 87–127.
[7] , “DTNSRDC revised standard submarine equations of motion,”
David Taylor Naval Ship Research and Development Center, Tech. Rep.,
Defense Tech. Inform. Center Doc. A071 804, June 1979.
[8] T. I. Fossen, Guidance and Control of Ocean Vehicles. New York:
Wiley, 1994.
[9] M. Gertler and G. Hagen, “Standard equations of motion for subma-
rine simulation,” David Taylor Naval Ship Research and Development
Center, Tech. Rep., Defense Tech. Inform. Center Doc. 653 861, June
1967.
[10] K. Goheen and E. Jefferys, “The application of alternative modeling
techniques to ROV dynamics,” in Proc. IEEE Int. Conf. Robot. Automat.,
May 1990, pp. 1302–1309.
[11] A. Goodman, “Experimental techniques and methods of analysis used in
submerged body research,” in Proc. 3rd Symp. Naval Hydrodynamics.
Washington, DC, 1960, pp. 379–449.
[12] A. J. Healey, “Model-based maneuvering controls for autonomous un-
derwater vehicles,” ASME J. Dyn. Syst., Meas. Contr., vol. 114, no. 4,
pp. 614–622, Dec. 1992.
[13] A. J. Healey and D. Lienard, “Multivariable sliding mode control for au-
tonomous diving and steering of unmanned underwater vehicles,” IEEE
J. Oceanic Eng., vol. 18, pp. 327–339, July 1993.
[14] A. J. Healey, S. M. Rock, S. Cody, D. Miles, and J. P. Brown, “Toward an
improved understanding of thruster dyamics for underwater vehicles,”
IEEE J. Oceanic Eng., vol. 20, pp. 354–361, Oct. 1995.
[15] P. Hsu, M. Bodson, S. Sastry, and B. Paden, “Adaptive indentification
and control for manipulators without using joint accelerations,” in
Proc. IEEE Int. Conf. Robotics Automation, Raleigh, NC, 1987, pp.
1210–1215.
[16] A. Isidori, Nonlinear Control Systems II. London, U.K.:
Springer-Verlag, 1999.
[17] N. Leonard, “Stability of a bottom-heavy underwater vehicle,” Auto-
matica, vol. 33, no. 3, pp. 331–346, March 1997.
[18] A. T. Morrison, III and D. R. Yoerger, “Determination of the hydrody-
namic parameters of an underwater vehicle during small scale, nonuni-
form, 1-dimensional translation,” in Proc. IEEE/MTS OCEANS’93, Oct.
1993, pp. 277–282.
[19] K. Narendra and A. Annaswamy, Stable Adaptive Systems. New York:
Prentice-Hall, 1988.
[20] S. Sastry and M. Bodson, Adaptive Control: Stability, Convergence, and
Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989.
515
[21] J. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs,
NJ: Prentice-Hall, 1991.
[22] J.-J. E. Slotine and W. Li, “Composite adaptive control of robot manip-
ulators,” Automatica, vol. 25, no. 4, pp. 509–519, July 1989.
[23] D. Smallwood, R. Bachmayer, and L. Whitcomb, “A new remotely op-
erated underwater vehicle for dynamics and control research,” in Proc.
11th Int. Symp. Unmanned Untethered Submersible Technol., Durham,
NH, 1999, pp. 370–377.
[24] D. A. Smallwood and L. L. Whitcomb, “Preliminary experiments in the
adaptive identification of dynamically positioned underwater robotic ve-
hicles,” in Proc. 2001 IEEE/RSJ Int. Conf. Intelligent Robots Systems,
Maui, HI, 2001, pp. 1803–1810.
[25] , “Toward model based dynamic positioning of underwater robotic
vehicles,” in Proc. IEEE OCEANS’ 2001, Honolulu, HI, 2001, pp.
1106–1114.
[26] , “The effect of model accuracy and thruster saturation on tracking
performance of model based controllers for underwater robotic vehi-
cles: Experimental results,” in Proc. IEEE Int. Conf. Robot. Automation,
Washington, DC, 2002, pp. 1081–1086.
[27] E. D. Sontag, “Smooth stabilization implies coprime factorization,”
IEEE Trans. Automat. Contr., vol. 34, pp. 435–443, Apr. 1989.
[28] E. D. Sontag and Y. Wang, “On characterizations of the input-to-state
stability property,” IEEE Syst. Contr. Lett., vol. 24, no. 5, pp. 351–359,
Apr. 1995.
[29] , “New characterizations of input to state stability,” IEEE Trans.
Automat. Contr., vol. 41, pp. 1283–1294, Sept. 1996.
David A. Smallwood received the B.S. degree from the University of Miami,
Coral Gables, FL, in 1997 and the M.S. degree from The Johns Hopkins Univer-
sity, Baltimore, MD, in 1999, where he is currently pursuing the Ph.D. degree
in the Dynamical Systems and Controls Laboratory, Department of Mechanical
Engineering.
His Ph.D. research involves the dynamics and control of underwater robotic
vehicles. As part of this work, he designed and built the JHUROV, a new ROV
used for research in underwater vehicle navigation, dynamics, and control. His
research interests are in the dynamics and control of electromechanical systems.
Louis L. Whitcomb (S’87–M’87–SM’02) received the B.S. and Ph.D. degrees
from Yale University, New Haven, CT, in 1984 and 1992, respectively.
From 1984 to 1986, he was an R&D Engineer with the GMFanuc Robotics
Corporation. From 1992 to 1994, he was a Postdoctoral Fellow with the Woods
Hole Oceanographic Institution, Woods Hole, MA, and the University of Tokyo,
Tokyo, Japan. He is an Associate Professor with the Department of Mechanical
Engineering, with joint appointment in the Department of Computer Science,
The Johns Hopkins University, Baltimore, MD. He is codeveloper of naviga-
tion and control systems for underwater vehicles including the robotic vehicles
JHUROV, Jason II, DSL120A, Isis, Hercules, and the D.S.V. Alvin inhabited
submersible. His research focuses on the navigation, dynamics, and control of
robotic vehicles and systems.