.............................................
........................................
AMADEUS is a dexterous subsea robot hand incorporating force and slip
contact sensing, using
fluid
filled tentacles for fingers. Hydraulic pres-
sure variations in each
of
three flexible tubes (bellows) in each finger
create a bending moment, and consequent motion or increase in
contact force during grasping. Such fingers have inherent passive
compliance, no moving parts, and are naturally depth pressure-com-
pensated, making them ideal for reliable use in the deep ocean. In
addition to the mechanical design, development of the hand has also
considered closed loop finger position and force control, coordinated
finger motion for grasping, force and slip sensor development/signal
processing, and reactive world modeling/planning for supervisory
‘blind grasping’. Initially, the application focus is for marine science
tasks, but broader roles in offshore oil and gas, salvage, and military
use
are
foreseen. Phase
I
of the project
is
complete, with the construc-
tion
of
a
first prototype. Phase
I1
is
now underway, to deploy the hand
from an underwater robot arm, and carry out wet trials with users.
Keywords: Robot hands, undersea robotics, AMADEUS
.....................................................................................
n marine geology and benthic science, current practice for
I
sampling rocks, sediment and fauna beyond diver depth is
crude, often relying on grabs, corers and dredgers deployed
from surface vessels. Such techniques are not selective,
imprecise in sample location, disturb the surrounding envi-
ronment during the sample, and usually result in over or
under sampling. The use of Unmanned Underwater Vehicles
(UWs) (Figure
1)
presents the possibility of a cost effective
solution
to
these problems. However, the manipulative abili-
ties of such vehicles are currently primitive, using manipula-
tors with no dexterity or tactile feedback in their end
effectors (Figure
2).
The
MADEUS
project focuses
on
improving the dexterity
and sensory abilities of underwater systems for grasping and
manipulation
of
delicate and other objects. The practical
needs of scientists in the ocean are driving technological
developments in hand and tactile sensor design, position and
contact control systems, and supervisory control of grasping
for operation in poor visibility. In Phase
I
of
the project (com-
pleted in May
1996),
a prototype dexterous three-fingered
underwater dexterous gripper (Figure
6)
was developed,
incorporating force and slip sensors. Techniques for sliding
mode control of finger vibration, task function control of fin-
ger position/contact, and finger coordination have been
34
IEEE
Robotics &Automation Magazine
1070-9932/97/$10.0001997
IEEE
December
1997
Authorized licensed use limited to: Universitat de Barcelona. Downloaded on February 13, 2009 at 04:08 from IEEE Xplore. Restrictions apply.
Figure
1:
Unmanned Underwater Vehicle. Heriot- Watt
ANGUS
002
demonstrated, to allow grasping of objects up to 150”
diameter and 5Kg mass.
A
“blind grasping” mode of supervi-
sory operation has also been developed, using a reactive world
modeling and task planning architecture, utilizing only finger
contact sensing. The initial prototype operates successfully in
the laboratory tank.
Phase
I1
(Figure
3)
is now under way, to develop two work-
cells for wet trials with scientist users. The first will employ a
more rugged dexterous gripper design, mounted on an under-
water robot arm. The second will develop a two-arm system
for coordinated grasping and manipulation of larger or heav-
ier objects (Figure
4).
We will summarize some
of
the achievements of phase
I,
and the prospects for phase
11,
in each of the technology areas.
SYSTEM
ARCHITECTURE
To integrate the various technological developments, a
functional system architecture has been employed (Figure
5).
This architecture provides both the context for each
partner’s activities, and the hierarchy within which hard-
ware and software integration takes place.
At
the lowest level is the dexter-
ous gripper mechanism itself
(HWU-MCE) including actuators.
Sensory information from the dex-
terous gripper (HWU-CEE), is used
at several levels in the hierarchy.
For control (DIST), force data and
estimates of finger position and
velocity are used with strategies
for high bandwidth vibration con-
trol, and control
of
finger position
and force. This low level control is
driven from a medium level cou-
pled control, coordinating finger
movements
for
grasping and
manipulation. This in turn is dri-
ven by grasp planning (HWU-
CEE), which models the observable
Figure
2:
Crude Gripper
of
Typical Underwater Manipulator
tain stable grasps (tele-assistance, supervisory control). Finally,
the human computer interface
(IAN)
provides a graphical user
interface to observe and control the dexterous gripper in both
tele-operation and tele-assistance modes. Control, planning and
HCI are linked bi-directionally, to allow observable failures in
execution to propagate upwards for corrective action.
The architecture is implemented on a mixture of
HP
and PC
UNIX workstations and a multiprocessor VME
68040
system
under the VMEexec real-time operating system (Figure
8).
All
processors are interconnected through ethernet using UNIX
sockets, in addition to the VME bus connection for real-time
boards. Within the project, software interfaces between each
partner’s functional modules were defined, and semi-rigorously
enforced, to ease the final integration task. Of particular impor-
tance was the use of the MATLAFVSIMULINK real-time exten-
sions, to allow rapid prototyping of low and medium level
control system design in the transition from simulation studies
to the real robot. Only a very limited number of additional hand
written C-code drivers for handling inter-board communica-
tions and
VO
were required,
as
well as a few specific routines to
model the finger deflections and Jacobians.
geometry and physical properties
of a grasped object, and reactively
AMADEUS
Phase
111996-1999
plans actions to obtain and main-
Figure SAMMEUS Project Structure
December
1997
IEEE
Robotics &Automation Magazine
35
Authorized licensed use limited to: Universitat de Barcelona. Downloaded on February 13, 2009 at 04:08 from IEEE Xplore. Restrictions apply.
MECHANICAL DESIGN
gripper incorporates knuckle joints. These joints are driven in
A
fundamental innovation in the project has been the concert from a small single central hydraulic cylinder via a
mechanical design of the dexterous gripper mechanism. simple linkage, and provides each finger with
40"
of angular
Existing dexterous hands for use in air generally use articulat- rotation. This knuckle movement is sufficient to enable the
ed joints, with tendon and pulley mechanisms for actuation. finger tips to touch and grasp small objects (diameter
(0)
They are not well suited for use in the ocean, through per- 10") or move apart
ceived difficulties with seals, corrosion, ingress
of
grit and
so
thatlarger objects
water, and hence reliability. The
MADEUS
system, however, (up to 63150") may
utilizes an "elephant's trunk" finger design
[
11,
which has be considered
largely no moving parts, is naturally pressure compensated
A
small hydraulic
for use at depth, and has passive compliance for robustness. cylinder
(O.D.
22mm),
The dexterous gripper consists of three modular sections; with both ports at the
body, finger, and fingertip (Figure
6),
with hydraulic and con- closed end, is used to
trol systems as separate units. The modular approach allows a drive the machined
range of components, materials and geometries to be evaluat- aluminum knuckle
ed in the laboratory without extensive reworking of the entire mechanism. Concen-
dexterous gripper system, and follows sound underwater tric guide bearing
design principles. cylinders located
about the body of
Finger Design the hydraulic cylinder
minimize any side
loads and reduce the
risk of the mechanism
jamming. Extension
of the piston causes
Workcell
The operation of elephant's trunk fingers relies on the elastic
deformation of cylindrical metal bellows with thin convoluted
walls. The convolutions ensure that the assembly is signifi-
cantly stiffer radially than longitudinally, and that longitudinal
extension
is
therefore much greater than radial expansion
when subjected to internal pressure. Currently,
phosphor bronze bellows (14.3" outside diam-
eter
(OD)
and length 125") with 52 active con-
volutions and 0.28mm wall thickness are used.
Each finger is made up from three bellows
placed in a parallel arrangement forming the ver-
tices of an equilateral triangle (pitch center
diameter (PCD) 30"). The proximal end of the
triad
is
attached to the knuckle joint
of
the dex-
terous gripper body, the other to an end-plate,
which connects each bellow to the other two
members of a particular finger. Utilizing a differ-
ent pressure in each bellow creates a range of
extension forces causing the finger to bend
according to the constraints provided by the end
plate (Figure
7).
The larger the differential pres-
sure, the larger the resulting fingertip deflection.
In addition to bending, the triangular arrange-
ment enables the direction of fingertip move-
ment to be controlled.
There
is
a minimum radius
of
curvature
which can be produced by a finger. This radius is
due to the wall thickness, convolution pitch, and
the material used in the bellow actuators. The
larger the maximum deflection required at the
finger tip, the longer the normal length of the
actuator must be.
A
series of plastic belts cov-
ered by neoprene slewing supports the actuators
along their length, reducing the risk of contact
damage or buckling.
Palm
Body
To allow the dexterous gripper to grasp a wide
1
range of object
sizes,
the palm of the dexterous
Figure 5:AMADEUSFunctional System Architecture
36
9
IEEE
Robotics &Automation Magazine
December
1997
Authorized licensed use limited to: Universitat de Barcelona. Downloaded on February 13, 2009 at 04:08 from IEEE Xplore. Restrictions apply.
Figure
6.
MADEUS
Phase IPrototype Dextrous gripper
Overall height (including fingers):
365
mm
Peak finger tip force (straight):
15.45
N
Mass
(of
gripper):
Knuckle movement
+Zoo
(minimum):
910
mm
3.5
kg
Target object dimensions:
SENSOR DESIGN AND SIGNAL PROCESSING
Grasping and manipulation of delicate objects requires
reliable sensing in the finger tips. As a minimum this
should measure the magnitude of any applied force, but
should further include direction for more complex
manipulations. Measurement of slipping may also be of
benefit in reactively maintaining a grasp in the presence
of disturbances.
As
with the hand design, simplicity,
robustness and tolerance to changing pressure and tem-
perature are essential for use in the ocean.
Contact force and slip sensors are currently included
within each fingertip of the dexterous gripper design
(Figure
9),
encapsulated within a compliant silicon rub-
ber compound, using a two-stage injection molding
process. The force sensor uses strain gauges mounted on
a skeleton at appropriate angles. Slip sensing relies on
voltage variations in a piezoelectric material (PVDF) as
slipping causes vibration at its surface.
A
small fifty way push fit connector has been developed
to split sensor feedback at the finger tip interface, allowing
rapid substitution of the finger til, unit as
required.
A
machined
housing
protects
this connector and provides the basic form for the
remainder of the finger tip. Water ingress into the
Finger deflection (maximum):
Maximum frequency response:
20"
(maximum):
5.5
Hz
,0150~~
housing is prevented by a nitrile O-ring seal.
Currently, no sensor is incorporated to mea-
been identified). For closed loop position control, therefore, a
calibrated model of the finger motion is used, driven by pres-
sures measured from sensors within each tube.
the dexterous gripper to flex by movement of a sliding
rotary joint on each of the knuckle manifolds. Another rota-
tional joint on each knuckle manifold is attached to the sta-
tionary outer bearing cylinder of the mechanism.
Polyamide resin plain bearings and slide rings minimize the
friction between all moving surfaces. The angle of the
knuckle joint is measured by a rotary potentiometer driven
from the knuckle mechanism via a rack and pinion. The
potentiometer and drive assembly is housed in a robust oil
filled casing, with quad and O-ring seals to prevent water
ingress or oil leakage.
The knuckle joint alters the dexterous gripper configura-
tion prior to object contact, while the individual finger
motions are reserved for grasping and fine manipulation.
Large changes in position or orientation of grasped objects
must be performed by the arm or wrist onto which the dexter-
ous gripper is mounted.
Hydraulic System
The hydraulic system (Figure
8)
uses a fixed displacement
gear pump with pressure reducing pilot valve to maintain a
system pressure of
30
bar. The pressure in each bellow actua-
tor is controlled using
a
solenoid operated proportional con-
trol valve with spring return, which has a sigmoid shape static
response, with some hysteresis. Due to valve leakage, the min-
gers are usually operated above 12 bar to ensure that
operation is within the central linear portion.
The control valve for the knuckle joint
is
a solenoid operated
three position direction control valve with spring center align-
ment. This enables the flow to the knuckle joint to be either
extended, retracted
or
switched off.
imum
pressuve
which
can
be
delivered
is
7.5
bar
and
the
fin-
Force Sensor Design
An aluminum skeleton approximating the final shape of the
fingertip was constructed, (Figure
lo),
around which the
compliant material is mounted. When stress is applied to the
finger the structure deforms, and these deformations are
measured using an array of
12
strain gauges strategically
mounted on the skeletal structure. From these deformations
it is possible to deduce the force on the fingertip. The choice
of covering is important; if the material is too stiff, there will
be a loss in spatial resolution, whereas
if
it is too compliant
transmission of the forces to the strain gauged elements will
be poor. Currently, the best candidate material for the cover-
ing is a silicone elastomer.
Two quite different methods for determining the forces at
the fingertip from the raw strain gauge readings were devel-
oped. One uses the finite element approach and the theory of
structural stiffnesses, and the other a less mathematically rig-
orous but potentially more accurate approach using a fixed
gain Kalman filter. These methods, including an analysis of
the sensor performance may be found in
[3]
and
[5].
Slip
Sensor
Embedded lmm below the surface of the compliant covering
of the sensor is a thin (52pm) layer of piezoelectric film
(PVDF) (Figure
11).
Piezoelectric film has the property that a
charge is developed at
its
surface when subject to a deforma-
tion. This is a dynamic property in that once it stops deform-
ing, the charge built up on the surface quickly decays to zero.
December 1997
/E€€
Robotics &Automation Magazine 37
Authorized licensed use limited to: Universitat de Barcelona. Downloaded on February 13, 2009 at 04:08 from IEEE Xplore. Restrictions apply.
It is thus ideal for the measurement of transient phenomena
but, on its own, unsuitable for measuring steady state proper-
ties. Since the film develops a (transient) charge at its surface
when deformed, with suitable signal processing
it
has the
ability to act as a vibration pickup.
When used as a slip sensor, the relative motion between
two surfaces causes mechanical vibration in a direction
normal to the plane of motion. The
PVDF
produces a
charge related to these vibrations, and hence can indicate
when a grasped object is slipping. Furthermore, the nature
of the signal detected depends
on
the properties of the
grasped object, the grasp force, the speed at which slip is
distribution, which have remarkable geometric properties.
The first subset is responsible for the motion of the manip-
ulated object, and can be expressed exclusively in terms
of
“tangential motion forces.” This means that the net resultant
wrench applied to the object is due to the superposition of
forces directed along the tangent plane at each contact point.
The second subset has the role of ensuring the feasibility and
the robustness of the grasp, with respect to model uncertain-
ties related with friction coefficients, contact distribution and
so
on. These forces, which typically span a subset of the
so
called “internal forces” space, are basically formed by the nor-
mal forces acting in correspondence with each contact point,
taking place and whether the motion is rotational or trans-
but not causing any motion,
[6].
lational. Analysis
of
both the time
and frequency domain signals
should thus provide further
information about the state of a
grasp and any slippage, with a
possible goal being the direct
control of slippage during dexter-
ous manipulation of an object.
LOW AND
MEDIUM
LEVEL CONTROL
SYSTEM
DESEN
Since the fingers have natural pas-
sive compliance, some grasping
and manipulation tasks can readily
be achieved open loop. However,
to provide more precision in posi-
tioning and applied contact force
(for delicate objects), three
areas of closed loop control have
Pressure
PI
...._-._.._._____..-......_____
(____
Longitudinal Extension
z
2
Fixed End
Resultant Bending
Pressures
Axis
Direction
Y
Radial (Reference Direction)
z
Tangential
Figure
7:
(a) Intemal Pressure Causes Mainly Longitudinal Extension;
(b)
Bending of Flexible Actuator
Caused
By
Internal Pressure Diffwential
been studied:
Low
level
Of
position
and
contact force
*
Medium level coordinated control of fingers for grasping.
Experimental high bandwidth actuation, sensing and con-
trol of finger position.
Low
Level Force
and Position
Control
The low level control module is responsible for positioning
the finger during grasping and manipulation, under the direc-
tion of the medium level controller. Control of contact forces
similarly takes place
here,
Positioning
the
finger during
grasping
is
made difficult since there is currently no position
sensor on each finger, and hence estimates of finger location
must be used, based on bellow pressures and calibrated mod-
els of finger motion. Positioning the finger during manipula-
tion uses the force sensors described in the previous section
on sensor design and signal processing.
The first problem was
to
devise a model for representing, in a
form suitable for both planning and control, the whole set
of
interaction forces acting on the surface of a manipulated object
during generic manipulation operations. To this end, a suitable
and original decomposition has been made, capable of repre-
senting any set of contact forces. In particular,
it
has been found
that there exist
two
subspaces generated by the contact forces
One advantage offered by this kind of decomposition has
been a clear geometrical insight into the structure of the
space of the contact forces during manipulation operations,
and the decoupling of forces responsible for object motion
and the grasp robustness. On the basis of these results, closed
loop robot control algorithms, including iterative learning
techniques (very effective in cases of completely unknown
robot dynamics), have been designed allowing proper control
of object motion and internal forces under the assumption of
proper position and contact forces feedback
171.
Stability and
robustness properties
of
the proposed control schemes, with
respect to possibly unknown robots dynamics were also
assessed. Simulation results also confirmed the effectiveness
of the proposed approach.
The second significant outcome of this phase of the
research program, has been the definition of the general
framework for the design
of
the control architecture for the
AMADEUS
dexterous gripper. In particular it has been neces-
sary to devise a control formulation which could take into
account the peculiarity of the mechanical design of the dex-
terous gripper (the elephant’s trunk design), and on the
other hand allow a “standard” formulation of significant
classes of robotic tasks.
38
IEEE
Robotics &Automation Magazine
December
1997
Authorized licensed use limited to: Universitat de Barcelona. Downloaded on February 13, 2009 at 04:08 from IEEE Xplore. Restrictions apply.