HAL Id: inria-00202698
https://hal.inria.fr/inria-00202698
Submitted on 7 Jan 2008
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entic research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diusion de documents
scientiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Experiments in Navigation and Mapping with a
Hovering AUV
George Kantor, Nathaniel Faireld, Dominic Jonak, David Wettergreen
To cite this version:
George Kantor, Nathaniel Faireld, Dominic Jonak, David Wettergreen. Experiments in Navigation
and Mapping with a Hovering AUV. 6th International Conference on Field and Service Robotics -
FSR 2007, Jul 2007, Chamonix, France. �inria-00202698�
Experiments in Navigation and Mapping with a Hovering AUV
George Kantor, Nathaniel Fairfield, Dominic Jonak and David Wettergreen
The Robotics Institute
Carnegie Mellon University
Pittsburgh, Pennsylvania 15213 USA
Email: {kantor, than, dom, dsw}@cmu.edu
Summary. This paper describes the basic control, navigation, and mapping methods and experiments a hovering autonomous
underwater vehicle (AUV) designed to explore flooded cenotes in Mexico as part of the DEPTHX project. We describe the low
level control system of the vehicle, and present a dead reckoning navigation filter that compensates for frequent Doppler velocity
log (DVL) dropouts. Sonar data collected during autonomous excursions in a limestone quarry are used to generate a map of the
quarry geometry.
1 Introduction
The DEPTHX (DEep Phreatic THermal eXplorer) project is a three-year NASA-funded effort whose primary objec-
tive is to use an autonomous vehicle to explore and characterize the unique biology of the Zacat
´
on cenote. Zacat
´
on,
the world’s deepest known limestone sinkhole, is a water-filled cavern that is at least 300 meters deep. The depths
of Zacat
´
on are geothermally heated with a high sulfur content and a lack of sunlight or dissolved oxygen, making
this an ideal place to search for exotic microbial life [Gary, 2002]. The robotic exploration and search for microbial
life in Zacat
´
on is an analog mission for the search for life in the liquid water ocean beneath the frozen surface of
Europa.
The DEPTHX robot (Figure 1) is a hovering autonomous underwater vehicle (AUV) designed to explore flooded
caverns and tunnels while building 3D maps, collecting environmental data, and obtaining samples from the water
column and cavern walls. To accomplish these tasks, the vehicle is equipped with a Doppler velocity logger (DVL),
a ring laser gyro-based inertial navigation system (INS), a depth sensor, and an array of 54 narrow beam sonar
transducers. In the Zacat
´
on mission, the vehicle will use all of these sensors together to perform simultaneous
localization and mapping (SLAM). This paper presents a preliminary field result where the INS, DVL, and depth
sensor are used together to generate a dead reckoned position estimate and the sonar array is used to create a map
based on that estimate. This technique was used to autonomously map a limestone quarry in Austin, Texas during
field tests conducted in January of 2007.
The paper is organized as follows: Related research activities in using AUVs for science and mapping are briefly
discussed in Section 2. Then we provide a description of the instrumentation of the DEPTHX vehicle in Section 3.
This is followed by a description of our implementation of a Kalman filter to merge DVL and IMU measurements,
allowing for high performance dead reckoned position estimation even in cases where the DVL frequently drops out
(Section 4). An overview of the vehicle control system is provided in Section 5. Experimental results are presented
in Section 7, followed by some conclusions and a discussion of future work in Section 8.
2 Related Work
There are more than 1000 commercial underwater vehicles in operation today, performing tasks as diverse as ar-
chaeological excavation to drill-rig maintenance to oceanographic survey [Whitcomb, 2000].
There are a variety of techniques employed for determining the position of underwater vehicles – unfortunately
GPS does not work underwater – and they can be divided into those that utilize emplaced infrastructure and those
that do not. (See [Leonard et al., 1998] for a comprehensive survey.)
2 George Kantor, Nathaniel Fairfield, Dominic Jonak and David Wettergreen
Fig. 1: Left: the DEPTHX AUV deployed in a test tank at Austin Research Laboratory. Right: A model of the DEPTHX vehicle
structure and components. Eleven pressure vessels house computing, batteries, sensors, and science instruments. Diameter is
approximately 2m.
c
Stone Aerospace, 2006.
When accurate position is needed underwater, most AUVs and many human-driven remotely operated vehicles
(ROVs) rely on a surveyed array of acoustic beacons, known as a long base-line (LBL) array. Acoustic beacons
provide a fixed frame of reference for positioning the vehicle [Whitcomb, 2000]. [Yoerger et al., 2007] have shown
detailed sea-floor mapping of subsea vents using an LBL array with the Autonomous Benthic Explorer.
Over large distances, or in underwater caves and tunnels, the performance of LBL systems is unknown due to
signal attenuation, reverberation and multipath. Without a fixed LBL infrastructure, an AUV uses a combination of
depth sensors, inertial sensors, and Doppler velocity sensors to compute a dead reckoned estimate of its position
while at depth. With high accuracy attitude and depth sensors the uncertainty in the AUV’s 3D pose (roll, pitch,
yaw, x, y, z) is primarily in x and y. Most underwater navigation systems are based on a Kalman Filter which
combines Doppler velocity and inertial measurements [Larsen, 2000]. These systems report navigation errors as
low as 0.015% of distance traveled, however error accrues drastically in situations where the DVL is unable to make
accurate velocity measurements. [Kirkwood et al., 2001] have developed an AUV for multi-day under-ice arctic
survey.
Eventually the dead reckoned estimate will drift from true, and when the accumulated error exceeds what
is required for the application, a correction must be made by (re)observing a known reference. A common ap-
proach is to surface and obtain position from a GPS. If LBL or surfacing is not an option, the position error can
be bounded by simultaneous localization and mapping (SLAM) [Williams et al., 2000] [Dissanayake et al., 2001].
[Williams and Mahon, 2004] provide an example of near-bottom mapping with sonars and cameras of coral reefs
using SLAM. [Roman, 2005] uses multibeam sonar maps to do SLAM over varied topography.
3 Vehicle Description
The DEPTHX vehicle is a hovering AUV that has been specifically designed for exploration of flooded caverns and
tunnels. It has an ellipsoidal shape that measures approximately 1.5 meters in height and 1.9 meters in both length
and width (Figure 1). Its dry mass is 1500kg. Four large pieces of syntactic foam are mounted on the top half of the
vehicle, passively stabilizing the vehicle roll and pitch. The vehicle can move directly in the remaining four degrees
of freedom (forward, starboard, down, and heading) using six thrusters driven by brushless DC motors. The cruising
speed of the vehicle is about 0.2 meters per second. The vehicle is powered by two 56-volt Lithium-Ion battery
stacks with a total capacity of 6.2 KWh, enough to supply the vehicle during an four-hour exploration mission.
The DEPTHX vehicle has a full suite of underwater navigation sensors, including a Honeywell HG2001AC
INS, two Paroscientific Digiquartz depth sensors, and an RDI Navigator 600kHz DVL. The specifications for the
INS are roll/pitch: 0.2
◦
2σ, yaw: 0.4
◦
2σ, for the DVL velocities 0.3 cm/s 1σ, and for the depth sensors 0.01% of full
range (10 cm for our 1000m rated sensor). The two depth sensors are tared, or zeroed with respect to atmospheric
pressure, at the start of each day. The DVL is mounted to the front of the vehicle facing forward and tilted down
Experiments in Navigation and Mapping with a Hovering AUV 3
30 degrees from horizontal, a nonstandard configuration for this instrument. The usual DVL configuration points
straight down so that it can achieve lock on the ocean floor. In our application, it is difficult to predict the relative
direction to surfaces useful for DVL lock. The top 280 meters of Zacat
´
on is known to be a chimney with a diameter
of approximately 80 meters [Fairfield et al., 2005], so the forward-looking configuration should allow the DVL to
lock on to one of the vertical walls in most situations. This configuration can cause the DVL to lose bottom lock in
more wide-open waters, such as the quarry that is the subject of this paper. Loss of bottom lock can also occur at
extremely short ranges, or when passing over highly irregular terrain.
For our purposes, the raw roll, pitch, and yaw measurements provided by the IMU and the depth measurements
provided by the depth sensors are accurate enough to be considered absolute measurements of those quantities. The
task of determining the location of the vehicle is then reduced to the two dimensional problem of estimating its
position in the lateral plane (x and y).
For mapping, the vehicle has an array of 54 2
◦
beam-width sonars that provide a constellation of range measure-
ments around the vehicle. This array is in the shape of three great circles, a configuration that was arrived at after
studying the suitability of various sonar geometries for the purposes of SLAM [Fairfield et al., 2005]. The sonars
have long ranges (some 100m and others 200m) and the accuracy of the range measurements is fairly high (about
10cm), however the low resolution, update rate, and point density makes the mapping problem significantly more
difficult than it is with ranging sensors like a laser scanner that provide fast, accurate, high-resolution range images.
4 Dead Reckoning Localization
The concept of dead reckoning from DVL velocities is straightforward: integrate the velocities to obtain a position
estimate. Here we lay out the details of how this is done on the DEPTHX vehicle.
4.1 Dead Reckoning with DVL Velocities
The raw DVL measurement is a vector specifying the velocity of the DVL in the DVL’s own coordinate frame.
Using the notation of [Spong et al., 2006], we denote this vector by v
dvl
dvl
. We actually want to know the velocity of
the origin of the vehicle frame (which is located at the centroid of the vehicle) represented in the world frame, which
we denote by v
w
v
. To get this, we first compute the velocity of the vehicle in the vehicle frame:
v
v
v
= R
v
dvl
v
dvl
dvl
− ω
v
v
× o
dvl
.
Here, R
v
dvl
is the constant rotation matrix describing the orientation of the DVL relative to the vehicle frame;
ω
v
v
= [ω
x
ω
y
ω
z
]
T
are the angular velocities in vehicle frame (as measured by the IMU); and o
dvl
is the position of
the DVL in the vehicle frame (or lever-arm.) It is then a simple matter to rotate the vehicle frame velocities into the
world frame
v
w
v
= R
w
v
v
v
v
,
where R
w
v
is the rotation matrix describing the orientation of the vehicle relative to the world frame, as measured
by the IMU. The x and y components of v
w
v
are then numerically integrated to get the new dead reckoned x and y
position estimates using Euler’s method.
4.2 Patching DVL Dropouts
As discussed above, the forward-looking configuration of the DVL together with uneven bottom conditions causes
it to frequently lose bottom lock in open water situations. During these times, no DVL velocity measurements are
available, making it impossible to update the dead reckoned position estimate. Another unfortunate symptom of
losing bottom lock is that the last few DVL measurements before and the first few measurements after the loss of
lock are likely to be noisy. In order to filter out these measurements, we used a simple first difference filter to discard
DVL measurements which indicated a change in speed of more that 0.05 m/s – the vehicle does not have a very high
acceleration.
Our approach to solving this dropout problem involves using the velocity estimates provided by the IMU to
continue the dead reckoning solution when DVL measurements are unavailable. However, since the IMU velocity
4 George Kantor, Nathaniel Fairfield, Dominic Jonak and David Wettergreen
estimates result from integrating measurements from the IMU accelerometers, they are subject to significant drift.
To address this problem, we use a Kalman filter to estimate the drift of the IMU velocity estimates during the times
when the DVL measurements are good (i.e., when the DVL has a lock).
We implement this with two independent and identical Kalman filters. One filter is used to estimate the IMU
velocity and drift in the world frame x direction, the other is used to estimate the same parameters in the y direction.
The state of each filter is q = [v, d]
T
, where v is the world frame vehicle velocity in the relevant direction, and d is
the associated IMU drift term. If we define q
k
to be the value of the state at the time t and q
k+1
to be the value of
the state at the time t + ∆t, then the state transition model that maps (q
k
, a
k
) to q
k+1
is
q
k+1
=
1 ∆t
0 1
q
k
+
1
0
a
k
+ w
k
, F
k
q
k
+ Ga
k
+ w
k
,
where the process noise w
k
is assumed to be white, zero-mean, and Gaussian with covariance matrix Q. The DVL
provides a measurement of the velocity v, so the measurement model is
y
k
=
1 0
q
k
+ ν
k
, H
k
q
k
+ ν
k
,
where the measurement noise ν
k
is assumed to be white, zero-mean, and Gaussian with variance R. From here,
the filter is implemented using the standard Kalman equations (see, e.g., [Maybeck, 1979]), with the small caveat
that prediction steps are executed whenever an IMU measurement is received (about 50 Hz) and update steps are
executed whenever a valid DVL measurement is received (about 4 Hz).
5 Vehicle Controller
The DEPTHX vehicle has a three-level control system that is used to guide the vehicle on its mission. The lowest
level, aptly named the low level control system (LLCS), employs velocity feedback from the DVL and IMU in
order to generate the thrust necessary to track a desired vehicle frame velocity command. The middle level, named
the navigator, issues velocity commands to the LLCS in order to achieve the immediate goal. In this paper, the
immediate goal is to drive the vehicle to a specified waypoint, however the navigator also is capable of executing
more general behaviors such as wall following and obstacle avoidance. At the highest level is the system executive
that, among other things, issues a series waypoint commands to the navigator in order to accomplish the overall
mission.
5.1 Low Level Control System
The LLCS performs two basic functions: it uses velocity feedback to convert a vehicle frame velocity command
into the vehicle frame thrust needed to track that velocity and it implements a mixing table in order to convert the
vehicle frame thrust command into the necessary shaft torque
1
commands to each of the individual thrusters.
Velocity feedback is implemented in four independent loops, one for each of the vehicle’s four degrees of free-
dom. Each loop contains an experimentally tuned PI controller. Note that this structure assumes that the components
of vehicle frame velocity are not coupled by the dynamics of the vehicle, an assumption which not true. In particular,
the forward and sideways velocity components of the vehicle will be highly coupled when the vehicle simultane-
ously rotates and moves in the lateral plane. Hence we can enforce the decoupled assumption by simply avoiding
these type of motions, a restriction which is compatible with the slow, deliberate types of missions that the vehicle
will undertake. In practice, however, the controller performs well even when such coupling motions are executed.
The thrust mixer maps the vehicle frame thrust command vector F
v
= [F
ω
, F
x
, F
y
, F
z
]
T
into a vector of n
individual thruster commands T = [τ
1
, τ
2
, τ
3
, . . . , τ
n
]
T
, where n is the number of thrusters. It is implemented as
a matrix multiplication, i.e., T = M F
v
, where M is computed as follows. First, let A be the 6 × n matrix whose ith
column is given by
A
i
=
p
i
× D
i
D
i
,
1
The relationship between shaft torque and thrust is very nearly linear, and we rely on the DriveBlok
T M
controller produced
by MTS Systems Corp to implement the desired shaft torque on the brushless DC thruster motors.
