System Identification and Adaptive Input Estimation on the

Jaiabot Micro Autonomous Underwater Vehicle

Ioannis Faros1† and Herbert G. Tanner1†
1 Mechanical Engineering Department, University of Delaware, 130 Academy St, Newark,
19716, DE, USA.

Contributing authors: [email protected]; [email protected];

arXiv:2504.02005v1 [eess.SY] 1 Apr 2025

† These authors contributed equally to this work.

Abstract
This paper reports an attempt to model the system dynamics and estimate both the unknown internal
control input and the state of a recently developed marine autonomous vehicle, the Jaiabot. Although
the Jaiabot has shown promise in many applications, process and sensor noise necessitates state
estimation and noise filtering. In this work, we present the first surge and heading linear dynamical
model for Jaiabots derived from real data collected during field testing. An adaptive input estimation
algorithm is implemented to accurately estimate the control input and hence the state. For validation,
this approach is compared to the classical Kalman filter, highlighting its advantages in handling
unknown control inputs.

Keywords: Input estimation, System identification, Underwater Vehicles

1 Introduction Peng et al., 2021; Li and Du, 2021). These AUVs

can be deployed from shore and offer advantages
Over the last decades, the effects of climate in terms of affordability and open source access.
change, including rising sea levels, disruptions One such micro-AUV that shows great promise
to marine life, and declining water quality, have for environmental monitoring application is the
become increasingly prominent (Liu et al., 2016; Jaiabot (Fig. 1). This vehicle is approximately
Zereik et al., 2018). As a result, the demand for one meter long, weighs about 3 kg, and is capa-
accurate environmental monitoring and in situ ble of achieving speeds of nearly 5 m/s with
measurements around coastal regions has grown a range of 11 km. In order to perform vertical
significantly. To observe such phenomena and col- dives or general movement within the sea, the
lect data through traditional methods, which pri- Jaiabot is equipped with a single propeller, rud-
marily rely on in situ measurements from station- der, GPS, and compass, inertial measurement unit
ary platforms or manual data collection, can either (IMU), plus with sensors for collecting data such
yield very sparse data sets or be time consuming as salinity and temperature. Few attempts at con-
and costly. From this perspective, such data col- structing dynamical models for such micro-AUVs
lection tasks can be serviced by automating these have been reported. The first known dynamical
processes by using one or multiple autonomous system for a Jaiabot is known for its vertical dive
underwater vehicles (AUVs) (Lim et al., 2023; motion (Tanner et al., 2024), based on which, a

1
 estimation (RCIE) has been developed (Ansari
and Bernstein, 2018; Sanjeevini and Bernstein,
2022). RCIE formulates a retrospective cost opti-
mization problem, where the coefficients of the
Fig. 1: The first generation of the Jaiabot AUV. input estimator are recursively adjusted to mini-
mize a (retrospective) cost function. By doing so,
RCIE effectively builds an internal model of the
control for diving is redesigned and developed with unknown input that estimates the later, which is
new safety constraints, specifically with respect to subsequently fed into the Kalman filter. The esti-
overshoot in terms of assigned depth. Subsequent mator coefficients are continuously adapted using
design modifications made on this AUV require the innovations (differences between predicted
new system parameter identification, while the and observed measurements) as the error metric,
aforementioned approach applies still. thereby enhancing the accuracy and robustness of
While the controlling the Jaiabot during ver- state estimation. RCIE has been extensively stud-
tical dive without overshoot has been addressed, ied and modified for nonminimum-phase discrete
open questions still remain regarding its surface time systems, linear time varying (Sneha and Den-
maneuvering. Thus in this paper, the Jaiabot is nis, 2024) and invariant systems (Rahman et al.,
operated mainly as an autonomous surface vehi- 2016). It has also been applied in the area of
cle (ASV). Specifically, noisy data collected from signal and process and especially in numerical dif-
the miniature onboard GPS and IMU may result ferentiation (Verma et al., 2024) and integration
in inaccurate positioning. In classical state esti- (Sanjeevini and Bernstein, 2023).
mation, the Kalman filter and its variants give The contributions of the work reported in this
well-established techniques for estimating unmea- paper are outlined as follows:
sured states by taking advantage of knowledge • Modeling surge and heading dynamics: The
about the system dynamics, its input, as well the surge and heading dynamics of Jaibot are
process and sensor noise. However, while the PID derived from field data.
control loop that regulates thruster speed and • Adaptive input estimation (AIE): The adaptive
rudder configuration of this ASV is known, the input estimation based on the RCIE (Verma
actual thrust generated to propel the ASV or the et al., 2024), is implemented to the derived mod-
torque generated to make it turn is unknown. This els of Jaiabot to estimate the internal control
presents challenges to the implementation of a input and hence the state of the system using
Kalman filter on the surface maneuvering dynam- real data.
ics. Specifically, here both the state of the system
and its input have to be estimated simultaneously. The remaining of the paper is organized as fol-
This is not a new problem. Numerous meth- lows. Section 2 frames the process of modeling
ods have been used for state estimation in noisy the surge and heading dynamics from real data.
environments (Paull et al., 2013). However, when Section 3 provides the simulation results using real
the input signal is deterministic but unknown, data in support of the derived models, the imple-
obtaining unbiased state estimates becomes cru- mented AIE algorithm, and give a comparative
cial. Solutions to this problem include unbiased analysis between AIE and the classical Kalman fil-
Kalman filters, unknown input observers, and ter. Section 4 summarizes the research outcomes
sliding-mode observers (Kitanidis, 1987; Darouach and outlines directions for future work.
and Zasadzinski, 1997; Veluvolu and Soh, 2009).
An alternative input estimation approach con- 2 Technical Approach
siders the unknown input as the output of an
auxiliary system with known dynamics, perturbed 2.1 System Identification
by white noise. This estimated input can then be
incorporated into the state estimator to yield more The Jaiabot is a miniature marine vehicle that
accurate state estimates. Over the last decade, a can move fast on the surface of the water. To
new technique known as retrospective cost input perform principled control design for automated

2
maneuvering for this vehicle we constructed mod- 2.1.2 Heading dynamics
els based on experimental data. There can be
The identification of the heading (yaw) dynam-
several general model templates for surface vehi-
ics of the Jaiabot was performed using a standard
cle kinematics and dynamics (Fossen, 2021), and
maneuvering test for marine vehicles (Fossen,
most advanced such models incorporate coupling
2021). The vehicle moves with a fixed rudder con-
between longitudinal (surge) and lateral (yaw)
figuration over a period of time (Fig. 2b), and the
vehicle motions (Panagou et al., 2015). Given
response of the yaw rate is measured (in this case,
the choice of parameterizing the model based on
based on the vehicle’s IMU and GPS readings).
experimental data, we opted to initially ignore the
Assuming a similar second order model for the yaw
coupling between surge and yaw (see also (Fossen,
dynamics
2021)) and built separate models for surge and
I θ̈ + c θ̇ = r (2)
yaw motion.
where I is in the role of the vehicle’s moment of
inertia along the vertical axis, c is a hydrodynamic
2.1.1 Surge dynamics drag coefficient and r expresses the rudder input
The surge dynamics of the vehicle were assumed command, a least squares fitting approach on the
(Fossen, 2021) to take the form of a second order step response for this model gives I = 4.896 and
linear system c = 9.087.
m ẍ + d ẋ = u (1) In the section that follows, the surge and yaw
where m denotes the mass of the vehicle, d is dynamics, (1) and (2) respectively, are expressed
in the role of a hydrodynamic drag coefficient in discrete time, in order to be integrated into a
along the surge direction, and u expresses for- Kalman filter.
ward thrust. Equivalently, (1) can be regarded as
a first order system in surge speed; either way, 2.2 Discrete-time Vehicle Dynamics
the step response of this model is known analyti- Jaiabot’s motion, just like most surface vehicles,
cally and thus knowing the mass of the vehicle, the is along curves; neither straight lines nor pure
coefficient d can be directly determined through rotations. Thus none of the models of (1) or
a least squares process from experimental data (2) can, in isolation, describe the motion of the
(Fig. 2a) The fitting process, accounting for input vehicle. What is more, to be integrated into a
Kalman filter, these models need to be associated
with measurable outputs. And while the vehicle’s
orientation is directly measured via its IMU or
compass, the a single GPS measurement does not
directly inform about the length of the path trav-
eled, which is practically what is being tracked by
(1). To overcome the latter challenge we construct
discrete-time models that describe the evolution of
(a) Surge tests (b) Yaw tests path length traversed by the vehicle and its change
in orientation between two consecutive time steps.
Fig. 2: Maneuvering experiments for motion
dynamics system identification. (a) Experimen- 2.2.1 Heading
tal data for surge dynamics were obtained from
We start with the incremental heading dynamics
indoor tests in a long water tank. (b) Data for yaw
because they are more straightforward. If we name
dynamics were collected during outdoor tests in
z1 a variable capturing the vehicle’s current head-
lake Allure, PA
ing and z2 its yaw rate, then the discrete-time
heading dynamics derived by analytic integration
of (2) for a time step T 1 results in
scaling, yields parameter values for (1) as follows:
m = 0.469, and d = 0.311.

1
In implementation, T is set to 0.546 seconds.

3
 h i
1 0.5487(1−e−1.82249T )
[ zz12 ]k+1 = [ zz12 ]k with ZOH over rk , with the distance between the
0 e−1.82249T
vehicle at consecutive steps, ∆S̄, and express the
0.5487+0.301072(e−1.82249T )
h i
+ 0.5487(1−e−1.82249T )
rk latter as
q
We now augment the heading state vector at ∆Sk+1 ≈ ∆S̄k+1 = ∆Sx 2 + ∆Sy 2 (7)
step k + 1 to include the heading at step k as so
where ∆Sx and ∆Sy can be computed from GPS
measurement differences after projection to the
 h
1 0.5487(1−e−1.82249T ) 0
h z1 i z1 i
z2
z3 k+1
= 0 e−1.82249T 0
z2
z3 k body frame at time step k + 1 (see Fig. 3).
1 0 0
0.5487+0.301072(e−1.82249T )
 
+ 0.5487(1−e−1.82249T ) rk (3)
0
∆θ̄
and define the output of this discrete-time linear y x
system to be ∆θ̄ which relates to the heading state ∆S̄
z through step k step k + 1
∆S
h z1 i ∆Sx
∆θ̄k+1 = 0.200474 [ 1 0 −1 ] z2 (4) ∆Sy
z3 k+1

The output variable can be considered directly Fig. 3: Geometric relation between surge model
measurable in the form of the difference of com- output and sensor measurements.
pass readings between two time steps.

2.2.2 Surge A remaining challenge, however, is that during

deployment and due to the vehicle’s control archi-
The discrete-time augmented dynamics for surge tecture and interface, neither the input thrust u
is constructed along the same lines. We denote z1 in (1) nor the input moment r are directly known.
the length of the path traveled by the vehicle, z2 This is where retrospective cost input estimation
its speed along this path, and z3 the path length comes in to provide real-time estimates of uk and
at the previous time step. With these we arrive at rk through an adaptive estimation process.
 h
1 1.50625(1−e−0.66397T ) 0 2.3 Adaptive Input Estimation
h x1 i x1 i
x2
x3 k+1
= 0 e−0.66397T 0
x2
x3 k
1 0 0
This section provides a mathematical overview of
1.50625T +2.26879(e−0.66397T )
 
+ 1.50625(1−e−0.66397T ) uk (5) the AIE technique Verma et al. (2024). Consider
0 the linear discrete-time system

Similarly, the output for this discrete-time system xk+1 = A xk + Buk

is defined as the difference ∆S of path lengths
yk = C xk + V
between consecutive steps
h x1 i where xk ∈ Rn is the system state, uk ∈ R is the
∆Sk+1 = 1.4162 [ 1 0 −1 ] x2
x3 k+1
(6) control input (assumed unknown), and V ∈ R is
zero-mean Gaussian sensor noise. Matrices A ∈
The challenge here, however, is that ∆S is not Rn×n , B ∈ Rn×1 , and C ∈ R1×n are known. The
provided as a sensor measurement. Nonetheless, goal of AIE is to estimate simultaneously both
this quantity can be derived directly through geo- uk and xk . AIE is practically comprised of an
metric conditions that depend explicitly on com- input estimation subsystem, and a Kalman filter
pass and GPS measurements. Without excessive (Thacker and Lacey, 1998).
loss, therefore, we approximate the incremental
path length, ∆S, assumed at the scale of the time
step to be adequately captured by a circular arc

4
2.3.1 Input estimation subsystem where now Hi (k) is defined as
To obtain the estimated input ûk , we construct 
the input estimation subsystem of order ne > 1 as CB
 k≥i=1
follows Hi (k) ≜ C Āk−1 · · · Āk−(i−1) B k≥i≥2

0 i>k

ne
X ne
X
ûk = Mi,k ûk−i + Ni,k zk−i , (8)
i=1 i=1 with Ā ≜ A(I + Kk C), and Kk being the Kalman
filter gain that is included in the update step of
where the Mi,k ∈ R, Ni,k ∈ R, and zk is the the state filtering process. To find the coefficient
residual in the prediction step of Kalman filter. vector θk , we construct an optimization prob-
Subsystem (8) can be written in form of lem; more specifically a retrospective optimization
problem where the coefficient vector will denote
ûk = Φk θk , (9) the optimization variable. To this end, define the
retrospective variable as
where the regressor matrix Φk is defined as
zr,k (θ̂) ≜ zk − (ûf,k − Φf,k θ̂)
Φk ≜ [ûk−1 · · · ûk−ne zk · · · zk−ne ] ∈ R1×lθ (10)
where now the θ̂ represents the optimization argu-
the coefficient vector θk is ment. Next, define the retrospective cost function

θk ≜ [M1,k · · · Mne ,k N0,k · · · Nne ,k ] ∈ Rlθ (11) k

X
2
Jk (θ̂) ≜ (θ̂−θ0 )⊺Rθ (θ̂−θ0 )+ Rz zr,i (θ̂)+Rd (Φi θ̂)2
with lθ ≜ 2ne +1. The order of this subsystem, ne , i=0
must be chosen large enough to properly develop
the internal model for the input estimation. The where Rz ∈ (0, ∞), Rd ∈ (0, ∞) are scalar opti-
objective now becomes to update the coefficient mization gains, and Rθ ∈ Rlθ ×lθ is a positive
vector θk , in order to derive an estimate control definite gain matrix. Note that the regularization
input. To do so, we first define the backward-shift term (θ̂−θ0 )⊺ Rθ (θ̂−θ0 ) weighs the initial estimate
operator for a discrete signal Yk as and ensures that the θk+1 has a unique global min-
imizer (Islam and Bernstein, 2019). Define now
q −1 Y (k) = Y (k − 1) P0 ≜ Rθ−1 . Then for all k ≥ 1, the unique global
minimizer θk+1 , is given by the recursive least
and thus we define the following filtered signals squares (RLS) update

Φf,k ≜ Gf,k (q −1 ) Φ(k − i) Pk+1 =Pk − Pk Φ̃⊺k Γk Φ̃k Pk

ûf,k ≜ Gf,k (q −1 ) û(k − i) θk+1 =θk − Pk Φ̃⊺k Γk (z̃k + Φ̃k θk ) (12)

Pnf −i where
. By defining Gf,k (q−1 ) ≜ i=1 q Hi (k), with
nf ≥ 1 being the window length of filter, the
−1

Φf,k

filtered signals can be written as Γk ≜ R̃ +−1
Φ̃k Pk Φ̃⊺k Φ̃k ≜
Φk
zk − dˆf,k
   
nf Rz 0
X z̃k ≜ R̃ ≜
Φf,k = Hi (k) Φ(k − i) 0 0 Rd
i=1
nf By using (12) and replacing the k + 1 with k in (9)
X
ûf,k = Hi (k) û(k − i) we derive the estimated input. We choose θ0 = 0
i=1 which implies û0 = 0.

5
Remark 1. To properly implement the AIE algo- 3.1 Results for Surge Dynamics
rithm, it is essential to first specify all the hyper-
First, the AIE algorithm has been implemented
parameters ne , nf , Rz , Rd , Rθ , typically done
for the surge dynamics. Figure 4 gives a compre-
empirically through trial and error.
hensive view of the results obtained from AIE.
Remark 2. This subsystem is highly sensitive to
As expected, the coefficient vector θk in (11) con-
hyperparameter variations, and small changes in
verges after approximately 100 iterations (Fig. 4,
them can either yield the desired results or result
top). Nine coefficients (lθ = 2ne + 1) to achieve
in significantly high values for the estimated input.
proper convergence for this subsystem. Figure 4,
Remark 3. Large values of the hyperparameter
middle, illustrates the estimated input for the
nf or ne do not necessary lead to better filtering
surge dynamics. The dashed red line shows the
of the signals or better estimate control input. On
average of the values that the input takes. It
the contrary, they might cause divergence of the
can be observed that the control input fluctuates
input estimate.
within a small range ([−0.05, 0.30]), an observa-
tion that also aligns with the behavior of the
3 Validation estimated state, shown in Fig. 4, bottom. Note
here that the output (and measurement) for the
This section presents simulation results and
surge dynamics represents the length of the chord
numerical analysis that supports the theoretical
of the arc along which the Jaiabot is moving.
predictions on (a) estimation of the determin-
Hence, small values in this context indicate small
istic control input as it is presented in Section
steps along the motion path.
2.3, and (b) the effect of the estimated input to
On the other hand, Fig. 5 displays a compar-
the Kalman filter to estimate the state of the
ison of state estimation results. The red dashed
Jaiabot. The approach is applied to the problem
curve represents measurements, the blue curve
of estimating the states of surge (5)–(6) and head-
shows the state estimate obtained from the
ing model (3)–(4). To achieve this, the control
Kalman filter that uses a nominal (step) control
inputs for both the surge and heading dynam-
input set at uk = 1, and the black curve illustrates
ics must first be estimated, because the operator
the output of AIE. It is seen that AIE provides sig-
has no direct knowledge of the thrust and torque
nificantly more effective noise filtering compared
inputs. (The relationship between operator com-
to the standalone Kalman filter.
mands and thrust/torque inputs can be empiri-
cally established via hydrodynamic experiments.)
Theta Convergence
For comparison reasons, a standard Kalman filter 0.30
0.25
Theta[0]
Theta[1]
Theta[2]
is also implemented with the assumption that the 0.20
0.15
Theta[3]
Theta[4]
Theta[5]
nominal operator control input is the one that is 0.10
0.05
Theta[6]
Theta[7]
Theta[8]
0.00
implemented by the Jaiabot. 0.05
0 25 50 75 100 125 150 175
Part of the implementation of the algorithm Adaptive Input Estimation
Estimated Input
for estimation of control inputs and the states of 0.25
Average Input
0.20
the two systems, is the process of tuning all the 0.15
0.10
hyperparameters. In our tests, we set the following 0.05
0.00
values. For the surge dynamics: ne = 4, nf = 8, 0.05
0 25 50 75 100 125 150 175
Rz = 1, Rd = 50, Rθ = 10−0.01 I9 . For the heading State Estimation
5 Data
dynamics: ne = 3, nf = 4, Rz = 1, Rd = 0.1, 4
AIE

Rθ = 10−2 I7 . 3
2
The algorithms used experimental data col- 1
0
lected in Lake Allure, PA, where the Jaiabot was 0 25 50 75 100 125 150 175
Timesteps
deployed in a series of turning maneuvers (see
Fig. 8). Fig. 4: Adaptive input estimation results for
surge.

6
 struggles to properly filter the process during sig-
Adaptive Input Estimation
Data
Kalman
nificant fluctuations without accurate knowledge
5 AIE
of the system input. One the other hand, AIE
4 framework demonstrates improved performance in
scenarios involving large fluctuations, providing
3
more robust and accurate estimates compared to
2
the standalone Kalman filter.

1 Theta Convergence
1.00 Theta[0]
0.75 Theta[1]
0
Theta[2]
0.50 Theta[3]
0 25 50 75 100 125 150 175 0.25 Theta[4]
Timesteps 0.00 Theta[5]
Theta[6]
0.25
0.50
Fig. 5: Comparison of the estimated state for 0 25 50 75 100 125 150 175

surge against experimental data, for the Kalman Adaptive Input Estimation
150 Estimated Input
Average Input
filter and AIE. 100
50
0
50

3.2 Results for heading dynamics 100

0 25 50 75 100 125 150 175
State Estimation
Data
In this subsection we report on the AIE imple- 300
200
AIE

mentation for the heading dynamics. Figure 6 100

0
100
summarizes the findings. Note that for this model 200
300
the measurement is the heading (angle) which is 0 25 50 75 100 125 150 175

obtained directly from the AUV sensors. Similarly Timesteps

to the AIE implementation for the surge dynam- Fig. 6: Adaptive input estimation results for
ics, the coefficient vector θk in (11) converges after heading.
approximately 100 iterations (Fig. 6, top). For
heading estimation, only 7 coefficients are needed
to achieve proper convergence. Figure 6, middle,
shows the estimated input for the heading dynam-
ics. The dashed red curve represents the average Adaptive Input Estimation
Kalman
of the values that the input takes over the whole 300
Data
AIE
period of application. In this execution, it can be
200
observed that the control input fluctuates within a
range ([−100, 170]), in correlation with the behav- 100

ior of the estimated state, shown in Fig. 6, bottom, 0

which depicts the estimated state. As seen, during 100

instances of significant fluctuations, such as within 200

the range [80, 110], the black line representing the
300
estimation effectively filters out the initial noisy
0 25 50 75 100 125 150 175
behavior. Timesteps
Lastly, Fig. 7 displays a comparative analy-
sis of the state estimation results. In Fig. 7, the Fig. 7: Comparison of the estimated state
red dashed line represent system measurements, for heading against experimental data, for the
while the blue lines show the estimated states Kalman filter and AIE.
obtained from the Kalman filter, with the nom-
inal control input set to uk = 1. The black line
shows the estimated state from the AIE algorithm. Application of AIE suggests that the DC gain
It can be observed that while the Kalman fil- of the surge dynamics (5) is 4.68 while that for the
ter provides a good estimate for the heading, it heading dynamics (3) is 0.125.

7
3.3 Trajectory reconstruction based 4 Conclusion
on estimates
The proper choice of the control input in a sys-
This section makes a combined spatial compara- tem is of the essence for an effective control of
tive analysis between AIE and standalone Kalman a dynamical system. In cases where the control
filter, as the ASV motion path is reconstructed input is unknown, it becomes necessary to esti-
based on the corresponding estimates and com- mate it in conjunction to the state of the system.
pared to experimental (GPS) data. In both cases This paper implements this concept for the first
(AIE and standalone Kalman filter) the trajectory time on an ASV to generate filtered pose (position
is reconstructed as & orientation) estimates using linear models for
  its surge and heading dynamics, identified inde-
∆θ pendently based on experimental data. Results
Xk+1 =Xk + ∆S sin
2 indicate promise for AIE as a method for providing
 
∆θ accurate and consistent real-time state estimates
Yk+1 =Yk + ∆S cos
2 for an ASV, the actual control inputs of which are
opaque to its operator.
where Xk and Yk are the north-south and east-
Acknowledgements. The authors thank
west coordinates of the ASV at time step k. Here,
Chanaka Bandara for collecting and sharing the
∆θ is derived from the heading dynamics, either
ASV data at Lake Allure PA. Funding for this
using the AIE or the Kalman process, while the
work is provided in part through the USGS Next
∆S is estimated similarly. Figure 8 depicts the
Generation Water Observing System (NGWOS)
results of this reconstruction for the ASV trajec-
Research and Development program via award
tory. It can be observed that AIE outperforms the
#G23AC00149-00 and by NSF through award
Kalman filter since it is adjusting the control input
#2234869.
to be suitable for estimation of the state. On the
other hand, Kalman filter presents big fluctuations
along the trajectory, primarily due to the control References
input that was selected from the user, which does
not accurately reflect the actual internal control Ansari, A., Bernstein, D.S.: Input estimation for
input implemented by the ASV. nonminimum-phase systems with application
to acceleration estimation for a maneuvering
vehicle. IEEE Transactions on Control Systems
Trajectory Technology 27(4), 1596–1607 (2018)
Data
15 AIE
Kalman Filter Darouach, M., Zasadzinski, M.: Unbiased min-
10
imum variance estimation for systems with
5
unknown exogenous inputs. Automatica 33(4),
0
717–719 (1997)
Y

10 Fossen, T.I.: Handbook of Marine Craft Hydro-

15 dynamics and Motion Control, 2nd edn. John
20
0 5 10 15 20 25 30
Wiley & Sons, ??? (2021)
X
Fig. 8: Comparison of the trajectory reconstruc- Islam, S.A.U., Bernstein, D.S.: Recursive least
tion based on the Kalman filter and the AIE squares for real-time implementation [lecture
relative to experimental data. notes]. IEEE Control Systems Magazine 39(3),
82–85 (2019)

Kitanidis, P.K.: Unbiased minimum-variance lin-

ear state estimation. Automatica 23(6), 775–
778 (1987)

8
Li, D., Du, L.: Auv trajectory tracking models and control of linear time-varying systems. IEEE
control strategies: A review. Journal of Marine Systems and Control Letters 185, 105744
Science and Engineering 9(9), 1020 (2021) (2024)

Lim, H.S., Filisetti, A., Marouchos, A., Khosoussi, Tanner, H.G., Bandara, C.T., Gyves, M.C.: Par-
K., Lawrance, N.: Applied research directions tial system identification and sensor fusion with
of autonomous marine systems for environmen- the jaiabot micro autonomous underwater vehi-
tal monitoring. OCEANS 2023-Limerick, 1–10 cle. In: Proceedings of the 2024 IEEE Mediter-
(2023) ranean Conference on Control and Automation,
pp. 167–172 (2024)
Liu, Z., Zhang, Y., Yu, X., Yuan, C.: Unmanned
surface vehicles: An overview of developments Thacker, N., Lacey, A.: Tutorial: The kalman
and challenges. Annual Reviews in Control 41, filter. Imaging Science and Biomedical Engi-
71–93 (2016) neering Division, Medical School, University of
Manchester 61 (1998)
Paull, L., Saeedi, S., Seto, M., Li, H.: Auv naviga-
tion and localization: A review. IEEE Journal Veluvolu, K.C., Soh, Y.C.: High-gain observers
of Oceanic Engineering 39(1), 131–149 (2013) with sliding mode for state and unknown input
estimations. IEEE Transactions on Industrial
Panagou, D., Tanner, H.G., Kyriakopoulos, K.J.: Electronics 56(9), 3386–3393 (2009)
Nonholonomic control design via reference vec-
tor fields and output regulation. ASME Journal Verma, S., Sanjeevini, S., Sumer, E.D., Bern-
of Dynamic Systems, Measurement and Control stein, D.S.: Real-time numerical differentiation
137(8), 081011–0810119 (2015) of sampled data using adaptive input and state
estimation. International Journal of Control,
Peng, Z., Wang, J., Wang, D., Han, Q.-L.: An 1–13 (2024)
overview of recent advances in coordinated con-
trol of multiple autonomous surface vehicles. Zereik, E., Bibuli, M., Mišković, N., Ridao, P.,
IEEE Transactions on Industrial Informatics Pascoal, A.: Challenges and future trends in
17(2), 732–745 (2021) marine robotics. Annual Reviews in Control 46,
350–368 (2018)
Rahman, Y., Xie, A., Hoagg, J.B., Bernstein,
D.S.: A tutorial and overview of retrospective
cost adaptive control. In: Proceedings of the
2016 IEEE American Control Conference, pp.
3386–3409 (2016)

Sanjeevini, S., Bernstein, D.S.: Decomposition of

the retrospective performance variable in adap-
tive input estimation. In: Proceedings of the of
the 2022 IEEE American Control Conference,
pp. 242–247 (2022)

Sanjeevini, S., Bernstein, D.S.: Random-walk

elimination in numerical integration of sensor
data using adaptive input estimation. In: Pro-
ceedings of the 2023 IEEE American Control
Conference, pp. 4791–4795 (2023). https://doi.
org/10.23919/ACC55779.2023.10155990

Sneha, S., Dennis, S.B.: Performance-variable

decomposition in retrospective cost adaptive