A Novel Type of Compliant and

Underactuated Robotic Hand for

Dexterous Grasping
Raphael Deimel Oliver Brock
August 24, 2015

Abstract
The usefulness and versatility of a robotic end-
effector depends on the diversity of grasps it can
accomplish and also on the complexity of the con-
trol methods required to achieve them. We believe Figure 1: The RBO Hand 2 is a compliant, underac-
that soft hands are able to provide diverse and robust tuated robotic hand, capable of dexterous grasping.
grasping with low control complexity. They possess It is pneumatically actuated and made of silicone rub-
many mechanical degrees of freedom and are able to ber, polyester fibers, and a polyamide scaffold.
implement complex deformations. At the same time,
due to the inherent compliance of soft materials, only
very few of these mechanical degrees have to be con-
trolled explicitly. Soft hands therefore may combine 1 Introduction
the best of both worlds. In this paper, we present
RBO Hand 2, a highly compliant, underactuated, Dexterous grasping is a prerequisite for task-
robust, and dexterous anthropomorphic hand. The dependent manipulation. By the term dexterous, we
hand is inexpensive to manufacture and the morphol- refer to the postural variability of the hand: the
ogy can easily be adapted to specific applications. To higher this variability, the more dexterous we con-
enable efficient hand design, we derive and evaluate sider a hand (for examples of grasping postures, refer
computational models for the mechanical properties to the grasp taxonomies presented in Cutkosky [1989]
of the hand’s basic building blocks, called PneuFlex and Feix et al. [2009]). Such variability enables ver-
actuators. The versatility of RBO Hand 2 is evalu- satile grasping and manipulation: Small objects can
ated by implementing the comprehensive Feix taxon- be picked up with pincer grasps, large objects with
omy of human grasps. The manipulator’s capabilities enveloping power grasps. Depending on the task, a
and limits are demonstrated using the Kapandji test cylindrical side-grasp can be used to pick up a glass
and grasping experiments with a variety of objects of for drinking, whereas a disk grasp from above is ap-
varying weight. Furthermore, we demonstrate that propriate to lift it off a cluttered table.
the effective dimensionality of grasp postures exceeds In robotic hands, dexterous grasping capabilities
the dimensionality of the actuation signals, illustrat- are traditionally realized through complex, multi-
ing that complex grasping behavior can be achieved jointed structures and sophisticated actuation mech-
with relatively simple control. anisms. Such hands are expensive and difficult to

1
design. They also require complex sensing and con- used to evaluate thumb dexterity in human hands af-
trol. Recently, underactuated hands with passively ter surgery. In addition, we show that the hand is
compliant parts have become a popular alternative capable of enacting 31 out of 33 grasp postures of
in robot hand design. These hands perform cer- the human hand from the comprehensive Feix taxon-
tain grasps robustly, have simpler mechanics, and re- omy [Feix et al., 2009]. We evaluate the space of hand
quire simpler control due to underactuation. How- posture exhibited by humans and the RBO Hand 2
ever, there is one commonly assumed drawback of which we find to be similar. We also show that
compliant hands: underactuation and passive com- four actuation degrees of freedom suffice to achieve
pliance seem to render dexterous grasping difficult or a postural space with more than these for dimen-
even impossible. The experiments performed with sions. This implies that the variability in grasping
our novel hand indicate the opposite. posture is only partially generated by the hand’s ac-
We present a novel type of compliant and underac- tuation. The remaining variability is the result of
tuated hand based on soft robotic technology. This interactions between hand and object. These interac-
hand, called RBO Hand 2, is capable of dexterous tions, we claim, are greatly simplified and enriched by
grasping, it is easy to build, robust to unanticipated the extensive use of passive compliance in the hand’s
impact, inherently safe, low-cost, and easy to con- design. These results indicate that dexterous grasp-
trol. These advantages are achieved by building al- ing is easier to achieve with passively compliant than
most the entire hand out of soft, compliant materials with traditional, stiff-linked hands.
or structures, rather than of rigid parts. We believe This paper extends our previous work [Deimel and
that the combination of dexterous grasping capability Brock, 2014] in several important ways. In Sec-
and ease of manufacturing make the presented hand tion 6.3, we extend the analysis of the hand’s dex-
well-suited for enabling novel advances in grasping terity by comparing its postural diversity to that of
and manipulation. the human hand, given a set of diverse grasps. This
Our design, shown in Fig. 1, purposefully maxi- is important, as it provides support for the state-
mizes the hand’s passive compliance, while ensur- ment that compliance enhances dexterity and ensures
ing sufficient structural support to lift objects. We that the RBO Hand 2 has as diverse postures as
believe that this design choice is critical for robust the human counterpart. We also present and vali-
grasping: First, passive compliance facilitates obtain- date a novel model for the actuator’s behavior in Ap-
ing force closure in power grasps [Dollar and Howe, pendix A. This enables the validation of novel actua-
2010, Deimel and Brock, 2013]. Second, passive com- tor designs prior to manufacturing. Furthermore, we
pliance facilitates the use of contact with the envi- significantly extended the description of the hand de-
ronment to aid attaining a grasp. This strategy— sign and production process in Section 4. Finally, we
the exploitation of environmental contact to reduce publish the experimental data sets, videos, and high
uncertainty—has been shown to increase grasp per- resolution images of the human and robotic experi-
formance in humans as well as robots [Deimel et al., ments related to the Feix taxonomy in Multimedia
2013, Kazemi et al., 2014]. For these reasons, passive Extensions 1 to 6.
compliance is a key ingredient for robust grasping. In
this paper, we want to show that in addition to the
two aforementioned advantages, passively compliant 2 Related Work
hands can also perform dexterous grasping. Indeed,
our results indicate that passive compliance even fa- Many highly capable robotic hands exist. A his-
cilitates dexterous grasping. torical overview, surveying robotic hands from over
An opposable thumb is important to achieve dex- five decades, provides an excellent overview [Con-
terity in human and robotic hands. We evaluate the trozzi et al., 2014]. An analysis of robot hand designs
thumb dexterity of the RBO Hand 2 using the Ka- with respect to grasping capabilities was recently pre-
pandji test [Kapandji, 1986]. This test is commonly sented by Grebenstein [2012]. As the notion of com-

2
pliance is central to our hand design, we will limit our is matched by theoretical approaches to analyze and
discussion to hand designs that deliberately include evaluate their dexterity [Prattichizzo et al., 2012,
this concept. Gabiccini et al., 2013]. The most promising mod-
We distinguish between two types of hands, ac- els rely on the hypothesis of a low-dimensional rep-
tively and passively compliant. The former can be resentation of grasp postures, called synergies [San-
achieved by using active control on fully actuated tello et al., 1998]. Odhner et al. [2014] simplify the
or even hyper-actuated systems, where every degree underactuated hand mechanics into compliance el-
of freedom can be controlled. Examples are the im- lipsoids at possible locations of contact points. How-
pressive Awiwi hand [Grebenstein, 2012], the Shad- ever, these approaches require accurate knowledge of
owRobot Shadow Dexterous Hand, and the SimLab grasp posture, contact point locations and contact
Allegro Hand [Bae et al., 2012]. These hands achieve forces. Given current sensor technologies, this infor-
dexterity and compliance through fast and accurate mation is difficult to obtain. Interestingly, humans
control, which comes at the price of mechanical and are able to grasp under comparable conditions with
computational complexity. As a result, these hands strongly impaired perception, e.g. with blurred sight
tend to be mechanically complex and expensive to and wearing a glove [Deimel et al., 2013]. This sug-
manufacture. Mechanical complexity can also in- gests that there is an alternative, perceptually less
crease the probability of hardware failure. demanding representation of compliant behavior.
The alternative is to make hands passively compli- The inclusion of compliance into the design of
ant by including elastic or flexible materials. Build- robotic hands has led to significant improvements in
ing a passively compliant joint is much cheaper than performing power grasps. Very little work has ex-
building an actively controlled one in terms of costs, amined the effect of compliance and underactuation
spatial volume, and mechanical complexity. Passive on the dexterity of a robotic hand. Closing this gap
compliance limits impact forces, a crucial property will be the focus of this paper. Tavakoli et al. [2014]
for an end-effector designed to establish contact with recently characterized the influence of reducing the
the world. More degrees of freedom can better adapt number of actuated degrees of freedom on the num-
to the shape of an object greatly enhances grasp suc- ber of possible grasps for the ISR-SoftHand. This
cess and grasp quality. At the same time, the hand anthropomorphic hand relies on extensive compliance
can be underactuated, effectively offloading control using elastomeric joints and deformable finger pads.
to the physical embodiment of the hand. They found that about four to six actuated degrees
A pioneering work in grasping with passive com- of freedom are enough to enact a broad set of hu-
pliance was the soft gripper by Hirose and Umetani man grasps and that an opposable thumb is crucial
[1978]. Recently, a whole range of grippers and to achieve this, which corroborates our own findings.
hands were built using passive compliance: the FRH-
4 hand [Gaiser et al., 2008], the SDM hand and its
successor [Dollar and Howe, 2008, Ma et al., 2013, 3 PneuFlex Actuators
Odhner et al., 2014], the starfish gripper [Ilievski
et al., 2011], the THE Second Hand and the Pisa- The RBO Hand 2 uses a highly compliant, pneumatic
IIT Soft Hand [Catalano et al., 2014], the ISR- continuum actuator design, called PneuFlex, which
SoftHand [Tavakoli and Almeida, 2014], the Pos- was first presented in Deimel and Brock [2013]. Pneu-
itive Pressure Gripper [Amend et al., 2012], the Flex actuators can be manufactured within a day and
RBO Hand [Deimel and Brock, 2013], and the use materials that are cheap and non-toxic. Fig. 2 il-
Velo Gripper [Ciocarlie et al., 2013]. A different lustrates the working principle. When inflating the
source of inspiration was taken by Giannaccini et al. contained chamber with air the pressure forces the
[2014], who built a compliant gripper inspired by the hull to elongate along the actuator. The bottom
octopus arm. side contains an inelastic fabric to prohibit elonga-
The practical realization of underactuated hands tion. This causes a difference in length between the

3
 (a) An inflated PneuFlex ac- (b) Cut of a PneuFlex continuum actuator. (c) Functional parts
tuator in front of deflated
ones.

Figure 2: Working principle and structure of a PneuFlex actuator: When inflated, the top of the finger,
consisting of translucent silicone, extends, thereby bending as its motion is constrained by the bottom of
the finger, into which an inelastic fabric is embedded. The helically wound threads stabilize the actuator
shape and relieve the rubber of nonfunctional strains, i.e. the inflation leads to bending rather than to radial
expansion.

top and bottom side and the actuator bends. Ra- After unmolding, a silicone tube is inserted at a
dial fibers stabilize the actuator’s shape and greatly convenient position into the top part and bonded to
increase the attainable curvature. it. The tube enables us to easily connect the actuator
to the pneumatic control.
Manufacturing Actuators A distinguishing fea- Afterwards the air chamber is closed by placing the
ture of the PneuFlex actuator is the integrated design top part on a thin (1 − 2 mm) sheet of freshly cast
pipeline, which enables rapid prototyping of actua- silicone that embeds a bendable but inextensible PET
tors with widely varying properties without changing fabric.
the production process. The steps of the process are When the bottom layer has cured, a sewing thread
illustrated in Fig. 3. is wound around the actuator in form of a double
To create an actuator from scratch, first a set of helix. To fixate the thread in place, a thin layer of
planes is defined along a line or curve, on which addition-cure silicone is applied to the top and bot-
the local cross section of the actuator is defined. tom side. This step finishes the actuator.
The principal shape parameters for each cross sec- The presented process enables us to freely change
tion (height, width, hull thickness) are determined width and height of the actuator and the thickness of
from the desired actuation ratio and stiffness at each the rubber hull. We can also create straight as well
point along the actuator, e.g. by using the model pro- as curved actuators, as long as the bottom layer stays
vided in Appendix A. in a plane.
In the second step, the set of cross sections is trans- The actuator design space can be explored by vary-
lated into a 3D model of a two-part mold for casting ing shape and size of the actuator, and by varying the
the rubber body of the actuator. The model is pro- thickness of the silicone hull at the top, side and bot-
duced on a 3D printer. Because we need to separate tom. Additionally, available silicone types let us vary
the mold from the cast, the bottom side of the actu- the shear modulus by an order of magnitude. All of
ator is not included. these parameters affect the bending behavior, stiff-
In the next step, the top part is cast using the ness, and limits of the actuator. If needed, a bellows-
printed mold and addition-cure silicone. shaped hull extends the design space towards larger

4
curvatures and lower stiffnesses. Curved actuators
can realize three-dimensional motion. For example,
the two palm actuators (actuators 6 and 7 in Fig. 4)
are used to achieve thumb opposition. Differential in-
flation of the two actuators provides additional dex-
terity.
Actuators can also be packaged together, similarly
to muscle fibers, allowing for redundant actuation or
variation of the actuation strength. Joining actuators
also enables the implementation of multiple deforma-
tion modes or the deliberate mixing of deformation
modes [Bishop-Moser et al., 2012].
Besides their flexible production process, PneuFlex
actuators are robust to impact and blunt collisions,
are inherently safe, and are not affected by dirt, dust,
or liquids. However, they can easily be cut or pierced.

Modeling Actuators The PneuFlex actuator Figure 4: The seven actuators of the soft anthropo-
shares many properties with other recently published morphic hand: four fingers (1–4), thumb (5), and the
continuum actuators, most notably the fast PneuNet palm, consisting of two actuators (6, 7)
actuators [Mosadegh et al., 2014] and the actuator
by Galloway et al. [2013]. In contrast to these actua-
tor designs, the PneuFlex design and production pro- achieved, if necessary, as we will explain in Section 7.
cess is optimized to provide freely adaptable cross sec-
tions which determine actuation ratio and stiffness, Morphology The design space of possible hands
the ability to include multiple separate air chambers, is very large. For this hand, we chose an anthropo-
and to provide access to its internal space for the morphic design in shape and size for three reasons.
integration of sensors and wiring. First, we know the human hand form enables dexter-
We provide a detailed analysis of basic actuator be- ous grasping in humans. By starting with a human-
havior in Appendix A and also propose design rules hand-like morphology, we start with a proven hand
for successful actuator design in Section A.11. Ad- design. Second, many objects have been built for
ditional insights can be drawn from related research manipulation by a human hand and match the an-
on continuum actuators. For example, Bishop-Moser thropomorphic form factor. Third, we can use well-
et al. [2012] characterize all basic motions attainable established grasp taxonomies and compare our de-
by changing inclinations of the reinforcement helices. signs with humans and many other robotic hands.
Others proposed approximate numeric models based
on twisted, one dimensional beams [Renda et al.,
2012, Giorelli et al., 2012]. Control Pneumatic control of the PneuFlex actu-
ators is based on a simple linear forward model for
computing valve opening times. The model takes into
4 Hand Design account the regulated supply pressure to achieve a de-
sired channel pressure which corresponds to a desired
In this section, we describe the components of our soft bending radius or contact force. Alternatives to this
anthropomorphic hand (RBO Hand 2, see Fig. 1). digital control are cylinder-based continuous control
The entire hand weighs 178 g and can carry a pay- systems [Marchese et al., 2014]. Renda et al. [2012]
load of up to 0.5 kg. Higher payload can easily be demonstrate a computationally simple forward model

5
Figure 3: Production steps for making a PneuFlex actuator with a custom stiffness and actuation ratio
profile.

scribed by the local curvature κ and torsional stiff-

ness M around the bottom side for short segments
of the actuator and is determined by the geometry of
the actuator cross section (see Fig. 5). Appendix A
contains an analysis of this simplified model of the
actuator, which provides surprisingly simple rules to
Figure 5: Illustration of the finger geometry and its design the ratio of curvature κ to pressure p at each
principal parameters: κ (local curvature), M (mo- segment along the actuator by varying the hull thick-
ment around the passive layer at the interface), d ness d:
(top side rubber thickness) and c (circumference)
∆κ 1
≈ ,
∆p G·d
for an artificial octopus arm. This model can also be where G is the shear modulus of the rubber and con-
used for PneuFlex actuators. stant within an actuator.
For the experiments, control was implemented with Translational forces between actuator segments are
industry-grade air valves and a separate air supply. mainly transmitted by the inelastic fabric of the bot-
For a mobile system, the control system can easily be tom layer and therefore don’t need to be considered.
optimized for size and weight, because the required The model in Appendix A also yields a rule of thumb
air flows are an order of magnitude smaller than pro- for the torsional stiffness of an actuator segment.
vided by industrial grade valves. For the same reason, For an approximately squared cross section, stiffness
pressurized air can be effectively sourced by small scales with circumference c:
compressors and small tanks. In systems where elec-
∆M
tric energy is scarce, high-pressure tanks can provide ≈ G · d · c3 .
storage of air with high energy density [Wehner et al., ∆κ
2014]. As an example, for the fingers of the hand we chose
to increase stiffness linearly from tip to base and keep
the actuation ratio constant. Such a profile has also
Fingers The five fingers of our hand are single been used in the Soft Gripper [Hirose and Umetani,
PneuFlex actuators (see Fig. 4). The index, middle, 1978]. Using x as the distance from the tip along the
ring, and little finger are 90 mm long and of identical actuator, the cross section at this point is defined by:
shape, the thumb actuator is 70 mm long. All fingers
get narrower and flatter towards the finger tip. By d(x) = const,
using actuators as fingers, we can exploit the excel- c(x) ∝ x 3 .
1

lent compliance and robustness of the actuators and

greatly simplify the design. The resulting geometry of the actuator is illus-
The mechanical behavior of the finger can be de- trated in Fig. 5. According to our model, the aspect

6
ratio of the cross section does not strongly influence
stiffness and actuation ratio. So to give the finger a
visually more appealing shape, we set set the width
to:
1
width(x) ∝ x 8 .

Palm A key feature of the human hand is the op-

posable thumb. We realize it in our hand by actu-
ating the palm (see Fig. 4). The palmar actuator
compound consists of two connected actuators. Its
base shape is a circular section of 90◦ with 78 mm
outer and 25 mm inner radius. The actuator curves
perpendicular to the passive layer. The stiffness as
well as the actuation ratio remain constant along the
curved actuator. They are also designed to be twice
as stiff as the fingers to account for the fact that the
two actuators in the palm oppose four fingers. Fig. 8
provides an impression of the possible thumb motions
when the two palm actuators are inflated either to- Figure 6: Difference in thumb configuration and fin-
gether or differentially. gertip use during a pincer grasp between a human
In addition to enabling thumb opposition, the palm hand and the robotic hand
also provides a compliant surface that, together with
the fingers, is used to enclose objects in various power
grasps. To augment this function, the fingers and the
palmar actuator are connected by a thin sheet of fiber negative curvatures. As both sides of the PneuFlex
reinforced silicone, covering the gap between palm actuator have similar surface characteristics (unlike
actuators and fingers (shown in Fig. 1, but absent in human thumbs), this choice will not affect grasp qual-
in Fig. 4). This sheet transmits tensile forces between ity.
fingers and palm, and between adjacent fingers. This
stabilizes the underlying scaffold during power grasps Scaffold The fingers and the palm are connected
or for heavy loads, as shown in Fig. 11. to the wrist by individual, flexible struts as part of
a 3-D printed polyamide scaffold (2 mm thick, see
Thumb Like the other fingers, the thumb consists Fig. 1). The intentionally flat cross section of the
of a single PneuFlex actuator. The thumb is shorter struts enables deformation modes, such as arching
and twice as stiff, but also features a linear stiffness the palm and spreading the fingers. Space for the re-
profile. A faithful imitation of how humans use their spective actuator is provisioned, but was not added
thumb would require a negative curvature close to to the hand described here. The struts decouple dis-
the tip, as shown in Fig. 6, and would significantly placement between fingers, further increasing passive
increase complexity of manufacturing the thumb. We compliance of the hand. The flexibility of the struts
therefore deviated here from the human hand. In- limits impact forces, while providing sufficient stiff-
stead of the inside of the thumb, we use the backside ness for heavy payloads without excessive deforma-
(dorsal side) as the primary contact surface for pincer tion (see Fig. 11).
grasps. This effectively changes the contact surface The fingers and the palmar actuator compound are
orientation by about 45–60◦ , relative to the orienta- bonded to the supporting scaffold as shown in Fig. 1.
tion found in a human thumb, avoiding the need for The palm is supported by parts of the scaffold to

7
increase its torsional stiffness during opposition with
the fingers.

Strength between thumb and fingers A flexi-

ble, but inextensible band connects the base of the in-
dex finger to that of the thumb (see Fig. 4). Similarly
to a muscle in human hands (adductor Pollicis), it
enables increased contact forces between thumb and
opposing finger, by reducing torques on the struts
at the wrist. The sheet connecting fingers and palm
serves a similar role, especially for power grasps of Figure 7: The Kapandji test counts the number of
cylindrical objects with large diameter. indicated locations that can be contacted with the
thumb tip.

5 Grasp Dexterity
In this section, we evaluate the dexterous grasping sition the thumb as desired. The six most important
capabilities of the proposed hand. The most appro- postures of the hand performing the test are shown
priate evaluation would of course be in full-fledged, in Fig. 8. The thumb tip could reach all but one lo-
real-world grasping experiments. However, this re- cation. Location 1 was not possible to reach because
quires the integration of hand and control with per- it would require a backwards bending of actuator 5
ception and grasp planning and would effectively be (thumb). Still, the hand scores seven out of eight
an evaluation of the integrated system. Here, we fo- points, indicating a high thumb dexterity.
cus on evaluating the capabilities offered by the hand.
Furthermore, we have to resort to empirical methods. 5.2 Grasp Postures
Accurate simulation of the complex, nonlinear defor-
mations encountered in such a heterogeneous and soft A common way of assessing the dexterous grasping
structure is difficult to conduct and anyways requires capabilities of hands is to demonstrate grasps for a set
empirical experiments to validate the results. of objects. For example, the THE Second Hand was
evaluated with four objects and two grasp types [Gri-
oli et al., 2012], the SDM hand on ten objects and
5.1 Thumb Dexterity
a single grasp type [Dollar and Howe, 2008], the
Medical doctors employ the Kapandji test [Kapandji, Velo Gripper on twelve objects and a single grasp
1986] to assess thumb dexterity during rehabilitation type [Ciocarlie et al., 2013], and the Awiwi hand on
after injuries or surgery. This test was also used by eight objects and 16 grasp types [Grebenstein, 2012].
Grebenstein for evaluating and improving the thumb We follow these examples in our evaluation.
dexterity of the Awiwi hand [Grebenstein, 2012]. For We select grasp types and objects based on the
the Kapandji test, the human subject has to touch a most comprehensive grasp taxonomy to date, the Feix
set of easily identifiable locations on the fingers with taxonomy [Feix et al., 2009]. It covers the grasps most
the tip of the thumb. These locations are shown in commonly observed in humans and therefore is a real-
Fig. 7. The total number of reachable locations serves istic reference for assessing the dexterity necessary for
as an indicator of overall thumb dexterity. A thumb common grasping tasks. The taxonomy encompasses
is considered fully functional if it is able to reach all 33 grasp types, out of which the first 17 are identical
locations. to the grasps in the Cutkosky taxonomy [Cutkosky,
To perform the Kapandji test on our hand, we 1989]. To demonstrate these 33 grasps, the original
manually selected actuation pressures that would po- publication illustrates 17 different object shapes [Feix

8
Figure 8: The RBO Hand 2 succeeds in the Kapandji test for all but one position (position 1, lower right,
showing best effort).

5: Light Tool reference

B C D
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4 A
0.2 0.2 0.2
0.0 0.0 0.0
0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8

(a) Failed Light Tool grasp: no force clo- 0.8 0.8

sure 0.6 0.6

0.4 0.4 B
19: Distal Type reference
0.2 0.2
0.0 0.0
0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8

0.8
0.6
0.4 C
0.2

(b) Failed Distal Type grasp: functionally 0.0

0.0 0.2 0.4 0.6 0.8
defective
Figure 10: Scatter plots of the four actuation chan-
Figure 9: Grasping postures not successfully attained nels for the actuation patterns of the 31 successful
by the robotic hand grasps. Darker color indicates overlapping dots.

9
et al., 2009]. We therefore used 17 objects and 33
grasp types to evaluate our hand.
We implemented the grasps from the Feix taxon-
0.54 kg
omy by defining appropriate actuation pressures and
actuation sequences. When, due to collisions, simul-
taneous actuation of all channels was not sufficient
to reach the desired posture, we added an appro-
priate pre-grasp posture. The commanded actuation
pattern was then modified and tested iteratively to
improve the quality of the grasp in terms of grasp
stability and robustness against external forces, and (a) Cylinder off center
8N
to ensure the proper types and locations of contact.
Grasp quality was judged by manually rotating and
translating the hand, and by testing several repeti-
tions of the actuation pattern.
To simplify the search for appropriate actuation
patterns, we combined the control of the seven actu-
ators into four actuation channels. Channel A drives
actuators 1, 2, and 3 (small, ring, and middle fingers), (b) Tolerated disturbance
channel B drives actuator 4 (index finger), channel C 7N

drives actuators 5 (thumb) and 7 (inner palm), and

channel D controls actuator 6 (outer palm). These
channels can be understood as the hand’s four grasp-
ing synergies.
To perform a grasping experiment for a particu-
lar grasp type, the experimenter triggers the actua-
tion sequence to attain the pre-grasp posture, holds
the object in the seemingly most appropriate location (c) Tolerated disturbance
relative to the hand, and then triggers the actuation 6N
sequence for the grasping motion. The resulting pos-
tures for each empirical actuation pattern are shown
in Fig. 12; high resolution images are provided in Ex-
tension 1. Out of 33 grasp types, the hand is able to
perform 31 repeatably (three consecutive successful
trials). The two grasps that failed are the light tool 1.65 kg
grasp and the distal type grasp. (d) Tolerated disturbance (e) Support strength
The light tool grasp fails because the hand does
not possess finger pulp that fills the cavity formed by
the maximally bent fingers, which causes the object Figure 11: Illustrations of grasping force capabilities:
to slip. The distal type grasp fails because the result- (a) finger strength and palm support strength, (b)–
ing grasp is nonfunctional with respect to proper use (d) tolerated disturbance forces in different directions
of the scissors, even though it is possible to put the for grasp 1, and (e) strength of the support provided
soft fingers through the scissors’ holes. Both grasp by the scaffold
failures are shown in Fig. 9.
Fig. 10 shows a scatter plot of the actuation pat-
terns for the 31 successfully achieved grasp types of

10
the Feix taxonomy. The actuation patterns relate to 6 Compliance Benefits Dexter-
final grasps, not pre-grasp postures. The plots indi-
cate an even distribution of activation for all channels
ous Grasping
and do not reveal obvious correlations that could be In the previous section, we showed that our under-
leveraged to further simplify actuation. actuated and compliant hand is capable of dexterous
The evaluation presented in this section demon- grasping. In this section we will investigate whether
strates the hand’s ability to assume a variety of grasp its compliance and underactuation are beneficial or
postures. This ability is comparable with that of detrimental to attaining different grasp postures. If
other hands presented in the literature. We therefore beneficial, control will be simpler than the result-
believe that dexterous grasping and compliance can ing behavior, i.e. actuation space smaller is smaller
indeed be combined in a highly capable, compliant, than grasp posture space. The dimensionality of the
underactuated robotic hand. posture space that exceeds the dimensionality of the
actuation space can be explained by the compliant
5.3 Grasping Forces interactions between hand and object.
While grasp quality and grasp strength was not the
driving design criterion for the hand, it is important 6.1 Postural Diversity of the Feix Tax-
to verify that a compliant hand is capable of lifting onomy
objects of reasonable weight. To give the reader an
To assess the dimensionality of the attainable grasp
intuition on the capabilities of the hand, we provide
posture space, we first have to assert that the grasp
a few tests regarding grasping forces.
set that we use to sample from that space is diverse
The heaviest objects used in the Feix taxonomy
enough, i.e. that the employed grasps span the space
grasps were the rectangular plate in grasp 22 (156 g),
of possible grasps. For this, we recorded humans do-
the metal disc in grasp 10 (181 g), the wooden ball
ing Feix taxonomy grasps using the method published
in grasp 26 (183 g), and the circular plate in grasp 30
by Santello et al. [1998], and compare the results
(240 g). Note that in grasps 26 and 30, the shown
to existing published data sets (the data published
posture offers the least structural support of possi-
in Santello et al. [1998], the UNIPI data set1 , and
ble hand poses. Fig. 11 shows two additional heavy
UNIPI-ASU data set2 .
objects, a wooden cylinder (541 g) and a lead ball
In the experiment, we asked five healthy human
(1.650 g). Fig. 11 also shows three different di-
participants to enact every grasp of the Feix taxon-
rectional disturbance forces on a cylinder which is
omy five times while wearing a Cyberglove II data
power-grasped with grasp 1. If forces above 6-8 N
glove, using exactly the same objects as used for
are applied, the cylinder will slide in the hand.
the experiment in the previous section. Participants
were allowed to use the other hand to assist in as-
5.4 Grasping in Realistic Settings suming the grasp posture, but had to achieve a suc-
To further illustrate the effectiveness of the pro- cessful grasp in the sensorized hand without addi-
posed hand, we performed experiments with com- tional support. The resulting postures were sam-
plete grasping sequences, shown in Fig. 13 and in pled 50 times within 500 ms and averaged over sam-
Extension 2. In these experiments, a human operator ples and episodes. We then performed dimensionality
selects the appropriate grasp, triggers the pre-grasp reduction by applying principal component analysis
posture of the hand, places the hand in the appropri- (PCA) for each subject individually to exclude inter-
ate location, and then executes the grasp. These ex- subject variance in accordance with the data analysis
periments demonstrate that the proposed hand, given used in Santello et al. [1998]. We then compared the
appropriate perception and grasp planning skills, is 1 Data set available at http://handcorpus.org
able to perform real-world grasps. 2 Data set available at http://handcorpus.org

11
 1: Large Diam- 2: Small Diam- 3: Medium 4: Adducted 5: Light Tool
eter eter Wrap Thumb (failed)

6: Prismatic 4 7: Prismatic 3 8: Prismatic 2 9: Palmar 10: Power Disk

Finger Finger Finger Pinch

11: Power 12: Precision 13: Precision 14: Tripod 15: Fixed Hook
Sphere Disk Sphere

16: Lateral 17: Index F. 18: Extension 19: Distal 20: Writing
Ext. Type (failed) Tripod

21: Tripod 22: Parallel 23: Adduction 24: Tip Pinch 25: Lateral
Variation Ext. Grip Tripod

26: Sphere 4 27: Quadpod 28: Sphere 3 29: Stick 30: Palmar
Finger Finger

31: Ring 32: Ventral 33: Inferior

Pincer

Figure 12: Enacted grasps of the Feix taxonomy, using empirically determined actuation patterns: Grasps
are numbered according to the Feix taxonomy [Feix et al., 2009]; the hand failed to replicate grasps 5 (Light
Tool) and 19 (Distal Type, Scissors)
12
Figure 13: Performing grasps using the grasp postures 25, 1, 9, 28 and 18: a human places the hand and then
triggers the actuation of the appropriate grasp; top: pre-grasp posture, middle: executed grasp, bottom:
lifting object to show success.

resulting residual unexplained variances with data 6.2 Choice of Representation

from literature.
Because the RBO Hand 2 does not have discrete
joints, it is not possible to assess its postural diver-
The results are shown in Fig. 14. For four out
sity using the method of Santello et al. [1998], even
of five participants, the unexplained variances were
though the hand is anthropomorphic. An alternative
higher than those of the three independently pub-
representation of hand posture compatible with soft
lished data sets, suggesting that the grasps span the
hand mechanics has been used for the Human Grasp
space of possible grasp postures more effectively than
Database3 experiment. In this experiment, Romero
the considerably larger but less structured set of ob-
et al. [2010] represented hand posture in terms of fin-
jects that was used for the data sets we compare with.
gertip position and orientation relative to the back of
the hand.
The discrepancy between the published data sets
Before using this representation, we first have to es-
and ours may be explained by the fact that the San-
tablish that for assessing postural diversity it is com-
tello and UNIPI data sets were recorded on grasping
parable to using joint angles. Fig. 15 shows a compar-
imagined objects, while our data and the UNIPI-ASU
ison of the residual unexplained variances of a Prin-
data set are recorded while grasping real objects.
cipal Component Analysis for fingertip position and
The additional variance observed may come from the
joint angles (from the previous experiment), both ac-
interaction between hand and object, whereas with
quired on human hands using the Feix taxonomy (31
imagined objects, hand posture may be more related
grasps, excluding grasps 19 and 23 in Fig 12). The
to actuation pattern than the actual grasp posture.
graph shows that fingertip positions may be a more
This interpretation of existing data is consistent with
compact representation than joint angles, but most
our hypothesis that hand control can be simpler than
effected posture. 3 Data set available at http://grasp.xief.net/

13
 0.8
participant 1
participant 2
0.7 participant 3
participant 4

Residual unexplained variance

0.6 participant 5
Human Grasp Database
(fingertip positions)
0.5

0.4

0.3
0.8
participant 1
participant 2
0.2
0.7 participant 3
participant 4 0.1
Residual unexplained variance

0.6 participant 5
Dataset in Santello
(1998) 0.0
0.5 Dataset UNIPI-ASU 1 2 3 4 5 6
(May 2011) Principal components
Dataset UNIPI
0.4 (Oct. 2011)
Figure 15: Residual variances of PCA on grasp pos-
tures enacted by humans using data from different
0.3
acquisition methods, joint angles and fingertip posi-
0.2
tions. Solid line denotes data from the Human Grasp
Database [Romero et al., 2010], dotted lines denote
0.1 the data acquired using the Feix taxonomy and the
objects shown in Fig. 12
0.0
1 2 3 4 5 6
Principal components

Figure 14: Residual variances of PCA on joint mea- importantly not worse. Therefore we can use finger-
surements taken with data gloves. Solid lines denote tip position data as a conservative estimate of joint
published data sets using the method of Santello et al. posture posture diversity. Omitting fingertip orien-
[1998], dotted lines denote the data acquired using tation data (represented as quaternions) from the
the Feix taxonomy and the objects shown in Fig. 12. Principal Component Analysis did not significantly
decrease unexplained variance. It was therefore ex-
cluded to simplify acquisition of comparable data on
the RBO Hand 2 using a motion capture system.

Concluding the human experiments and compara-

tive study of published data, the Feix taxonomy ap-
pears to be a good proxy for diverse grasp postures.
Additionally, we found that fingertip position is a vi-
able alternative to joint angles for assessing postural
diversity of human hands and therefore also of an-
thropomorphic robot hands.

14
 0.8
Human Grasp Database
(fingertip positions)
0.7 RBO Hand 2
(fingertip positions)

Residual unexplained variance

RBO Hand 2
0.6 (actuation signal)

0.5

0.4

0.3

0.2
Figure 16: Screen shot from video Extension 4, show-
ing the marker placement on the fingertips for Motion 0.1
Capture.
0.0
1 2 3 4 5 6
Principal components

6.3 Postural Diversity of RBO Hand 2 Figure 17: Dimensionality of fingertip positions for
a human hand, the RBO Hand 2, and its respective
The previous subsection gives us a justification to use control signal. The gray region indicates where the
fingertip positions of grasps from the Feix taxonomy dimensionality of fingertip posture exceeds control di-
to assess the RBO Hand 2 in terms of grasp posture mensionality.
diversity. To acquire the fingertip positions we used
a motion capture system. We placed 3 mm retrore-
flective markers on the backside of the fingertips of A second observation can be made by considering
the four fingers. On the thumb tip, the marker was the dimensionality of control required to implement
placed on the side to not interfere with the pincer the grasp postures, which is indicated by the dotted
grasp. The marker placement can be seen in Fig. 16 line and gray area in Fig. 17. The space of grasp pos-
as in the video Extension 4. Note that additional tures is of higher dimensionality than the actuation
markers were attached to the hand during recording, space, which is of dimension four. Where does this
but only the five fingertip markers were used for the increase come from? It must be introduced by the
experiment. diverse shapes of the grasped objects. The interac-
The motion capture system has a spatial precision tions between the hand and the object differentiate
of 0.05 mm, as measured using 205 frames recorded different postures. This differentiation is facilitated
for a stationary hand (at 50 Hz). For each grasp, two by the hand’s ability to adapt to objects compliantly.
trials are recorded, fingertip positions are extracted The differences between imagined and real grasps in
and averaged over the trials before applying Principal human experiments, discussed in Section 6.1, corrob-
Component Analysis. The data set from the Human orate this hypothesis. Tavakoli et al. [2014] also ob-
Grasp Database is subjected to the same procedure. tain the minimum of four to six actuated degrees of
The results are shown in Fig. 17. The postural di- freedom to implement dexterous grasping for their
mensionality exhibited by the RBO Hand 2 matches anthropomorphic, compliant hand.
the human hand data well. This strongly supports The diversity and consistency of evidence we pre-
our claim that our robotic hand is as dexterous as a sented here strongly suggests that compliant hands
human hand with respect to attainable grasp posture, benefit dexterous grasping.
taking the Feix taxonomy as a reference. The presented evaluation, based on published data

15
sets, measurements of human grasps, and of grasps points between object and hand. It also reduces the
with the RBO Hand 2, paints a consistent picture: probability of slip when objects collide or upon jerky
The robotic hand performs similar to a human hand, wrist motion. At the same time, it also places a limit
both in terms of grasp posture diversity, and in terms on the forces the hand can exert on the environment
of covering sets of grasps, again, using the Feix tax- through the grasped object, for example. This nega-
onomy as a reference. tively affects tasks where those forces must be high.
This diversity in grasping posture is achieved by We therefore need to balance and these two compet-
the RBO Hand 2 with only four actuated degrees ing properties.
of freedom. This is possible because the interaction In case grasp stiffness proves to be the limiting
of the compliant degrees of freedom with the diverse factor for certain tasks, there are several methods
objects introduces additional variance in the posture. available to selectively increase actuator stiffness [V.
et al., 2015]. But they increase design complexity and
production costs, and therefore should be avoided if
7 Limitations possible. Another simple method to increase grasp
stiffness is to select power grasps instead of precision
Adopting a novel technology in a new application,
grasps.
like continuum actuators in the design of soft dexter-
In the context of grasp stiffness, we can understand
ous hands, opens up new possibilities but also leads
the exploration of soft hand designs as searching for
to new limitations and challenges that need to be
a lower bound on grasp stiffness that still is able to
considered carefully.
provide sufficient grasp force. Future research will
investigate how grasp stiffness can be increased while
Grasp Forces and Payload Continuum actua- maintaining compliance where necessary.
tors, when constructed with reinforced rubber and
actuated hydraulically, are in principle capable of ex-
erting extremely large forces. For example, an ac- Pneumatics Our hand relies on additional exter-
tuator made out of car tire rubber with steel-fiber nal pneumatic components for control. These com-
reinforcements and hydraulic actuation would proba- ponents are cheap and readily available in industry-
bly be able to exert grasping forces exceeding 100 N. grade quality. However, they are are over-sized for
It would also be straightforward to make a much the low pressures, small volumetric flow, and size con-
stronger hand with the current production process straints of robotic hands. Miniaturizing and integrat-
by choosing stiffer rubbers and thicker hulls. In the ing electrically actuated valves directly into the hand,
current hand design, we chose to use very soft actu- possibly even into the actuator, would greatly sim-
ators, to investigate the effect of compliance, to in- plify integration into predominantly electromechanic
crease safety, and to make manufacturing convenient. robots.
We chose pneumatic actuation over other fluidic op- Long-term autonomy arguably is easier to achieve
tions as it is much simpler and cleaner to operate in with pneumatic systems. In contrast to electrical
a lab environment. power systems, where no good solution for long-term
untethered operation exists, the technology exists to
Grasp Stiffness While grasp forces, as discussed make small, quickly refillable air tanks or even to di-
above, refer to the magnitude of forces exerted on rectly convert chemical energy [Wehner et al., 2014].
the object, grasp stiffness refers to the hand’s abil- Mobility can easily be obtained by using compact,
ity to maintain a grasp posture in the presence of small compressors, as the average rate of airflow for
external forces. Naturally, a soft hand built for max- operating the RBO Hand 2 is very low and peaks can
imum compliance does not create an extraordinarily be serviced by small air tanks. The use of electrical
stiff grasp. Low grasp stiffness has the advantage of energy storage also often simplifies the integration
reducing the peak forces encountered at the contact into existing system.

16
Precision and Repeatability of grasps While key mechanical properties of PneuFlex actuators in
actuation patterns can be reproduced with high pre- Appendix A. While that model is not able to es-
cision, the interaction with the object or features of timate all parameters necessary to create a faithful
the environment during a grasp can introduce sub- simulation, it can be used to straightforwardly cus-
stantial variations in the final grasp posture. While tomize deformation modes and stiffness profiles of the
this is often understood to be a disadvantage, we view actuators before building them. The Appendix also
it as an important feature of the design, leading to evaluates the influence of several nonlinear phenom-
robust grasping performance. Therefore, we believe ena on actuator behavior and suggests several, easy
one must carefully differentiate between precision and to follow design guidelines to avoid common failure
repeatability versus grasp robustness. The hand pre- modes.
sented in this paper deliberately trades the former for Yet another consequence of the complex and highly
the latter. variable deformations that happen during grasping
is that most existing grasp planners cannot be ap-
Sensing The compliance of the materials makes plied, as they rely on complete and accurate geomet-
local, dense sensing for proprioception and contact ric and kinematic models of hand and environment.
forces important but also very difficult. Also, sensor While it might be possible to perform simulations
technologies compatible with soft actuators are cur- of the hand’s deformation using finite element meth-
rently not available commercially. While it is easy to ods, these are computationally too complex to em-
integrate air pressure sensors, it would be very desir- ploy them in search-based grasp planning.
able to integrate strain and touch sensors too. This is
a topic of active research as current electronic sensor
technology is predominantly designed for rigid struc- 8 Conclusions
tures, but working solutions for stretchable electron-
We presented a compliant, underactuated, and dex-
ics start to appear, making an integration with Pneu-
terous anthropomorphic robotic hand based on soft
Flex actuators feasible in the near future [Rahimi
robotics technology. The hand is able to achieve 31
et al., 2014, Culha et al., 2014, Gerratt et al., 2014].
of 33 grasp postures from a state-of-the-art human
grasp taxonomy. To evaluate the dexterity of the op-
Modeling Hand Mechanics Modeling the whole posable thumb, we performed the Kapandji test, in
hand poses certain challenges: an accurate mechan- which the hand achieves seven out of eight possible
ical state is difficult to obtain due to nonlinearities points. We illustrated the hand’s excellent payload to
arising from large deformation and anisotropic struc- weight ratio, as it is able to lift objects of nearly three
ture of its components. Additionally, the actuators times its own weight. We also presented real-world
intentionally provide a large number of deformation grasping experiments to demonstrate the hand’s ca-
modes, which increases hand complexity even further pabilities in a realistic setting.
and makes sensing the hand’s complete mechanical We believe that compliance is crucial to enable ro-
state very difficult. bust grasping in robotic hands. We provided support
Because of these features, we are currently not able for this statement by showing that the dimensional-
to provide a quantitative analysis of grasp quality ity of the achievable postural space is significantly
based on mechanical models as it is state of the art larger than the dimensionality of the hand’s actu-
for hands with rigid links. We therefore qualitatively ation space. We explain this observation with the
assess grasp quality in Section 5.3 and also provide hand’s ability to mechanically comply to the shape
two videos in Multimedia Extension 5 and 6 that il- of the grasped object: The final grasping posture is
lustrate the attainable quality of grasps. the result of the hand’s actuation together with com-
To start closing the gap in modeling soft contin- pliant interactions between the hand and the object.
uum actuators, we present a model for computing We found that several grasping experiments with hu-

17
mans are consistent with this interpretation. We
therefore conclude that compliance in robotic hands,
when used correctly, can facilitate not only robust-
ness in power grasps but also dexterity.
In addition to enhancing dexterity, the use of soft
robotic technology renders the hand robust to impact
and blunt collisions and makes it inherently safe and
suitable for working environments containing dirt,
dust, or liquids. The effort, complexity, and cost of
building the hand are significantly lower than for ex- Figure 18: Simple model of an actuator segment
isting hand technologies. The hand presented here
can be built in two days, using materials worth less
than 100 US$. Both actuator and hand structure
are easily adaptable to specific application domains.
A.1 Formalization of the PneuFlex
We therefore believe that this novel way of building Actuator Geometry
robotic hands significantly lowers the barrier to entry Fig. 18 shows the parameterization of a small seg-
in the field of grasping and manipulation research. ment of the actuator. To simplify the model, we ig-
nore the side walls and assume a rectangular cross
section. Let x, z, d be the length, width, and thick-
9 Acknowledgments ness of a segment of the actuator respectively.
Width and height of the actuator are assumed to
We gratefully acknowledge financial support by the not change due to the helical thread around the ac-
Alexander von Humboldt foundation through an tuator. This assumption is facilitated by selecting an
Alexander von Humboldt professorship (funded by approximately circular cross section (e.g. a square)
the German Federal Ministry of Education and Re- for the shape of actuator cross section, as the radial
search). We are equally grateful for the funding pro- fibers will always deform the shape into circle to bal-
vided by the SOMA project (European Commission, ance the radial pressure.
H2020-ICT-645599) and the German Research Foun- Due to its embedded fabric, the bottom layer also
dation (DFG, award number BR 2248/3-1). has a fixed length x. The only possible deformation
left is to stretch the silicone layer while bending the
bottom layer, which is illustrated in Fig. 19
The interfaces of the segment are rotated to each
A Derivation of an Approxi- other by the angle ϕ and the bottom layer curves with
mate PneuFlex Model the radius r = ϕx . As we also have configurations with
ϕ = 0 (when the actuator is deflated), we will rather
To facilitate the design of PneuFlex actuators, we express equations in terms of the curvature κ of the
present a computational model for its deformation. bottom layer:
Overall, the deformation of such an actuator is com-
1 ϕ
plex. Nevertheless, surprisingly simple design rules κ= = ⇒ ϕ = κ·x
can be derived for designing the most important ac- r x
tuator properties, namely, actuation ratio and rota- We can also express the curvature on the top side
tional stiffness around the actuated axis. This ap- of the actuator:
pendix proposes a suitable formalization, an analy-
ses of simplifications taken, and a basic experimental 1
κtop = 1
evaluation of the resulting equations. κ + h

18
Figure 19: Parameterization of an actuator segment, illustrated at different curvatures. x, Vsil and h stay
constant while λ, r and ϕ change.

Both κ and κtop are defined on the same angle ϕ. The total volume of air contained in the deflated
Therefore we can compute a relationship between an- actuator plus any volumes connected to it, such as
gle ϕ and stretch λ1 : supply tubes:

Vair = Vsupply+V ch = Vsupply + Vact − Vsil

ϕ =
ϕtop
1 The Symbol V will denote the actual volume of the
κ·x = 1 · λ1 x air chamber, which is dependent on the deformation.
κ +h Therefore it is a function of actuator curvature. The
λ1 − 1
κ = (1) total actual volume of the actuator is V + Vsil .
h
λ1 = 1 + κ · h (2) Volume Change For computing the energy stored
x
ϕ = · (λ1 − 1) (3) by the compressed gas (air) within the actuator, we
h need to compute the actual volume with respect to
actuator curvatures. We can do this by first calculat-
Volumes To aid readability of the analysis, we de- ing the total volume of a flexed actuator segment:
fine several constant volume that do not change under
2  2 !
deformation. These volumes can be computed from

ϕ x x
the basic actuator geometry: V + Vsil = z · · π· +h −π·
2π ϕ ϕ
The volume of the effectively incompressible sili-  
cone in the active layer of the actuator: hϕ
= Vact · 1 +
2x
Vsil = z · x · d
As ϕx = κ, we can express actuator curvature in
The volume of the deflated air chamber: terms of air chamber volume and initial geometry as:

Vch = z · x · (h − d) 
h

V + Vsil = Vact · 1 + · κ
2
The total volume of the deflated actuator:  
2 V + Vsil
κ = −1
h h Vact
Vact = z·x·h = · Vsil  
d 2 V − Vch
= Vch + Vsil = ·
h Vact

19
 With Equation 1 we can also compute the relation- And because of the incompressibility assumption:
ship between λ1 and the actuator volume:
J3 = 0
 
V − Vch
λ1 = 1 + 2· (4) Simplifications and Limitations To keep the
Vact model simple, many potentially important effects
This leads to the first insight: actuator curvature were not included:
and stretch of the silicone rubber are linearly propor- We assume a uniform strain energy density within
tional to the gas volume: the rubber hull, which is acceptable for moderately
thin rubber hulls (i.e. d < 0.5h). This assumption is
modeled and discussed in Section A.6.
2 We use a Neo-Hookean material model. This ig-
∂κ = · ∂V
Vact · h nores higher order deformation effects. The error is
2 less than 2% though, as discussed in Section A.7.
∂λ1 = · ∂V
Vact We ignore material stiffening. The consequences
are discussed in Section A.8. The resulting error is
Strain Tensor Invariants The energy stored in typical less than 7.7%.
the rubber during deformation is modeled using Finally, we also ignore the side walls, i.e. hull parts
strain tensor invariants [Gent, 2012]. The strain ten- of the actuator which are stretched only at fractions
sor invariants are related to the orthogonal stretches of λ1 . The ramifications are discussed in Section A.9.
λ1 , λ2 , λ3 : The model can also be invalidated by compressive
forces applied externally. They remove fiber tension
and therefore usually lead to buckling. This limits
J1 = λ21 + λ22 + λ23 − 3 the usefulness of the model for simulation and plan-
J2 = λ21 λ22 + λ21 λ23 + λ22 λ23 − 3 ning. The main purpose is to provide simple equa-
J3 = λ1 · λ2 · λ3 − 1 tions for designing actuator behavior though. Here
the limitations are acceptable.
To compute them, we first define actuator-specific
relations between stretches λ1 , λ2 , λ3 in three prin- A.2 Statically Stable Actuator Con-
cipal directions, which are aligned to x, h and z re- figurations
spectively.
Silicones are effectively incompressible, therefore Using the Minimum Potential Energy Principle we
we can assume a constant volume: can derive statically stable configurations for the ac-
tuator First, we need to define all relevant forms of
λ1 · λ2 · λ3 = 1 work in the system:
The actuator’s radial size does not change either • Gas compression
because of the reinforcement helices. We therefore
set the circumferential stretch λ3 = 1, and get the • Elastic rubber deformation
relationship: • Work added by an external load.
λ2 = λ−1
1 The total potential energy of the actuator is:
This deformation is also called pure shear. Using
W = Wair + Wsil + Wload (6)
these relations, both J1 and J2 reduce to:
Equilibrium is reached, when the gradient of work
J2 = J1 = λ21 + λ−2
1 −2 (5) is zero, which is in our case very simple as we will

20
describe the actuator state with the single variable
G
Z
λ1 : Wsil = · J1 · δV
∆W 2
=0 Vsil
∆λ1
The remainder of this section derives each work G is the material’s shear modulus, and Vsil the
component from the definitions of the previous sec- volume of the silicone. Choosing this simple model
tion and yields the equation for statically stable con- over more complex ones is discussed in Section A.7.
figurations. We will further assume a uniform deformation, and
thus uniform strain energy density (i.e. uniform J1 )
throughout the hull, as justified in Section A.6. We
Gas Compression Work For computing the gas can then calculate the total strain energy as:
compression work, we need to differentiate between
two different control regimes: G
In the mathematically simpler case, pressure is held Wsil = · J1 · Vsil
2
constant under deformation: p(V ) = p. This is either
done actively by using a control system or passively Using Equation 5, we can express the work gradient
by using a big reservoir connected to the actuator w.r.t. stretch λ1 :
volume, which attenuates the effect of volume change
within the actuator on pressure. G
In the second, more common case the total en- Wsil = · (λ21 + λ−2
1 − 2) · Vsil
2
closed gas mass in the system is held constant e.g. δWsil
= G · λ1 − λ−3


when using pneumatic valves. Pressure changes ac- 1 · Vsil (8)
δλ1
cording to the ideal gas equation: p(V ) = nRT · V1 .
A closed gas volume increases the stiffness of the ac-
tuator slightly. But for the sake of brevity, the latter Load Work For a given actuator segment, external
case will not be derived here. load is applied on the interfaces to the two adjacent
In both cases, the work done by changing the vol- segments. By attaching our frame of reference to one
ume of a gas from V1 to V2 is: interface, load work can can be computed by only
considering the motion and force of the other inter-
ZV2 face.
Wair = WV 1 − p(V ) · dV The load work can further be split up into the
V1
work done by translatory forces and rotary moments.
Translatory forces are transmitted by the inelastic
In the case of constant gas pressure, we get a simple fibers of reinforcement helix and passive layer. If we
equation: assume completely inelastic fibers, those forces do not
contribute any work. Rotations, on the other hand,
Wair = W0 − p · (V ) do contribute work, and we can integrate the contri-
δWair δV bution along the bottom layer:
= −p ·
δλ1 δλ1
Z
δWair 1 Wload = M (ϕ) · dϕ
= −p · Vact (7)
δλ1 2

Rubber Deformation Work The deformation For the model, it is more convenient to integrate
work of the silicone rubber Wsil is modeled as a Neo- over x instead of ϕ along the actuator segment. We
Hookean solid model with coefficient C10 = G 2: can rewrite the integral to:

21
 set z = h, as the width of the actuator is usually
Z
∂ϕ (x) approximately its height. We arrive at the equation
Wload = M (ϕ (x)) · dx describing the stiffness of the actuator given its shape
∂x
and deformation:
For short enough actuator segments (small x) we
can assume constant, averaged moment along the δMload
= h3 · G · d · 1 + 3λ−4


1
whole segment, i.e. M (x) = M . Additionally we can δκ
substitute the derivative of Equation 3 for ∂ϕ(x)
∂x : The last, nonlinear term predicts a strong stiffening
when a PneuFlex actuator is straight or even nega-
λ1 − 1
Z tively curved. This sudden stiffening is indeed ob-
Wload = M· · dx served with actual actuators. But they also tend to
h
buckle under such loads.
λ1 − 1
= M· · x + W0
h
Scaling Law When scaling an actuator, the ratios
d z
From this equation, we can compute the work gra- h and h stay constant. The equation then becomes:

dient w.r.t. stretch λ1 : δMload


d

−4
= h4 ·


G · · 1 + 3λ1
δWload x δκ h
=M· (9)
δλ1 h Stiffness therefore scales to the fourth power of ac-
tuator size. This gives us a powerful lever to adjust
Minimum Total Potential Energy By comput- the strength of an actuator.
ing the local minimum of Equation 6 w.r.t. λ1 and
substituting with Equations 7, 8 and 9, we obtain
the equation describing stable actuator states: Non Squared Cross Sections Because the stable
configuration of the helical thread from the applied
uniform radial pressure is a circle, the cross section
δWsil δWload
0 = δW air
δλ1 + δλ1 + δλ1 will always deform into one given high enough air
0 = −p · 2 + G · λ1 − λ1 Vsil + Mload · h(10) pressure.
Vact −3 x


It therefore makes sense to use the the actuator cir-
cumference c instead of height and width to compute
A.3 Stiffness actuator stiffness:
The Stiffness of an actuator segment is expressed by
the change in moment Mload w.r.t. curvature κ. From δMload  c 3
· (G · d) · 1 + 3λ−4


= 1
Equation 10 we can compute Mload it explicitly: δκ 4
For designing actuator stiffness, we can approxi-
2
h ·z mate the nonlinear term with 1:
· p − λ1 − λ−3


Mload = 1 · G · h · d · z
2 δMload  c 3
As we assume a constant pressure regime, the first = · (G · d) (11)
δκ 4
term vanishes when differentiating:
δMload A.4 Actuation Ratio
= −G · h · d · z · (1 + 3 · λ−4
1 )
δλ1 The actuation ratio δκδp is the change of curvature
We then substitute with the derivative of Equa- given an increase in pressure while assuming zero
tion 2 which is ∂λ1 = h · ∂κ. Additionally, we can load. It can also be computed from Equation 10:

22
 [2013]. We believe that this can be caused by a non
circular cross section.
d
p = λ1 − λ−3


1 · G · (12) When pressure increases, the helical thread ten-
h sions and always deforms the actuator cross section
δλ1 1
= into a circle. But until then, the actuator expands in
1 + 3λ−4 · G · hd


δp 1 three dimensions instead of one, which increases the
strain energy in the rubber hull faster with respect to
Using Equation 2 we can substitute λ1 and δλ1 .
stretch λ1 . The effect is more pronounced with less
We get the actuation ratio:
circular cross sections, but can also be elicited by a
loosely wound helical thread. We can investigate the
δκ 1 1 resulting effects by introducing a correction factor for
= −4 · G · d (13)
δp 1 + 3 (1 + hκ) the strain energy gradient:

So according to the simple model, the actuation 2

ratio is inversely proportional to the silicone’s shear (λ1 − 1)
D (λ1 ) = (1 − k0 ) · 2 + k0 (16)
modulus, and its thickness. The actuation ratio also (λ1 − 1) + k12
has a nonlinear component with respect to the actual
curvature. At small curvatures the actuation ratio is The factor k0 ≈ [1 . . . 5.0] defines the relative in-
considerably lower than at higher curvatures. The crease of the gradient for a deflated actuator, while
nonlinearity can be linearized though by a non circu- the factor k1 ≈ 0.025 determines at which elonga-
lar cross section, as discussed in Section A.5. For ac- tion the effect is halved. The latter is probably de-
tuator design we can conveniently drop the nonlinear pendent on the wall’s thickness and rubber stiffness.
term and arrive at a simple design rule for computing The equation was chosen because unlike simpler mod-
the inverse actuation ratio: els the correction factor has a limited range, a finite
integral, and no poles.
D(λ1 ) is then plugged into Equation 12:
δκ 1
≈ (14)
δp G·d
d
p = D(λ1 ) · λ1 − λ−3


1 ·2·G·
Note that this equation is independent of both h h
and z (and therefore also circumference c). The ac-
The impact of different k0 on actuation ratio is il-
tuation ratio is not dependent on the size of the actu-
lustrated in Fig. 20. The actuator can behave almost
ator cross section! We can therefore set an actuation
linearly at higher pressures. An interesting applica-
ratio profile using thickness and shear modulus of the
tion of this effect may be to simplify actuator control.
rubber hull, and then set the stiffness profile with:
δMload  c 3 1
A.6 Justifying Simplification: Uni-
· δκ · 1 + 3λ−4


= 1 (15)
δκ 4 δp form Strain Energies Within
Rubber Hull
A.5 Justifying Simplification: Linear
In the model we assume a uniform strain energy den-
Actuation Ratio sity throughout the hull. This needs to be checked
Equation 12 contains the nonlinear term λ1 − λ−3 though, as the principal stretches are not uniform at


1 .
The model therefore predicts pressure to be non- all.
linearly related to actuator curvature. Interestingly
though, the actuation ratio can be linear at moder- Formalization For an infinitesimal volume of elas-
ate pressures, as it was observed in Deimel and Brock tomer within the hull, we define d to be the radial

23
 1.25 boundary of the hull yields:
model
k1 = 0.02
1.20 k1 = 0.03 h − d0
k1 = 0.04 λ3 =
linearization
h−d
1.15
Finally, we can derive λ2 by using the incompress-
λ1

1.10 ibility assumption λ1 λ2 λ3 = 1 and get:

1.05
h−d
λ2 =
1.00 κ · h2 + κ · d0 2 − (2 · κ · h + 1) d0 + h
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Normalized pressure p · h To normalize our calculations, we can express the
2·G·d
principal stretches in terms of the dimensionless ra-
0
tios dh , hd , and κ·h:
Figure 20: A non circular cross section effectively
linearizes the actuation ratio at moderate curvatures.
d0
 
λ1 = 1 + κ·h · 1−
h
distance of the volume from the outer boundary of
1 − hd
the hull in an undeformed actuator, while d0 denotes λ2 = 0 2 0
κ·h + κ·h · dh − (2 · κ·h + 1) dh + 1


the actual radial distance in the deformed actuator.
0
When the actuator bends, the volume moves ra- 1 − dh
dially outwards as the wall thins. This motion in- λ3 =
1 − hd
fluences the stretches encountered. As the radial
displacement is limited by the helical reinforcement With λ2 we can express the relation between d0
fibers, we can assume the outer boundary of the rub- and d as a linear differential equation using λ as the
2
ber hull to not move radially at all. gradient of displacement:
Due to the smaller radius the longitudinal stretch
is:
δd0
= λ2
δd
λ1 = 1 + (κ · h − κ · d0 ) (17)
δd0 1 − hd
= 2
The circumferential stretch can be calculated by δd d 0 0
κ·h − dh (2 · κ·h + 1) + 1 + κ·h


h
the radial displacement given a position d with re-
spect to the undeformed position d: This differential equation determines the position
d0 of a packet of rubber, and therefore its deforma-
tion.
π (h − d0 ) π (h − d0 − δd0 )
δλ3 = −
π (h − d) π (h − d) Strain Energy Distribution To analyze the
0
δd strain energy, the differential equation was solved nu-
= −
h−d merically. The boundary conditions were set on the
−d0 outer boundary of the hull at d0 = 0 to λ2 = 1 and
⇒ λ3 = +C λ3 = λ11 .
h−d
Fig. 21 shows the strain energy density w.r.t. the
normalized depth hd within the hull and at different
Setting the boundary condition λ3 = 1 at the outer normalized curvatures κ·h. We can see that the strain

24
 C01 , C02 and C20 though. The coefficients published
by Meier et al. [2005] have the following relations:
0.8
1
C01 ≈ · C10
0.6 50
1
J1

0.4
C20 ≈ · C10
κ·h = 0.5 500
κ·h = 0.4 C02 ≈ 0
0.2 κ·h = 0.3
κ·h = 0.2
κ·h = 0.1
κ·h = 0.0
0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Ignoring the coefficients C01 , C02 and C20 (i.e. set-
d ting them to 0) yields the Neo-Hookean model.
radial position h We can bound the error to strain energy when
assuming a maximum stretch of λ1 < 3 (which re-
Figure 21: Strain energy density distribution within lates to an actuator bending at the radius of half its
the rubber hull for different normalized curvatures. height), which given Equation 5 bounds the strain
tensor values to:
J1 < 7.11
energy density stays surprisingly flat even for moder-
ately thick hulls and at very strong curvatures. J2 < 7.11
The reason for this surprising result can be under- The terms dropped from the Mooney-Rivlin model
stood when looking at the principal stretches when are bounded to:
the rubber moves outwards radially (thinning the
wall). Circumferential stretch λ3 increases, but at the
same time radial stretch λ2 decreases as the packet C01 · J2 < 0.02 · C10 · J1
gets more compressed. Also λ1 decreases with the C20 · J12 < 0.014 · C10 · J1
distance from the bottom layer.
The analysis shows that as long as the hull thick- Which results in a total error of less than 2%.
ness is less than half the actuator height, we can as-
sume a uniform strain energy distribution for mod-
A.8 Justifying Simplification: Ignore
eling. Staying below this limit also avoids material
fatigue of the rubber on the inside of the air chamber. Material Stiffening
Elastomers exhibit a stiffening at large stretches.
A.7 Justifying Simplification: Neo- This is modeled by augmenting the strain energy
Hookean Deformation Model function with an additional parameter. The Gent
model [Gent, 2012] augments the strain energy func-
An alternative to the simple Neo-Hookean defor- tion with a logarithmic term:
mation model used in our model is the generalized
Mooney-Rivlin model for incompressible hyperelas-  
tic materials. It states a polynomial approximation G J1 (λ1 )
W = · J1 (λmax ) ln 1 − · Vsil
of strain energy density, using the strain tensor in- 2 J1 (λmax )
 
variants: G J1
= · Jm ln 1 − · Vsil
n
2 Jm
dW X
= Cij · J1i · J2j With the additional material parameter λmax
dv i,j=0
which is the stretch where the material exhibits un-
Silicone rubbers have rather small coefficients for limited stiffness.

25
 The rubber used for the PneuFlex actuator typi- equation for the actuation ratio, the second the equa-
cally has λmax ≈ 10, which results in: tion for actuator stiffness.

Jm ≈ 100
A.10.1 Actuators with Varying d
The model predicts that the actuation ratio is only
When computing the strain energy gradient using dependent on hull thickness d and the shear modulus
the Gent model, we get: of the rubber. To test this, we built four actuators
with the same shape but with a d of 2.5 mm, 3.5 mm,
Gent
δWsil 1 2λ1 − 2 · λ−3
1
4.5 mm and 5.5 mm respectively.
= C1 · Jm · Vsil · 2 +λ−2 −2
· Fig. 22 shows the actuators at a pressure of
δλ1 λ
1 − 1 Jm 1 Jm
46.3 kPa. Despite the change in height and width,
1 δWsil the curvatures of the bottom layers are constant along
= ·
λ21 +λ−2
1 −2 δλ the actuators, as indicated by the circle segments.
1− Jm
1
We can therefore validate the model’s prediction that
For plausible values of λ1 < 3 and Jm = 100, the height and width do not influence actuation ratio.
ignored stiffening factor can be bounded to: Please note that the circle segments are placed on top
of the edges of the bottom layers, where the fibers of
Gent
δWsil δWsil the embedded fabric are most stressed.
< 1.077 ·
δλ1 δλ1 Fig. 23 shows the relationship between pressure
So ignoring the material stiffening introduces an and fingertip orientation, which is an aggregated
error of less than 7.7%. measure of the curvature along the actuator. The
curves for each actuator show the nonlinear behavior
A.9 Justifying Simplification: Ignor- predicted by Eq. 13, i.e. an increase in actuation ratio
towards higher curvatures.
ing the Side Walls The dotted lines in Fig. 23 indicate the actuation
For analyzing the impact of ignoring the side walls, ratio when scaling the measurements of the thinnest
we can conceptually split the cross section of a real actuator (2.5 mm) to the thickness of the other actu-
Pneuflex actuator into many small parts, of which ators according to our model. Model and measure-
each behaves according to the simple model. ment agree well for 3.5 mm and 4.5 mm. At 5.5 mm
As there is no air below the side walls (only more it is clearly visible that the nonlinearity of the actua-
rubber), there is no additional force applied by gas tion ratio increases, making the actuator stiffer than
pressure. expected at low pressures.
The amount of deformation work δW sil
δλ1 does in-
crease, but the J1 also drops off quadratically when A.10.2 Actuator with Linear Stiffness
approaching the bottom layer. By using a constant
ratio between the thickness of the side wall and the To validate Equation 11 and 15, we can apply a force
top side of the actuator, the error being made when at the tip of an actuator. The actuator has a constant
computing the actuation ratio can be made constant. actuation ratio along its main axis. The contact force
For the stiffness, the side walls play even less of a will create a bending moment to segments of the ac-
role, as it scales with h3 . tuator that increases linearly with the segment’s dis-
tance from the contact point. At small curvatures,
i.e. an almost straight finger, the distance along the
A.10 Experimental Validation
bottom layer will approximate the euclidean distance
To validate the design rules we developed, we con- well. For a straight actuator, we can therefore as-
ducted two experiments. The first one validates the sume the moment along the actuator to be increasing

26
Figure 22: Example of four fingers with different actuation ratios, inflated to 46.3 kPa. The overlaid circle
segments indicate the constant curvature along the bottom layer.

linearly. If the actuator has a linear stiffness profile

along the actuator and a constant actuation ratio too,
then the load moment and actuated moment will can-
180 cel out and yield a constant curvature. The constant
actuation ratio is demonstrated in Fig. 22.
160 Fig. 24 shows an actuator at four different pres-
sures. The curvature stays constant during a large
140
range of inflation pressures, with the highest pres-
sure corresponding to about a 360◦ rotation if there
Tip orientation [◦ ]

120
was no contact. The actuator therefore has a linear
100 stiffness profile, which validates Equation 11.
The upwards rotation of the whole finger relative
80 to the fixture is caused by the rubber cap which closes
the air chamber at the base of the finger. This also
60 2.5 mm
causes the small curvature changes between different
3.5 mm
4.5 mm pressures. Fig. 24 (d) shows a slightly stronger cur-
40
5.5 mm vature close to the tip. This may be caused by dif-
3.5 (model)
20 ference between contact point location and the point
4.5 (model)
5.5 (model)
where the stiffness profile reaches zero stiffness, which
0 is exactly at the tip.
0 20 40 60 80 100
Pressure [kPa]
A.11 Summary
Figure 23: Tip orientation versus inflation pressure The presented theoretic model yields two simple de-
for different rubber hull thicknesses. Dotted lines in- sign rules (Equations 14 and 15) for designing actua-
dicate model estimates based on the 2.5 mm measure- tion ratios and stiffnesses along a PneuFlex actuator.
ment We also validated the scaling behavior of the model
in two experiments.
Based on the analysis, we can give the following
recommendations for choosing geometric parameters:

27
 (a) 0 kPa (b) 17.3 kPa (c) 41.6 kPa (d) 80.1 kPa

Figure 24: The finger is blocked by a fingertip contact while being inflated. An actuator with a linearly
decreasing stiffness profile will show a constant curvature along the actuator under such load.

• active layer thickness should stay less than half

the actuator height: hd < 0.5

• To achieve a minimum bending radius of 1.5

times actuator height, the rubber should have
an elongation at break of at least 500%
• Material stiffening can usually be ignored with
silicone rubber, the error is typically less than
7.7%

B Multimedia Extensions
Nr. Type Description
1 Images High resolution images of grasps
from the Feix taxonomy in
Fig. 12
2 Data Fingertip positions of
RBO Hand 2 for the Feix
taxonomy
3 Data Joint angles of five human partic-
ipants executing grasps from the
Feix taxonomy recorded with a
Cyberglove II
4 Video Video of how grasp postures for
grasps from the Feix taxonomy
were obtained
5 Video Video of example grasps from a
tabletop under human control
6 Video Video of manipulating two heavy
objects

28
References R. Deimel and O. Brock. A compliant hand based on
a novel pneumatic actuator. In IEEE International
J.R. Amend, E.M. Brown, N. Rodenberg, H.M. Conference on Robotics and Automation (ICRA),
Jaeger, and H. Lipson. A positive pressure uni- pages 2047–2053, 2013.
versal gripper based on the jamming of granular
material. IEEE Transactions on Robotics, 28(2): R. Deimel, C. Eppner, J. Alvarez-Ruiz, M. Maertens,
341–350, 2012. and O. Brock. Exploitation of environmental
constraints in human and robotic grasping. In
J. Bae, S. Park, J. Park, M. Baeg, D. Kim, and S. Oh. 16th International Symposium on Robotics Re-
Development of a low cost anthropomorphic robot search (ISRR), 2013.
hand with high capability. In IEEE/RSJ Interna-
tional Conference on Intelligent Robots and Sys- Raphael Deimel and Oliver Brock. A novel type of
tems (IROS), pages 4776–4782, 2012. compliant, underactuated robotic hand for dexter-
ous grasping. In Proceedings of Robotics: Science
J. Bishop-Moser, G. Krishnan, C. Kim, and S. Kota. and Systems, Berkeley, USA, July 2014.
Design of soft robotic actuators using fluid-filled
A. M Dollar and R. D Howe. Simple, reliable robotic
fiber-reinforced elastomeric enclosures in parallel
grasping for human environments. In IEEE Inter-
combinations. In IEEE/RSJ International Con-
national Conference on Technologies for Practical
ference on Intelligent Robots and Systems (IROS),
Robot Applications (TePRA), pages 156–161, 2008.
pages 4264–4269, 2012.
Aaron M. Dollar and Robert D. Howe. The highly
M. G. Catalano, G. Grioli, E. Farnioli, A. Serio,
adaptive SDM hand: Design and performance eval-
C. Piazza, and A. Bicchi. Adaptive synergies for
uation. The International Journal of Robotics Re-
the design and control of the pisa/IIT SoftHand.
search, 29(5):585–597, 2010.
The International Journal of Robotics Research, 33
(5):768–782, 2014. T. Feix, R. Pawlik, H. Schmiedmayer, J. Romero,
and D. Kragic. A comprehensive grasp taxon-
M. Ciocarlie, F. Mier Hicks, and S. Stanford. Ki- omy. In Robotics, Science and Systems: Workshop
netic and dimensional optimization for a tendon- on Understanding the Human Hand for Advancing
driven gripper. In IEEE International Conference Robotic Manipulation, 2009.
on Robotics and Automation (ICRA), pages 2751–
2758, 2013. M. Gabiccini, E. Farnioli, and A. Bicchi. Grasp
analysis tools for synergistic underactuated robotic
M. Controzzi, C. Cipriani, and M. C. Carozza. De- hands. The International Journal of Robotics Re-
sign of artificial hands: A review. In The Human search, 32(13):1553–1576, 2013.
Hand as an Inspiration for Robot Hand Develop-
ment, volume 95 of Springer Tracts in Advanced I. Gaiser, S. Schulz, A. Kargov, H. Klosek, A. Bier-
Robotics, pages 219–247. Springer, 2014. baum, C. Pylatiuk, R. Oberle, T. Werner, T. As-
four, G. Bretthauer, and R. Dillmann. A new an-
Utku Culha, Surya G. Nurzaman, Frank Clemens, thropomorphic robotic hand. In 8th IEEE-RAS In-
and Fumiya Iida. SVAS3: Strain vector aided sen- ternational Conference on Humanoid Robots (Hu-
sorization of soft structures. Sensors, 14(7):12748– manoids), pages 418–422, 2008.
12770, 2014.
Kevin C. Galloway, P. Polygerinos, C. J. Walsh, and
M. R Cutkosky. On grasp choice, grasp models, and R. J. Wood. Mechanically programmable bend
the design of hands for manufacturing tasks. IEEE radius for fiber-reinforced soft actuators. In Ad-
Transactions on Robotics and Automation, 5(3): vanced Robotics (ICAR), 16th International Con-
269–279, 1989. ference on, pages 1–6, 2013.

29
Alan N Gent. Engineering with rubber: how to design R. Ma, L. Odhner, and A. Dollar. A modular, open-
rubber components. Carl Hanser Verlag GmbH Co source 3D printed underactuated hand. In IEEE
KG, 2012. International Conference on Robotics and Automa-
tion (ICRA), pages 2737–2743, 2013.
A.P. Gerratt, N. Sommer, S.P. Lacour, and A. Bil-
lard. Stretchable capacitive tactile skin on hu- A.D. Marchese, K. Komorowski, C.D. Onal, and
manoid robot fingers; first experiments and results. D. Rus. Design and control of a soft and continu-
In Humanoid Robots (Humanoids), 2014 14th ously deformable 2d robotic manipulation system.
IEEE-RAS International Conference on, pages In IEEE International Conference on Robotics and
238–245, Nov 2014. Automation (ICRA), pages 2189–2196, May 2014.
M. E. Giannaccini, I. Georgilas, I. Horsfield, B. H. P. Meier, M. Lang, and S. Oberthr. Reiterated ten-
P. M. Peiris, A. Lenz, A. G. Pipe, and S. Dogra- sion testing of silicone elastomer. Plastics, Rubber
madzi. A variable compliance, soft gripper. Au- and Composites, 34(8):372–377, 2005.
tonomous Robots, 36(1–2):93–107, 2014.
M. Giorelli, F. Renda, A. Arienti, M. Calisti, B. Mosadegh, P. Polygerinos, C. Keplinger,
M. Cianchetti, G. Ferri, and C. Laschi. Inverse and S. Wennstedt, R. F. Shepherd, U. Gupta, J. Shim,
K. Bertoldi, C. Walsh, and G. M. Whitesides.
direct model of a continuum manipulator inspired
Pneumatic networks for soft robotics that actuate
by the octopus arm. In Biomimetic and Biohybrid
Systems, number 7375 in Lecture Notes in Com- rapidly. Advanced Functional Materials, 24(15):
puter Science, pages 347–348. Springer, 2012. 2163–2170, 2014.

M. Grebenstein. Approaching Human Performance - L. Odhner, L. Jentoft, M. Claffee, N. Corson, Y. Ten-

The Functionality Driven Awiwi Robot Hand. Dis- zer, R. Ma, M. Buehler, R. Kohout, R. Howe, and
sertation, ETH Zürich, Zürich, 2012. A. Dollar. A compliant, underactuated hand for
robust manipulation. The International Journal of
G. Grioli, M. Catalano, E. Silvestro, S. Tono, and Robotics Research, 33(5):736–752, 2014.
A. Bicchi. Adaptive synergies: an approach to
the design of under-actuated robotic hands. In D. Prattichizzo, M. Malvezzi, M. Gabiccini, and
IEEE/RSJ International Conference on Intelli- A. Bicchi. On the manipulability ellipsoids of
gent Robots and Systems (IROS), pages 1251–1256, underactuated robotic hands with compliance.
2012. Robotics and Autonomous Systems, 60(3):337–346,
2012.
S. Hirose and Y. Umetani. The development of soft
gripper for the versatile robot hand. Mechanism Rahim Rahimi, Manuel Ochoa, Wuyang Yu, and
and Machine Theory, 13(3):351–359, 1978. Babak Ziaie. A sewing-enabled stitch-and-transfer
F. Ilievski, A. Mazzeo, R. F. Shepherd, X. Chen, and method for robust, ultra-stretchable, conductive
G. M. Whitesides. Soft robotics for chemists. Ange- interconnects. Journal of Micromechanics and Mi-
wandte Chemie International Edition, 50(8):1890– croengineering, 24(9):095018, 2014.
1895, 2011. F. Renda, M. Cianchetti, M. Giorelli, A. Arienti, and
I. A. Kapandji. Cotation clinique de l’opposition C. Laschi. A 3D steady-state model of a tendon-
et de la contre-opposition du pouce. Annales de driven continuum soft manipulator inspired by the
Chirurgie de la Main, 5(1):68–73, 1986. octopus arm. Bioinspiration & Biomimetics, 7(2):
025006, 2012.
M. Kazemi, J Valois, J. A. Bagnell, and N. Pollard.
Human-inspired force compliant grasping primi- J. Romero, T. Feix, H. Kjellstrom, and D. Kragic.
tives. Autonomous Robots, 37(2):209–225, 2014. Spatio-temporal modeling of grasping actions. In

30
 2010 IEEE/RSJ International Conference on In- for underactuated anthropomorphic hands. In
telligent Robots and Systems (IROS), pages 2103– IEEE/RSJ International Conference on Intelligent
2108, 2010. Robots and Systems (IROS), pages 274–280, 2014.

M. Santello, M. Flanders, and J. F. Soechting. Pos- Wall V., R. Deimel, and O. Brock. Selective stiffen-
tural hand synergies for tool use. The Journal of ing of soft actuators based on jamming. In IEEE
Neuroscience, 18(23):10105–10115, 1998. International Conference on Robotics and Automa-
tion (ICRA), page (accepted for publication), May
Mahmoud Tavakoli and Anbal de Almeida. Adap- 2015.
tive under-actuated anthropomorphic hand: Isr-
softhand. In IEEE/RSJ International Conference M. Wehner, Y. Tolley, M. T.and Meng, Y. L. Park,
on Intelligent Robots and Systems (IROS), pages A. Mozeika, Y. Ding, C. Onal, R. F. Shepherd,
1629–1634, 2014. G. M. Whitesides, and R. J. Wood. Pneumatic
energy sources for autonomous and wearable soft
Mahmoud Tavakoli, Baptiste Enes, Lino Marques, robotics. Soft Robotics, 1(4):263–247, 2014.
and Anbal de Almeida. Actuation strategies

31