Content uploaded by Xing Wang
Author content
All content in this area was uploaded by Xing Wang on Mar 23, 2024
Content may be subject to copyright.
Fin-QD: A Computational Design Framework for Soft Grippers: Integrating
MAP-Elites and High-fidelity FEM
Yue Xie1,†, Xing Wang2,†, Fumiya Iida1, David Howard2
Abstract— Computational design can excite the full potential
of soft robotics, but it has the drawback of being highly
nonlinear in terms of material, structure, and contact. To
date, enthusiastic research interests have been demonstrated
for individual soft fingers, but the frame design space (how
each soft finger is assembled) remains largely unexplored.
Computational design remains challenging for the finger-based
soft gripper to grip across multiple geometrically distinct
object types successfully. Including the design space for the
gripper frame can bring huge difficulties for conventional
optimization algorithms and fitness calculation methods due to
the exponential growth of design space. This work proposes an
automated computational design optimization framework that
generates gripper diversity to individually grasp geometrically
distinct object types based on a quality-diversity approach.
This work first discusses a significantly large design space (28
design parameters) for a finger-based soft gripper, including
the rarely-explored design space of finger arrangement. Then,
a contact-based Finite Element Modelling (FEM) is proposed in
SOFA to output high-fidelity grasping data for fitness evaluation
and feature measurements. Finally, diverse gripper designs are
obtained from the framework while considering features such
as the volume and workspace of grippers. This work bridges
the gap of computationally exploring the vast design space of
finger-based soft grippers while grasping large geometrically
distinct object types with a simple control scheme.
I. INTRODUCTION
Recent research in soft robotics [1] has shown the potential
benefits of using constitutionally soft materials to design
artificial agents. They are frequently designed to mimic
biological organisms, such as the soft prosthetic hand [2],
soft robotic arm [3], and soft locomotion robot [4]. Benefiting
from Soft robotic grippers’ effectiveness and popularity in
industry applications, such as pick and place in assemble
line [5], underwater exploration [6], and agricultural har-
vesting [7]. Designing soft robotic grippers to meet such
applications’ demands also remains to be of great interest to
the research community. The design method has gradually
moved from human-dominant [8], bio-inspired approaches
to computational design [9]–[11].
The computational design shows remarkable potential to
output suitable and high-performance soft robotics once
given specific design requirements [10], [12]. It can be
initialised by defining a simplified design domain from a
human-driven or bio-inspired structure. The behaviour or
performance of those initial designs needs to be evaluated
using mathematical or physical simulation, which enables
†Equal contributions as first authors
1Department of Engineering, University of Cambridge, UK
2Robotics Design and Interaction Group, Data61, CSIRO, Australia
the designers to validate or refine the design without per-
forming the physical prototypes. Soft robotic grippers ben-
efit significantly from computational design, owing to their
intricate nonlinearity from constitutional material, structure
design, and contact, and multimodality in the search space,
which pose significant challenges in accurately modelling
and searching the design space.
However, to date, computational design has explored a
limited design space for finger-based soft grippers, mainly
on the design space of individual soft robotic fingers. A
significantly larger design space that includes a way to
map individual fingers is rarely researched for grasping. For
example, individuals are mostly assembled on a manually-
designed finger mount to form a multiple-finger-based grip-
per. Liu et al. [13] generated two optimal finger designs
from topology optimization and manually assembled them
symmetrically to form a two-finger gripper. The combination
of individuals and their relative positioning has not been
properly investigated. This unexplored design space is vast
and complex, and a combination of soft fingers results in an
exponential growth of design space.
To tackle those challenges, this study presents a com-
putational design framework to investigate the evolution of
soft robots while performing contact-based grasping tasks.
The work significantly expands the SOTA computational
design of finger-based soft robotic grippers, involving quality
diversity to a general evolutionary system (composed of a
rapid and high-fidelity soft robot simulator, a bio-inspired
generative encoding, and a quality-diversity EA). The vast
design space (28 design parameters) is covered and explored
by scripting the individual fin ray design and assembling
them in certain combinations. A high-fidelity contact-based
simulation that covers a complete grasping motion is pro-
posed in the SOFA framework, which outputs the grasping
success rate for performance evaluation.
This work evaluates how the values of different behaviours
affect the evolution of different combinations of design pa-
rameters and performance in accomplishing tasks. Our study
also demonstrates the potential of evolutionary techniques
to produce a variety of high-performance soft robots, and
our analysis may provide new insights into the evolution of
soft grippers. To summarise, the main contributions of this
research work are as follows:
•Propose a highly automated design optimization frame-
work that generates gripper diversity using a quality-
diversity optimization algorithm
•Explore and search for a significantly higher dimen-
sional design space that includes the domain for indi-
Fig. 1: Design optimization framework to generate diverse gripper designs
vidual robotic fingers and their combination
•Demonstrate a rapid and high-fidelity contact-based
FEM to evaluate the complete soft grasping across
multiple geometrically distinct objects
The remainder of this paper is organized as follows.
Section II describes the Related work in the research field.
Section III specifies the detailed components for our frame-
work. Then, Section IV covers the validation and results
on our optimized gripper. This study’s key conclusions and
future directions are summarized in Section V.
II. REL AT ED WO RK
Computational design has the potential to create optimal
soft robotics by leveraging the power of computation tools
despite their challenges of high dimensional design space,
nonlinearity and multi-modality search space. It attempts
to overcome these challenges by design parameterization,
rapid analytical, FEM or physical evaluation, and efficient
optimization algorithm [10]. The genetic optimization frame-
works are of interest here, even though some pioneering
frameworks only work for specific types of soft grippers and
have restricted the generalization of the method [14].
From the optimization algorithm perspective, traversing
the design space involves different levels of computational
cost depending on the degree of freedom in the parameteriza-
tion. Topology optimization utilizes a gradient-based method
to design the soft robots by optimizing the discrete material
distribution under predefined load and boundary constraints
[15]. The decision variables are usually directly encoded into
discrete elements, requiring tens of thousands up to billions
of parameters [16]. Josh et al. [17] proposed a pioneering
approach to topology optimization that creates soft grippers
with multiple materials and explores techniques for produc-
ing hermetically sealed designs. It explores a vast design
space with higher simulation accuracy. The disadvantages of
this algorithm include the necessity of detailed loading and
boundary conditions for initialization, the generation of a
single feasible design, the significantly high computational
cost and less convergence stability.
Bayesian optimization (BO) provides a feasible solution
to optimize objective functions that are expensive to analyze,
resulting in a more efficient traverse of the design space. Our
previous work presented a BO-based computational design
framework to search a continuous design parameterization
space [18]. A sophisticated FEM based on COMSOL was
implemented to find the compliance and normal contact
force. It focuses on the design space for individual fingers
only. It is not applicable for a complete multi-step grasping
motion due to the high expensive computational cost of
contact-based FEM. Bio-inspired optimization algorithms
demonstrate the capability of solving high dimensional opti-
mization problems that are challenging for BO that expe-
rience performance degradation. However, because of the
iterative, population-based nature of the solver, the compu-
tational evolution of soft robots tends to be computationally
expensive. It relies on low-fidelity solvers rather than high-
fidelity FEM for fitness evaluation. This often leads to soft
robot designs with notable reality gaps [10].
In this work, we aim to promote state-of-the-art of finger-
based soft grippers by implementing gradient-free EA, in-
cluding the quality diversity algorithm to explore the expo-
nential growth design space, the rapid and high-fidelity FEM
for grasping, and finally, the generation of soft grippers with
diverse features.
III. COM PUTATIONAL DESIGN FR AME WORK
The section details the working principle of the gradient-
free EA framework for soft gripper optimization, as shown
in Figure 1. The goal of the computational optimization is to
find the top-ranking gripper designs that demonstrate promis-
ing performance in grasping multiple geometrical-distinct
objects. A benchmark three-fingered gripper is designed with
fin ray structure due to its ubiquity throughout research and
industry, whose design can be varied by modifying one or
TABLE I: Design Variables, symbol, data types and their
ranges, i={1,2,3}.
Variables Symbol Design range
Finger length HiContinuous, [90,120]
Finger width LiContinuous, [28,40]
Finger thickness WiContinuous, [20,35]
Length of solid tip Ltipi Continuous, [20,30]
Thickness of the contact surface tflexi Continuous, [1,3]
Thickness of the non-contact surface tbacki Continuous, [1,3]
Number of linkage ribs NiDiscrete, [1,10]
Thickness of ribs tribi Continuous, [1,3]
Tilt angle of ribs Danglei Continuous, [−40,40]
G1: distance between fingers dmount Continuous, [30,40]
any combination of the total 28 design parameters. Once the
design is finalised, Gmsh exports high-quality tetrahedron
mesh and imports it as TetdrahedronFEMforcefield to the
SOFA simulator. The gripper design and configuration are
loaded to SOFA for fitness evaluation, which is set to be
the successful grasp rate and is calculated across all target
objects to evaluate the performance of the current gripper
design.
A. Problem statement
The optimization problem is formulated to discover the
mapping between dynamic grasping quality and the pa-
rameterised variables that can generate valid designs. The
grasping quality of a soft robotic gripper can be evaluated
by different objectives (fitness functions), such as grasping
rate, stability, disturbance resistance, dexterity, etc [19]. The
most intuitive and critical factor is to maximise the grasping
success rate, as shown in (1). It measures the effectiveness of
a robot’s ability to pick up and hold objects. It is calculated
by dividing the number of successful grasps by the total
number of attempted grasps. The grasping is classified as
successful using the Elasped-time threshold, which will be
detailed under Section Rapid simulation.
obj = arg max
x(f1(x)) (1)
s.t. lowi< xi< highi∀xi∈x(2)
While f1is the objective function (success rate), and xi
represents the design parameters.
B. Gripper Parameterisation
The basic gripper design comprises three distinct fin
ray fingers, a pneumatic linear actuator for actuation, and
rigid finger mounts that link individual fingers. Those rigid
finger mounts are programmed with changeable in-plane size
dmount to allow a tunable scale of the overall soft gripper. We
have added a Supplementary Video (SV1) to demonstrate
how the distance between fingers can be modified.
In addition, nine parameters are assigned for each finger
to define its design space, including finger length (H), width
(L), height (W), length of solid tip (Ltip), the thickness
of the contact surface (tflex ), the thickness of the non-
contact surface (tback), the number of linkage ribs (N), the
thickness of ribs (trib), and the tilt angle of ribs (tang le)
(Figure 2). The rich combination of three fingers is coded in
the evolutionary algorithm, ending with 28 design variables
(three fingers ×nine variables per finger plus one defining
the finger distance). Table I lists the design ranges of all
design variables.
Supplementary Video (SV1) is attached to provide a more
intuitive explanation of the vast design space of the soft
gripper. We replicate the same gripper design using the Open
Cascade Python library for smooth implementation in the
optimization framework. Note that the benchmark gripper
has three identical fingers with 9design parameters being
[94.5, 37.5, 21.3, 15, 0, 8, 2, 2, 2] [20].
C. Rapid simulation
An open-source physical simulator SOFA for high-fidelity
objective calculation is adopted in the framework to evaluate
the fitness and measure the behaviours of designed grippers
[21]. We evaluate the grasping success rate across geometri-
cally distinct objects for each design iteration. Figure 3 shows
the detailed setup for the high-fidelity SOFA simulation.
SOFA make the simulation a scene with an intrinsic
generalised hierarchy. It starts with a parent node, “root”,
and comprises more child nodes in a tree structure. Each
child node can represent one object (e.g. soft finger) in the
simulation scene, and various subnodes of the child node
are created to demonstrate the different representations of
the same object. For example, the same soft finger may be
imported with its visual, mechanical, collision and potentially
haptic representations in a simulation scene. Our simulation
simplifies and accelerates the evaluation by importing three
soft fingers and the target objects into the scene. The design
space for finger assembly is varied by reloading the relative
configuration of all fingers at the beginning of each simula-
tion. A simple controller that generates bending motion for
soft robotic fingers is added to the scene. The contact is set
up between the fingers and the target object using friction
contact constraint. Self-collision is enabled for soft fingers
to avoid self-penetration on the internal rib region.
The grasping simulation In SOFA is achieved by dis-
cretising the temporal domain into small time steps. The
ODESolver is used to construct a linear matrix of the system
for further solving. We implemented the EulerImplicitSolver
as the timeIntegrationSchema. The implicit solver calculates
the internal or external force on unknown positions at the
next time step, which shows better accuracy than the explicit
solver by constructing a more complex linear equation. Af-
terwards, iterative CGLinearSolver is implemented to solve
this equation. Each simulation is performed by specifying the
solving iterations (without GUI visualisation). The grasping
is considered successful if the object’s position remains
within the threshold distance of the gripper after a particular
elapsed time.
D. Quality-diversity approach
The design updating is represented as a combinatorial
optimization problem. We generate a diverse collection of fin
ray soft grippers through CMA-ME [22], a quality-diversity
Fig. 2: Gripper parameterization: details for the design variables on one fin ray finger
Fig. 3: SOFA simulation setup
algorithm combining the adaptation mechanisms of CMA-
ES [23] with the archiving mechanism of MAP-Elites [24]. It
specialises in continuous domains and is significantly more
sample-efficient than other QD algorithms. The workspace
and the volume of grippers are set as the features in the
algorithm. As present in Figure 1, the algorithm initialises
specific populations and evaluates those designs by checking
the preset threshold. The algorithm is terminal until it reaches
the computational budget, providing the gripper designs
with rich features. All those optimized grippers demonstrate
superior performance under the simulation environment.
Workspace is defined as the range of positions a robot
can reach to interact with its physical environment. Thus,
a robotic manipulator’s workspace consists only of all pos-
sible positions of the robot’s tip, or robotic gripper. Soft
grippers can interact with target objects at all points within
the volume swept due to their actuator deformations. All
passive and active deformations possible for the soft gripper
actuators contribute to the workspace. This swept volume or
workspace volume strongly correlates with the payload size
range and grasping versatility of the gripper [25]. Here, the
workspace of the soft gripper is varied when the fingers’
relative distance changes. The maximum area of the three
fingers’ equilateral triangle is calculated to represent the
maximum workspace, which is treated as the first feature.
fea1=3√3
4×d2
mount (3)
In addition, the other feature we are interested in is the
structure complexity, which can be represented using the total
volume of the fin ray design.
fea2=Vmount(dmount) + Vribs(N , trib)
+Vside(H, L, W, Langle, Ltip , tf lex, tback)(4)
IV. RES U LTS AND DISCUSSION
To retrieve the high-fidelity SOFA result, material proper-
ties must be obtained experimentally. We designed the dog-
bone elastomer samples and performed the universal tensile
test based on the ASTM D412 standard. An Instron 34SC-5
equipment was used to collect the experimental tensile results
5times, and the average strain-stress curve was plotted to fit
the linear elastic material model in SOFA. Several digitally
mixed materials made from the Aglius 30 and Vero family
were tested, and the one with a shore hardness of 85A was
eventually implemented in the simulation tests. The material
was assigned with an elastic modulus of 11.6 MPa and
passion ratio of 0.49 due to the incompressible nature of
rubber-like soft material.
The same quality of Tetrahedron mesh was applied across
all fingers, while the mesh for the objects was set to be
coarser to prevent penetration from the master nodes (rigid
objects) to slave nodes (soft fingers). The minimum and
maximum mesh ranges were set to be [0.5,1] and [2,4] using
Gmsh API, respectively. Modifying the mesh quality can also
allow a different fidelity of simulation results, which was
tested before the implementation of the current mesh quality.
We set up a maximum iteration of 1000 for the iterative linear
solver and a tolerance of 1e-6. The residual of the linear
matrix is reduced each time with an increase in iteration.
A complete SOFA simulation process is demonstrated in
Figure 4 and Supplementary Video SV2.
Fig. 4: SOFA simulation for a complete grasping process.
The grasping is treated as a success if the object is still
within the preset height in the scene.
(a) Archive coverage trade
through the evaluation
(b) QD score through the eval-
uation
Fig. 5: Evaluate indicates of the Quality diversity algorithm
The experiment is run on a Dell Precision Workstation
with an Inter I9-12900H CPU and is configured to the termi-
nal after 5iterations, considering time complexity and setting
the batch size to 15. Figure 5 presents the improvement trade
of the quality diversity algorithm via archive coverage and
QD score, where archive coverage presents how well the
archive represents the diversity and quality of the solutions
found so far. QD score [26] is a holistic metric which sums
the objective values of all cells in the archive. It can be seen
that even in 5generations, the archive’s coverage and the
quality of the resulting design grows.
Figure 6 presents the distribution of high-quality designs
of the final archive within the behavioural space and the
corresponding cell in the behaviour space of the bench
gripper. The presented heatmap clearly illustrates that the
Fig. 6: Distribution of high-performance designs and bench-
mark design
proposed framework can provide diverse gripper designs.
Considering volume (cost), the designs obtained by the
framework perform better than the benchmark and show
competitive performance in the workspace.
Moreover, Figure 7 respectfully presents the benchmark
designs and optimized designs’ success or failure. It can
be observed from the figure that all the grippers fail to
pick object 2and designs 2and 3perform the same as the
benchmark design with a lower cost. All grippers perform
well in grasping simple primitive objects and achieve a
grasping success rate of no less than 0.7.
V. CONCLUSIONS AND FUTURE WORK
This paper presents a computational design framework
for exploring the evolution of soft robots in contact-based
grasping tasks. The study focuses on finger-based soft robotic
grippers and aims to bridge the gap in the current under-
standing of the design space and performance evaluation.
The proposed framework introduces a highly automated
design optimization process that utilizes a quality-diversity
optimization algorithm, which allows for the generation of
gripper diversity by exploring a significantly higher di-
mensional design space. A rapid and high-fidelity contact-
based simulation is proposed within the SOFA framework
to evaluate the performance of the designed grippers, which
accurately captures the complete soft grasping motion and
measures the grasping success rate for performance eval-
uation. The framework also comprehensively evaluates the
automated design optimization process using both simulation
experiments. The experimental discussion demonstrates the
potential of evolutionary techniques in producing diverse and
high-performance soft robotic grippers. The proposed frame-
work can be a valuable tool for researchers and designers in
soft robotics, enabling them to explore and optimize broad
design spaces for various or specific object grasping.
Future work may explore the effect of finger numbers
and more irregular configurations on more complex object
datasets. The dataset covers geometrically symmetric target
objects, making it easier for such grippers with symmetrical
arrangements in the gripper mount. More challenges are
Fig. 7: Comparison of the grasping performance
faced for objects with irregular shapes or different material
properties (mainly soft materials). Additionally, it is worth
investigating further verity the fidelity of the simulation
model by conducting physical experiments to verify the real
grasping performance. This may includes more experimental
characterisation of the optimised soft grippers, such as the
individual finger stiffness and overall gripper payload capac-
ity. Finally, the gripper frame space configuration can also
accommodate grippers equipped with two or more fingers.
VI. ACKN OW LED GE M EN T
We acknowledge Lois Liow’s support in designing the
object database.
REFERENCES
[1] F. Iida and C. Laschi, “Soft robotics: Challenges and perspectives,”
Procedia Computer Science, vol. 7, pp. 99–102, 2011.
[2] G. Gu, N. Zhang, C. Chen, H. Xu, and X. Zhu, “Soft robotics enables
neuroprosthetic hand design,” ACS nano, 2023.
[3] Z. Xie, A. G. Domel, N. An, C. Green, Z. Gong, T. Wang, E. M.
Knubben, J. C. Weaver, K. Bertoldi, and L. Wen, “Octopus arm-
inspired tapered soft actuators with suckers for improved grasping,”
Soft robotics, vol. 7, no. 5, pp. 639–648, 2020.
[4] R. Chen, Z. Yuan, J. Guo, L. Bai, X. Zhu, F. Liu, H. Pu, L. Xin,
Y. Peng, J. Luo et al., “Legless soft robots capable of rapid, continuous,
and steered jumping,” Nature Communications, vol. 12, no. 1, p. 7028,
2021.
[5] J. Shintake, V. Cacucciolo, D. Floreano, and H. Shea, “Soft robotic
grippers,” Advanced materials, vol. 30, no. 29, p. 1707035, 2018.
[6] S. Aracri, F. Giorgio-Serchi, G. Suaria, M. E. Sayed, M. P. Nemitz,
S. Mahon, and A. A. Stokes, “Soft robots for ocean exploration and
offshore operations: A perspective,” Soft Robotics, vol. 8, no. 6, pp.
625–639, 2021.
[7] X. Wang, H. Kang, H. Zhou, W. Au, M. Y. Wang, and C. Chen,
“Development and evaluation of a robust soft robotic gripper for apple
harvesting,” Computers and Electronics in Agriculture, vol. 204, p.
107552, 2023.
[8] X. Wang and H. Kang, “Soft robotic finger with variable effective
length enabled by an antagonistic constraint mechanism,” Smart Ma-
terials and Structures, vol. 32, no. 5, p. 055001, 2023.
[9] X. Wang, A. Khara, and C. Chen, “A soft pneumatic bistable rein-
forced actuator bioinspired by venus flytrap with enhanced grasping
capability,” Bioinspiration & Biomimetics, vol. 15, no. 5, p. 056017,
2020.
[10] J. Pinskier and D. Howard, “From bioinspiration to computer gen-
eration: Developments in autonomous soft robot design,” Advanced
Intelligent Systems, vol. 4, no. 1, p. 2100086, 2022.
[11] D. Howard, A. E. Eiben, D. F. Kennedy, J.-B. Mouret, P. Valencia,
and D. Winkler, “Evolving embodied intelligence from materials to
machines,” Nature Machine Intelligence, vol. 1, no. 1, pp. 12–19, 2019.
[12] G. D. Howard, J. Brett, J. O’Connor, J. Letchford, and G. W. Delaney,
“One-shot 3d-printed multimaterial soft robotic jamming grippers,”
Soft Robotics, vol. 9, no. 3, pp. 497–508, 2022.
[13] C.-H. Liu, T.-L. Chen, C.-H. Chiu, M.-C. Hsu, Y. Chen, T.-Y. Pai, W.-
G. Peng, and Y.-P. Chiang, “Optimal design of a soft robotic gripper
for grasping unknown objects,” Soft robotics, vol. 5, no. 4, pp. 452–
465, 2018.
[14] A. T. Mathew, I. M. B. Hmida, C. Armanini, F. Boyer, and F. Renda,
“Sorosim: A matlab toolbox for hybrid rigid-soft robots based on
the geometric variable-strain approach,” IEEE Robotics & Automation
Magazine, 2022.
[15] H. Zhang, M. Y. Wang, F. Chen, Y. Wang, A. S. Kumar, and J. Y.
Fuh, “Design and development of a soft gripper with topology opti-
mization,” in 2017 IEEE/RSJ international conference on intelligent
robots and systems (IROS). IEEE, 2017, pp. 6239–6244.
[16] N. Aage, E. Andreassen, and B. S. Lazarov, “Topology optimization
using petsc: An easy-to-use, fully parallel, open source topology opti-
mization framework,” Structural and Multidisciplinary Optimization,
vol. 51, pp. 565–572, 2015.
[17] J. Pinskier, P. Kumar, M. Langelaar, and D. Howard, “Automated
design of pneumatic soft grippers through design-dependent multi-
material topology optimization,” in 2023 IEEE International Confer-
ence on Soft Robotics (RoboSoft). IEEE, 2023, pp. 1–7.
[18] X. Wang, B. Wang, J. Pinskier, Y. Xie, J. Brett, R. Scalzo, and
D. Howard, “Fin-bayes: A multi-objective bayesian optimization
framework for soft robotic fingers,” Soft Robotics, 2024.
[19] M. A. Roa and R. Su ´
arez, “Grasp quality measures: review and
performance,” Autonomous robots, vol. 38, pp. 65–88, 2015.
[20] X. Shan and L. Birglen, “Modeling and analysis of soft robotic fingers
using the fin ray effect,” The International journal of robotics research,
vol. 39, no. 14, pp. 1686–1705, 2020.
[21] J. Allard, S. Cotin, F. Faure, P.-J. Bensoussan, F. Poyer, C. Duriez,
H. Delingette, and L. Grisoni, “Sofa-an open source framework for
medical simulation,” in MMVR 15-Medicine Meets Virtual Reality,
vol. 125. IOP Press, 2007, pp. 13–18.
[22] M. C. Fontaine, J. Togelius, S. Nikolaidis, and A. K. Hoover, “Covari-
ance matrix adaptation for the rapid illumination of behavior space,”
in Proceedings of the 2020 genetic and evolutionary computation
conference, 2020, pp. 94–102.
[23] N. Hansen and A. Ostermeier, “Completely derandomized self-
adaptation in evolution strategies,” Evol. Comput., vol. 9, no. 2, pp.
159–195, 2001.
[24] A. Cully, J. Clune, D. Tarapore, and J. Mouret, “Robots that can adapt
like animals,” Nat., vol. 521, no. 7553, pp. 503–507, 2015.
[25] S. Jain, S. Dontu, J. E. M. Teoh, and P. V. Y. Alvarado, “A multimodal,
reconfigurable workspace soft gripper for advanced grasping tasks,”
Soft Robotics, vol. 10, no. 3, pp. 527–544, 2023.
[26] B. Tjanaka, M. C. Fontaine, and S. Nikolaidis, “Quantifying efficiency
in quality diversity optimization.”
