RESEARCH ARTICLE

www.advintellsyst.com

Modular Origami Soft Robot with the Perception of

Interaction Force and Body Configuration
Junda Huang, Jianshu Zhou,* Zhengyan Wang, Jones Law, Hanwen Cao, Yichuan Li,
Hongbo Wang,* and Yunhui Liu

Given that robots are supposed to work

Bio-inspired soft robots provide a promising solution for robots working in around humans, robots have to interact
human-centered scenarios and interacting with unstructured environments. with humans frequently through versatile
However, the functional versatility and multimodal sensing of soft robots still approaches such as grasping, delivering,
and collaborating on the same task. The
need improvements. On one hand, the configuration of a soft robot is predefined
second perspective is robot–environment
during manufacturing; on the other hand, the multimodal perception of the interaction.[3] Robots move out from struc-
deformable soft actuator is challenging. In this work, a reconfigurable and tured factory scenarios and enter unstruc-
proprioceptive soft origami module is presented, where two kinds of basic tured environments, such as households,
actuation modes (i.e., extension and bending) are realized, and multimodal shops, and restaurants. Ensuring the safety
perception is enabled using a novel foldable self-inductance sensor. As a result, of robots and the working environment is
essential.
the origami module can be reconfigured to assemble multifunctional robots that
Soft robots, which emerged and became
can measure interaction force, body configuration, and other environmental popular in the past decade, have many
information. Dedicated experiments are performed to validate the performance of advantages over their rigid counterparts.
the proposed origami module. An intelligent gripper assembled using three First, soft robots are known for their
origami modules is designed with the capabilities of grasping mode adjustment, inherent compliance, which completely
grasping force measurement, and the grasping target’s size measurement. An guarantees safety during human–robot
and robot–environment interactions.[4–6]
intelligent jellyfish is assembled using five origami modules, and equipped with
Second, the systems of natural organisms,
buoyancy adjustment and underwater grasping capabilities. The proposed which are different from those of well-
proprioceptive modular soft origami provides an effective solution for versatile developed rigid robots that are connected
and intelligent soft robot design. by isolated joint-link chains, can perform
multiple functions simultaneously.[7–9]
Similar to their natural counterparts, soft
robots can perform the functions of actu-
1. Introduction ation, execution, sensing, and control systems; therefore, they
can be designed with a compact and lightweight structure to
One of the prominent frontiers of modern robotics is the devel- achieve tasks distinctively and efficiently compared with tradi-
opment of robots working in human-centered and unstructured tional rigid robots.
environments.[1,2] In these scenarios, safety is the foremost pri- Numerous achievements in soft robotic technology, which
ority when designing robot performance. Safety can be viewed offers an effective robotic solution for tackling challenging tasks,
from two perspectives. The first one is human-robot interaction. have been reported in the past decades. Examples include soft

J. Huang, J. Zhou, H. Cao, Y. Li, Y. Liu J. Zhou, J. Law, Y. Liu

Department of Mechanical and Automation Engineering Hong Kong Center for Logistics Robotics
The Chinese University of Hong Kong HKCLR, 5-th floor, 17W, Hong Kong Science Park, Hong Kong 999077,
HKCLR, 5-th floor, 17W, Hong Kong Science Park, Hong Kong 999077, Hong Kong
Hong Kong SAR
E-mail: [email protected] Z. Wang, H. Wang
Department of Precision Machinery and Precision Instrumentation
The ORCID identification number(s) for the author(s) of this article University of Science and Technology of China
can be found under https://doi.org/10.1002/aisy.202200081. Rm.707, Mechanics Building V, West Campus, USTC, Huangshan Rd. 443,
Hefei, Anhui Province 230027, China
© 2022 The Authors. Advanced Intelligent Systems published by Wiley- E-mail: [email protected]
VCH GmbH. This is an open access article under the terms of the
Creative Commons Attribution License, which permits use, distribution
and reproduction in any medium, provided the original work is
properly cited.
DOI: 10.1002/aisy.202200081

Adv. Intell. Syst. 2022, 2200081 2200081 (1 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

robotic grasping,[10–14] soft robotic locomotion,[15–17] and novel origami. Furthermore, a dedicated inductive sensor with a fold-
robotic applications enabled by soft robotic technology.[18] able structure is developed to be fitted and attached to the ori-
However, the functional versatility and intelligence of soft gami unit. The sensor can feedback its self-configuration by
robots still need to be improved. On the one hand, the configu- measuring its inductance variation after preliminary calibration.
ration of soft robots is highly predefined by the mechanism of the Embedded in the origami sensor, the configurations of the ori-
composed soft actuator.[5] Although versatile soft actuators have gami modules, such as length and bending angle, can be mea-
been developed, one kind of soft actuator is designed to be dedi- sured in real-time. In a pneumatic origami system, if the input
cated to a typical task. If the requirement of the task is changed, pressure and body configuration are measured, the interaction
the mechanism and fabrication of the soft actuator have to be force of the origami can be calculated through the system’s
adjusted accordingly. On the other hand, sensing is challenging closed-loop model.
for a deformable soft actuator.[19] The large body deformation As a result, the origami module can be reconfigured to build a
and versatile topology of soft robots preclude the leveraging of versatile robot assembly with functional versatility and environ-
well-developed sensors, such as encoders and force gauges, ment awareness. For example, the assembled robot with inte-
for rigid robots. Thus, achieving multimodal perception is much grated sensors can measure its interaction force, body
more difficult for soft robots than for their natural counterparts. configuration, and environmental information. Different quadri-
Natural soft creatures are environmentally aware due to their lateral folds (Table 1) are explored to demonstrate the generality
sensing system, which dramatically benefits their intelligent of our concept. Dedicated experiments are conducted to validate
behaviors requiring a quick response to environmental stimuli. the performance of the proposed origami module. Moreover, an
This study presents an effective solution to developing versa- intelligent gripper assembled with three origami modules (two
tile and intelligent soft robots. The modular robotic system has bending and one extension) is designed with the capabilities
addressed the versatility concern in robot design,[20–24] and the of grasping mode adjustment, grasping force estimation, and
origami structure has been proven to be an effective means to grasping the target’s size measurement. An intelligent jellyfish
resolve soft robotic sensing challenges.[25–28] with buoyancy adjustment and underwater grasping capabilities
A reconfigurable origami modular robot with interaction force are also assembled using five origami modules, namely one
and body configuration sensing is, therefore, proposed in this extension and four bending. The proposed proprioceptive mod-
study. Two kinds of origami modules endowing primary motion ular soft origami provides an effective solution for the design of
to the robotics system are defined. One is a bending module per- versatile and intelligent soft robots.
forming bending/rotation motion, and the other is an extension The main contributions of this work are as follows: 1) A novel
module capable of linear elongation and contraction. The motion reconfigurable origami robot with interaction force and body
of the origami module is enabled by a pneumatic-cable-hybrid configuration proprioception is proposed; 2) Two origami
(PCH)-driven approach toward the origami structure. modules, namely bending and extension units, are developed
Pneumatic actuation provides the primary motion of the origami, comprehensively with design, fabrication, and modeling possi-
and a passive-driven cable provides the motion constraint of the bilities for versatile applications; and 3) A novel inductive

Table 1. Major parameters of the origami unit.

Chamber Feature Linear Module 1 Linear Module 2 Bending Module

(Paw of the Gripper) (Jellyfish Head) (Jellyfish Finger and Gripper Finger)
Material Anycubic TPU 1.75 mm Anycubic TPU 1.75 mm Anycubic TPU 1.75 mm
(Anycubic 3D Corporation) (Anycubic 3D Corporation) (Anycubic 3D Corporation)
Outer diameter [mm]: D0 27 90 27
Initial outer diameter [mm]: d0 17 66 17
Inner diameter [mm] d d d
Wall thickness [mm]: t 1.2 1.5 1.2
Number of convolutions: N 5.5 3 5.5
Outer diameter/inner diameter ratio α α α
Initial length of one convolution [mm]: h0 10 24 10
Tip area [mm2 ]: S 289 4356 289
String length [mm]: l N/A N/A 26
Distance of elastic force [mm]: r i N/A N/A 7.25, 27.25
Linear constant: k 0.4 N/A 0.4
Material hardness Shore 83A Shore 83A Shore 83A
Young’s modulus [Mpa] 32 32 32
Extension ratio 1.95 2.5 N/A

Adv. Intell. Syst. 2022, 2200081 2200081 (2 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

origami sensor is built to provide the origami module the pro- 2. Reconfigurable Origami Robot and Its
prioception capability. The proposed sensor can effectively tackle Proprioception
the soft robot sensing challenge.
The remainder of this article is organized as follows. Section 2 2.1. Prismatic and Bending Origami Modules
introduces the concept and design of the modular origami
robot and its embedded foldable sensor. Section 3 describes We select the basic pleated fold pattern to design our origami
the modeling of the origami module and its enabled robots. module, as shown in Figure 1a. This pattern is usually adopted
Section 4 presents the experimental validation of the basic in the bellows design with different cross-sectional shapes. The
origami unit, including motion–force calibration and simple pleated feature provides a relatively adaptable surface to
proprioception. Two robots composed of origami units are also attach foldable sensors compared with complex origami patterns.
developed, and their grasping performance is demonstrated. It also has a large extension and compression ratio, thus allowing
The conclusions and future work directions are summarized for a rich behavior. As shown in Figure 1b, four pieces of a
in Section 5. pleated surface are connected to assemble the main body of

Figure 1. Concept of the proprioceptive origami module. a) Planar sensor. The origami module can be made from a piece of a coupled inductive planar
sensor. b) Concept of the folded origami module. c) Top and front views of one convolution of the origami module. d) Realization of the proprioceptive
origami module. The blue dashes mark possible sensor positions. e) Configuration of origami modules and reconfiguration between linear and bending
modules. f ) Sketch of the used rectangular planar coil, its design process, and its multi-folding states. The coil is designed in AutoCAD, imported into
Altium Designer, and fabricated with off-the-shelf soft PCB fabrication technology. The smaller the folding angle is, the smaller the inductance is.

Adv. Intell. Syst. 2022, 2200081 2200081 (3 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

Table 2. Main parameters in the 3D printing process. flux around it changes due to the interaction of the magnetic
field. According to Wang et al.,[32,33] the inductance of a planar
Parameters Values coil decreases when it folds. To fit the design of the origami mod-
Extruder temperature for TPU filament [°C] 230 ule proposed in this study, we predefine a long rectangular coil
with multiple folds (Figure 1f ). The multi-folded planar coil can
Extruder temperature for conductive filament [°C] 220
be utilized for proprioception of the origami actuator in conjunc-
Build platform temperature [°C] 40
tion with the origami structures. As validated experimentally in
Nozzle size [mm] 0.2 previous research,[32,33] such inductive sensors are velocity-
Layer thickness [mm] 0.1 independent, hysteresis-free, stable, sensitive, and repeatable, so
Shell thickness [mm] 1.6 they can provide reliable data for proprioception. Furthermore, the
Print speed [mm s1] 30
sensor is insensitive to the property of the coil material and the soft
materials to be embedded, so it ensures the performance of soft
Print infill density [%] 100
robots reconfigured for various application scenarios.
The sensor used in this work is fabricated via the standard
flexible printed circuit (FPC) process by an outsource printed cir-
the origami module. A sealed pneumatic chamber is achieved by cuit board (PCB) manufacturer (Huaxin PCB Express, Shenzhen,
sealing the two ends of the origami body. The length of the ori- China). The patterns of the FPC coil are created in AutoCAD on
gami can be adjusted by tuning the pneumatic input. For exam- the basis of the design parameters mentioned earlier and then
ple, when the pneumatic input is increased, the origami is imported into Altium Designer to form a layer out of the conduc-
extended, the valley of the pleated feature is raised, and the tive trace. The copper coil trace has a thickness of 35 μm and a
mountain feature goes down. On the contrary, a negative pneu- width of 100 μm, with 100 μm space between traces. The dielec-
matic input causes the origami chamber to shrink and the moun- tric (65 μm) and insulating (50 μm) layers are made of polyimide
tain and valley features to fold together. The geometric design (PI, also called Kapton) films.
parameters of the origami body are depicted in Figure 1c. The
profile of the body is rectangular and has customizable param- 2.3. Reconfiguration and Assembly of Origami Modules
eters. The body of the origami module can be easily scaled to
versatile applications. Reconfiguration is a transformation of function and structure
To achieve two kinds of basic motion, we adopt the PCH- that origami modules undergo repeatedly during their functional
driven mechanism, as shown in Figure 1d. The prismatic life. By manually adjusting the restriction cable, we achieve the
module is achieved by the pneumatic origami actuator, and the reconfiguration of the modular origami. To flexibly integrate the
bending module is realized by adding cable restriction to the pris- reconfigurable modules into soft robots, we use 3D-printed rigid
matic module, as shown in Figure 1e. Given that the linear mod- parts as the connection between different actuators. Thus, the
ule has no elastic force at its original length and the direction of the reconfigurable origami module contains three elements: the ori-
restoring force is opposite to that of displacement, it is compressed gami module, the folding sensor, and the auxiliary rigid parts
when ΔP < 0, as shown in Figure 1e. With a restriction cable (Figure 5a). Two applications of the assembly of the origami actu-
across the cable route, the origami actuator can produce bending ator are provided in this study: the intelligent two-finger gripper
motion toward the restricted side. The limiting cable has two func- (Figure 4) and the jellyfish (Figure 5). The intelligent two-finger
tions. One is to limit the maximum motion range, and the other is gripper uses a compliant extension actuator as its paw, two bend-
to provide bending motion. Hence, two basic motions can be ing actuators as its fingers, and a rigid component at the tip as its
achieved by the proposed origami modules. fingertip. The intelligent jellyfish consists of five origami mod-
Two fabrication approaches are usually adopted for origami ules, where the head has a high-extension-ratio extension actua-
units. One is molding, and the other is 3D printing.[29–31] In this tor and the four legs have four identical bending actuators.
study, we print our prototype origami model by using a commer-
cially available 3D printer, Raise3D-E2-Pro, with TPU material at
83 A hardness (Table 2). The major design parameters of the pro- 3. Origami Module Modeling
totype origami are presented in Table 1. 3.1. Extension Module Modeling

2.2. Foldable Inductive Sensor Design and Fabrication The kinematic model of the extension module is derived from a
static analysis with the help of traditional bellows modeling[34]
Proprioception of the origami unit is challenging due to its fold- and finite element method (FEM) simulation. When the exten-
able body. We propose a novel foldable sensor based on the self- sion actuator is at the equilibrium state, the actuated pressure,
inductance of multi-fold planar coils to address this issue. Novel the elasticity of the origami material, and the external force
properties, including multiple foldings, long rectangular coil achieve an overall force balance as follows
design, and multi-sensor coupling, are introduced into the fold- ΔPS þ F ex ¼ FðxÞ (1)
able sensor for origami fold adaptation and large range
extension. where S is the tip area of the actuator, F ex is the external normal
The self-inductance of a coil can be defined by the ratio of force, ΔP is the pressure difference, x is the axial displacement of
magnetic flux to current. When the coil is folded, the magnetic the extension actuator, and F(x) is the elastic force.

Adv. Intell. Syst. 2022, 2200081 2200081 (4 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

Elastic force can be expanded as Equation (4). The difference between origami and circular bel-
lows is that circular bellows have a uniform force distribution
FðxÞ ¼ σðxÞ þ F axis ðxÞ (2) 0
F in their cross-section (Figure 1b) because of their symmetrical
structure. By contrast, the origami cross-section is quadrilateral,
which originates from two sources: the stress (σ) due to material so the stress at the four corners of the cross profile varies at a
strain when the material of the origami is compressed or pressure of 40 kPa in the current study. We also find that the
stretched (Figure 2b) and the axial elastic force (F axis ) due to stress is uniformly distributed at a pressure of 0 kPa and approx-
the tendency of each fold to deform inward or outward when imately uniformly distributed at a pressure of 40 kPa according
the origami is compressed or stretched as a whole (Figure 2b, to the simulation results (Figure 2d). Accordingly, instead of
d). Each black arrow in Figure 2b refers to each fold’s deforma- treating Δθ as a unit like in ref. [34], we take Δy, an equal distance
tion tendency. differential unit shown in Figure 1c, for analysis and assume that
A previous work[34] obtained the axial elastic force of thin bel- the force in the cross-section is uniformly distributed because
lows on the basis of Castigliano’s theorem. We refer to this our model operates in the pressure range of 30–20 kPa
work’s method and derive axial elastic force F axis in (Figure 2f,g). Thus, it can be expressed as

Figure 2. Non-contact model assumption and validation. a) Simulation results of the origami module. b) XY-plane cross-view: evidence on the existence
of hypothetical σðxÞ. The black dashes represent the original shape of the module. When pressurized, the mountains and valleys exhibit a trend of
extension or compression. The arrows imply extension or compression. c) Static analysis of the extension actuator. d) XZ-plane cross-view.
e) Static analysis of the bending actuator. Parts 1 and of the extension actuator are simplified into Springs 1 and 2 of the bending actuator, respectively.
f ) Non-contact model validation of the extension actuator. g) Non-contact model validation of the bending actuator. Although parameter k of the term
σðxÞ is estimated from the extension actuator, the validation of the bending actuator demonstrates the consistency and effectiveness of this term.

Adv. Intell. Syst. 2022, 2200081 2200081 (5 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

0
0 F Δy ðr 1 þ r 2 Þ X2
ΔF ¼ (3) ΔPS þ M ex ¼ Fi r i (8)
4d 2 i¼1

By substituting Equation (3) into the method used in ref. [34], xi l

we obtain axial elastic force F axis and diameter ratio α. The param-  ¼ r i , i ¼ 1, 2 (9)
ϕ ϕ
eters can be found in Table 1.
where M ex is the external moment acting on the actuator, xi is the
4Et3
F axis ðxÞ ¼ ⋅x (4) displacement of the ith spring, r i is the distance from O to the
Nð1  μ2 Þd2 ðα  1Þ3 origin of elastic force F i , and ϕ is the bending angle. With
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the feedback of the inductive sensor, the closed-loop model
D h20 þ ðD0  dÞ2  ðNx þ h0 Þ2 can be expressed as the mapping hclose ∶ðΔP, sÞ ! ðMex , ϕÞ.
αðxÞ ¼ ¼ 1 þ (5)
d d Similarly, if the external moment M ex is 0, we can verify our
open-loop model, as shown in Figure 2g.
In addition, because ref.[34] did not consider the stress caused
by the strain of the material at the folds when the wall thickness
cannot be ignored, this study introduces a linear term to describe 4. Results
this material stress. By performing linear regression on the dif-
4.1. Kinematic Validation
ference between F elasticity and F axis , we obtain the value of linear
constant k.
To demonstrate that the proposed methodology can achieve ori-
σðxÞ ¼ kx (6) gami full-body actuation and kinematic estimation, we conduct
two basic validation experiments, namely position estimation of
When external force F ex does not exist, we obtain an open-loop the linear module and angle estimation of the bending module.
kinematic model ΔPS ¼ FðxÞ. To verify our explanation of ori- Two shallow single-coil neural networks (SC-NNs) are trained
gami motion, we compare the predictions of the open-loop separately to learn functions g close and hclose for estimation. In
model with the actual data in Figure 2f. The errors are enhanced the training process of the SC-NNs, we generate around 200
near the length boundary on both sides because material stress σ training data by blocking the position of the tip and pressurizing
is nonlinear, especially near the length boundary. the module. The ground-truth kinematic data, position x, and
The open-loop model no longer applies when the actuator angle ϕ used for training and comparison are captured by
interacts with the environment. To estimate the posture and OptiTrack, and the estimated position and angle are derived from
force during an interaction, we refer to Equation (1), which g close and hclose , respectively, on the basis of the pressure sensor
implements the mapping g∶ðΔP, xÞ ! F ex . If we introduce a and coil sensor input ðΔP, sÞ.
folding sensor, because the sensor implies displacement, our The validation process involves letting the module track ran-
closed-loop model can be expressed as the mapping dom air pressure control signals while randomly adding
g close ∶ðΔP, sÞ ! ðF ex , xÞ, where s is the feedback inductance of unknown pushing at the end of the module (Figure 3). With
the sensor. Furthermore, the folding sensor has the potential the linear module as an example, a random external force is pro-
to achieve multidimensional configuration perception by attach- vided at the end of the linear module without pushing it off its
ing numerous sensors onto different sides of the origami mod- trajectory. The interference can be described as an unknown F ex ,
ule (Figure 4d). Detailed content will be discussed in the where F ex 6¼ 0 in the grey area of Figure 3e.
intelligent gripper part. As shown in Figure 3e,g, the actuators are initially actuated
without contact in the white area and provided with external force
F ex or torque M ex in the gray area. In Period 1, F ex and M ex are in
3.2. Bending Module Modeling the opposite direction of motion direction x and ϕ, respectively;
in Period 2, the actuators are actuated by this external force with-
By restricting one side of the extension actuator, the origami out inflation. The accuracy of motion prediction demonstrates
module can achieve bending motion (Figure 1e). As shown in the effectiveness of proprioception in the origami modules.
Figure 2e, the bending actuator is simplified as two springs Although external interference is provided randomly and
by cutting the linear actuator model into two halves. The elastic diversely, the sensors in the origami module maintain high accu-
force of the bending origami is close to the longitudinal folds (red racy in the prediction process.
lines in Figure 2e) due to the 2D-like bending resulting from the
quadrilateral cross profile. The elastic force F i of each spring is
4.2. Force Validation
estimated to be half of that of their linear counterpart in
Equation (1).
Force prediction uses the same method as that for motion predic-
1 tion; however, the motion capture system is replaced by an exter-
F i ¼ Fðx i Þ, i ¼ 1, 2 (7) nal force sensor stuck at the end for ground-truth data. Unlike the
2
linear module, whose normal force is measured in the same direc-
On the basis of the moment balance in Figure 2e, if the fixed tion as its moving trajectory, the force of the bending module is the
end O of the string is selected as the center of rotation, we can normal force against a fingertip’s surface at a fixed point. Although
build a static model of the bending actuator as a single folding sensor only performs force prediction at a fixed

Adv. Intell. Syst. 2022, 2200081 2200081 (6 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

Figure 3. Experiment setup and proprioception validation. a) Experiment setup and data-collecting devices. The setup contains: 1) a solenoid valve,
2) inductance reading PCB, 3) pressure sensor, 4) OptiTrack camera, 5) PWM amplifier, and 6) Arduino Mega2560. b) Force collection process.
c) Pressure control system that contains (1), (3), (5), and (6) in (a) and regulated voltage power for 24 V valve input. d,f ) Force prediction for the
extension and bending actuators, respectively, where ΔP ¼ 0 in the white area, ΔP > 0 in the blue area, and ΔP < 0 in the green area. e,g)
Motion prediction for the extension and bending actuators, respectively, where F ex ¼ 0 in the white area and F ex 6¼ 0 in the gray area. Random external
force F ex is provided by hand in the manner shown in (b).

point, we believe force can be predicted wherever contact happens Figure 3d,f shows that the colored area is predicted accurately,
with the module if several sensors are coupled. and the prediction in the white area has a larger error than that in
As indicated in Figure 3d,f, ΔP ¼ 0 at the beginning of the the colored area. We assume that this result is induced by two
experiment in the white area, after which the actuator is pressur- factors: the collected data points near ΔP ¼ 0 are not sufficient
ized and compressed in turn. In Period 1, the actuator is pres- for SC-NN to predict precisely and neural networks should con-
surized in the blue area and then compressed in the green area sider past states. Given that the soft material has viscosity, the
without the pose changing; in Period 2, the pose and pressure are force measured near ΔP ¼ 0 can vary greatly due to the internal
varied. force of the soft material. Therefore, a recurrent neural network

Adv. Intell. Syst. 2022, 2200081 2200081 (7 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

that considers past states can be helpful in achieving precise folding sensor coupling, we propose an intelligent gripper,
force and motion prediction. whose paw is composed of an extension actuator and whose fin-
gers are composed of two bending actuators. A multi-coil neural
4.3. Intelligent Gripper Performance network (MC-NN) is trained for paw motion prediction because
the paw is embedded with three similar folding sensors, which
To synthetically demonstrate the: 1) modular capability of the ori- are coupled in a simple manner.
gami actuator, 2) flexibility brought by the combination of two Four representative objects, namely baseball, bleach cleaner,
motion patterns, and 3) advanced proprioceptive sensation with glass cleaner, and mini soccer ball, are selected from YCB

Figure 4. Intelligent soft gripper enabled by the origami module. a,b) Compliant gripper grasping different YCB objects in sequence. Because the objects
in (a-1), (a-2), and (a-3) are small in size, the paw actuator is compressed to fit the size of the objects, whereas in (a-4) and (a-5), the paw is pressurized for
grasping large and heavy objects. c) Noise and applied inductance. The measured inductance is filtered using a moving average filter with a 0.1 s window.
d) Design of the compliant paw. Three pieces of the folding sensor are attached to the extension actuator to demonstrate 3D pose estimation for (a).
e) Experimental verification to show the effectiveness of proprioceptive sensing in both bending and compliant multi-sensor extension actuators. Scales
are obtained from folding sensors, pressure sensors, and the pretrained neural network.

Adv. Intell. Syst. 2022, 2200081 2200081 (8 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

Benchmark with different shapes, sizes, and weights. The are enlarged prominently in Figure 4(a-4,a-5), which demon-
weights of the objects are between 54.5 and 186 g, and their strates the effectiveness of the model prediction of multi-sensor
grasping diameters are between 3.5 and 17.2 cm. The bottleneck sensing. Different grasping points can influence force prediction
of the glass cleaner in Figure 4(a-3) has the smallest diameter, on the object because the bending module predicts force at a
whereas the football in Figure 4(a-5) has the largest size and fixed point. Grasping force prediction and estimation can provide
weight. abundant information via dedicated processing, such as the tar-
In this experiment, the paw is actively pressurized to change get objects’ features (rigid or soft) and weight (heavy and light).
its length for object size adaptation, and the fingers are pressur-
ized and then deflated every grasping instance. By tuning the
pressure of the paw and fingers within the range of 40 to 4.4. Intelligent Jellyfish Performance
40 kPa, the intelligent gripper can perform powerful and compli-
ant grasping. We leverage the prediction model introduced in the In addition to the intelligent proprioceptive gripper, a jellyfish
validation section to estimate the pose and tip force of the intelli- can realize functionalities, such as propulsion and predation,
gent gripper on the basis of the feedback on the inductance and by smartly controlling the fluidic interaction, as indicated in
pressure of each module, as shown in Figure 4e. The dynamic the previous literature.[35] The jellyfish’s flexibility, compactness,
grasping performance and real-time visualization of propriocep- and energy efficiency inspire us to explore underwater locomo-
tion are demonstrated in the attached video. tion, grasping, and manipulation. Hence, an intelligent jellyfish
The paw length in Figure 4e is defined as the arc length is developed in this study, as shown in Figure 5a, on the basis of
between two central points of ends, and it is sketched as black our module origami approach to achieve desirable performance
bending lines in Figure 4a. The results of paw length prediction in terms of buoyancy control, propulsion, and grasping.

Figure 5. Intelligent origami jellyfish enabled by the origami module. a) 3D construction of the origami jellyfish. b) Moving forward underwater.
c) Origami jellyfish sinking down to grasp a stone and floating upward. d) Three different objects are grasped. e) Two cycles of buoyancy experiment
on the jellyfish head (buoyancy chamber). The measured inductance is filtered using a moving average filter with a 0.1 s window. f ) Repeatability results of
the folding sensor and inductance–buoyancy relation in 22 cycles.

Adv. Intell. Syst. 2022, 2200081 2200081 (9 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

The intelligent jellyfish is composed of one extension actuator than on pressure is prominently manifested as inductance
(buoyancy chamber), four bending actuators (soft fingers), two change (Figure 5e). Buoyancy prediction can be affected by this
propellers, and an embedded electronic system. For the exten- lag and noise, as shown in Figure 5f.
sion actuator, by carefully selecting the design parameters, such
as the number of convolutions, origami diameter, wall thickness,
and folding depth, and by adjusting the position of adjacent folds
with one-half convolution (Table 1), we can achieve highly stable
5. Conclusions and Future Work Directions
buoyancy control and an enlarged extension ratio (Figure 5e). This study demonstrates a reconfigurable origami module with
While the extension actuator realizes up/down and floating interaction force and body configuration perception. The origami
motions, two propellers supplement it with planar motion. model is designed using a basic pleated pattern with a large
Specifically, the jellyfish moves forward or backward when the deformation ratio. Two basic modules, one for linear motion
two propellers propel in the same direction and rotate around and one for bending motion, can create versatile robots with dif-
the central axis when they propel in opposite directions. To miti- ferent configurations and applications. A dedicated foldable sen-
gate data noise underwater and the inductance induced by the sor is proposed and attached to the origami module to achieve
long electric wire, an electronic system consisting of an
proprioception of the robot, and it dramatically enhances the
STM32 minimum system board, a 3.7 V lithium battery, and
intelligence of the system by perceiving the body configuration
an inductance reader is embedded into the waterproof chamber.
and environmental interaction force. Models are also developed
This near-end data collection setup prominently reduces data
to explain how origami modules are actuated by pneumatic actu-
noise and avoids the extra inductance induced by the electric
ation. Dedicated experiments are conducted to validate the basic
wire.
performance of the origami modules with integrated actuation,
The actuation of the origami modules and propellers is pro-
execution, and proprioception capabilities. Moreover, an intelli-
vided by an external source using connected ultra-soft air tubes
gent soft gripper and an intelligent jellyfish are developed based
and electric wires. A full locomotion and grasping process is
on the origami modules, and they demonstrate promising intel-
shown in Figure 5b–d. Buoyancy control is directly related to
ligent performance in grasping, locomotion, and environmental
module length, but the function between module length and
interaction estimation.
pressure is rheonomous underwater due to the viscosity.
In future work, the variable stiffness feature could be inte-
Given that the function between module length and inductance
is scleronomous, SC-NN maps the self-inductance to buoyancy, grated with the origami module for forceful interaction. An
and Figure 5e indirectly shows the lag between module length embedded miniature electronic system could also be used to col-
and pressure. By sending a desired buoyancy force to the jelly- lect data on origami modules with high accuracy, thus avoiding
fish, the controller compresses or pressurizes the extension mod- long-distance sensing data transmission noise. Furthermore, an
ule with the feedback from SC-NN, such that it changes the untethered actuation system could endow the origami module
buoyancy. Then, by dynamically adjusting the pressure in the jel- with a highly diverse application range.
lyfish head and enabling two propellers, we achieve forward
movement while keeping the jellyfish floating underwater
(Figure 5b). When the jellyfish reaches a stone, we pressurize Supporting Information
four soft fingers and the jellyfish head to float up with the grasped
object (maximum of 320 g underwater), as shown in Figure 5c. Supporting Information is available from the Wiley Online Library or from
the author.
The successful grasping of objects with different weights, sizes,
and shapes demonstrates the compliance and robustness of the
soft bending actuators, as shown in Figure 5d. Therefore, the
intelligent jellyfish can flexibly adjust its motion underwater Acknowledgements
and achieve compliant and powerful grasping of underwater
objects. J.H. and J.Z. contributed equally to this work. This work is supported in
part by the Hong Kong RGC via projects T42-409/18-R and 14202918, in
To explore the accuracy of buoyancy feedback, two repeatabil- part by CUHK Direct Grant 4055171, in part by the Hong Kong Centre for
ity experiments involving input stability and buoyancy offset are Logistics Robotics, in part by CUHK Hong Kong-Shenzhen Innovation and
conducted on the buoyancy chamber, as shown in Figures 5e,f, Technology Research Institute (Futian), and in part by the VC Fund
respectively. Ground-truth buoyant force data are collected by a 4930745 of the CUHK T Stone Robotics Institute.
force sensor stuck at the top surface of the buoyancy chamber
(Figure 5e). In the input stability experiment, two representative
cycles of sensory inputs (Figure 5e), namely measured pressure
and inductance, are selected from 22 compression and extension Conflict of Interest
cycles. Given that noise on inductance is enhanced greatly under- The authors declare no conflict of interest.
water, a simple moving average filter with a window length of
0.1 s is introduced for signal filtering. A comparison of induc-
tance noise between in-air and underwater setups is shown in
Figure 4c, where the same moving average filter is applied to Data Availability Statement
the in-air setup. Moreover, due to the water greatly magnifying The data that support the findings of this study are available from the cor-
the viscidity of the origami module, the lag on the length change responding author upon reasonable request.

Adv. Intell. Syst. 2022, 2200081 2200081 (10 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH
www.advancedsciencenews.com www.advintellsyst.com

Keywords [15] M. Calisti, G. Picardi, C. Laschi, J. R. Soc., Interface 2017, 14, 20170101.
[16] G. Gu, J. Zou, R. Zhao, X. Zhao, X. Zhu, Sci. Rob. 2018, 3, eaat2874.
compliant manipulation, inductive sensor, origami, proprioception, soft [17] Y. Li, F. Fish, Y. Chen, T. Ren, J. Zhou, Bioinspiration Biomimetics 2019,
robot 14, 066008.
Received: March 21, 2022 [18] E. W. Hawkes, L. H. Blumenschein, J. D. Greer, A. M. Okamura, Sci.
Revised: May 16, 2022 Rob. 2017, 2, eaan3028.
Published online: [19] H. Wang, M. Totaro, L. Beccai, Adv. Sci. 2018, 5, 1800541.
[20] S. Murata, H. Kurokawa, IEEE Rob. Autom. Mag. 2007, 14, 71.
[21] M. Yim, W. M. Shen, B. Salemi, D. Rus, M. Moll, H. Lipson,
E. Klavins, G. S. Chirikjian, IEEE Rob. Autom. Mag. 2007, 14, 43.
[1] G. Z. Yang, J. Bellingham, P. E. Dupont, P. Fischer, L. Floridi, R. Full,
[22] K. G. Sheela, P. J. Menon, S. Swetha, C. M. Vandana, R. Mendez,
N. Jacobstein, V. Kumar, M. McNutt, R. Merrifield, B. J. Nelson,
B. Scassellati, M. Taddeo, R. Taylor, M. Veloso, Z. L. Wang, Mater. Today:. Proc. 2020, 24, 1918.
R. Wood, Sci. Rob. 2018, 3, eaar7650. [23] J. Zhou, X. Chen, J. Li, Y. Tian, Z. Wang, in 2018 IEEE Inter. Conf. on
[2] E. W. Hawkes, C. Majidi, M. T. Tolley, Sci. Rob. 2021, 6, eabg6049. Soft Robotics (RoboSoft), IEEE, Piscataway, NJ, April 2018, pp.
[3] R. Drechsler, U. Kühne, in 1st Inter. Summer School on Methods and 412–417.
Tools for the Design of Digital Systems, Bremen, Germany, June 2015, [24] J. Zhou, J. Yi, Z. Wang, in 2018 IEEE Inter. Conf. on Real-time
https://link.springer.com/content/pdf/10.1007/978-3-658-09994-7. Computing and Robotics (RCAR), IEEE, Piscataway, NJ, August
pdf. 2018, pp. 411–416.
[4] S. Kim, C. Laschi, B. Trimmer, Trends Biotechnol. 2013, 31, 287. [25] E. A. Peraza-Hernandez, D. J. Hartl, R. J. Malak Jr, D. C. Lagoudas,
[5] C. Laschi, B. Mazzolai, M. Cianchetti, Sci. Rob. 2016, 1, eaah3690. Smart Mater. Struct. 2014, 23, 094001.
[6] D. Rus, M. T. Tolley, Nature 2015, 521, 467. [26] C. H. Belke, J. Paik, IEEE/ASME Trans. Mechatron. 2017, 22, 2153.
[7] C. A. Aubin, B. Gorissen, E. Milana, P. R. Buskohl, N. Lazarus, [27] D. Rus, M. T. Tolley, Nat. Rev. Mater. 2018, 3, 101.
G. A. Slipher, C. Keplinger, J. Bongard, F. Iida, J. A. Lewis, [28] Z. Chen, S. Siu, S. Tin, Y. Liu, H. Cao, Y. Gao, J. Zhou, Y. Liu, in 2021
R. F. Shepherd, Nature 2022, 602, 393. IEEE Inter. Conf. on Real-time Computing and Robotics (RCAR), IEEE,
[8] J. Zhou, Y. Chen, D. C. F. Li, Y. Gao, Y. Li, S. S. Cheng, F. Chen, Y. Liu, Piscataway, NJ, July 2021, pp. 1379–1384.
in 2020 IEEE/RSJ Inter. Conf. on Intelligent Robots and Systems (IROS), [29] J. Yi, X. Chen, C. Song, J. Zhou, Y. Liu, S. Liu, Z. Wang, IEEE Trans.
IEEE, Piscataway, NJ 2020, pp. 9159–9165. Rob. 2018, 35, 114.
[9] J. Zhou, Y. Li, Y. Yang, H. Cao, J. Huang, Y. Liu, in 2021 IEEE Inter. [30] Z. Xie, B. Chen, J. Liu, F. Yuan, Z. Shao, H. Yang, A. G. Domel,
Conf. on Real-time Computing and Robotics (RCAR), IEEE, Piscataway, J. Zhang, L. Wen, IEEE Rob. Autom. Mag. 2021, 28, 90.
NJ, July 2021, pp. 20–26. [31] T. TolleyMichael, F. ShepherdRobert, C. GallowayKevin,
[10] J. Zhou, Y. Chen, Y. Hu, Z. Wang, Y. Li, G. Gu, Y. Liu, Soft Rob. 2020, J. WoodRobert, M. WhitesidesGeorge, A Resilient, Untethered Soft
7, 743. Robot, Soft robotics 2014.
[11] J. Zhou, S. Chen, Z. Wang, IEEE Rob. Autom. Lett. 2017, 2, 2287. [32] H. Wang, M. Totaro, S. Veerapandian, M. Ilyas, M. Kong, U. Jeong,
[12] J. Zhou, J. Yi, X. Chen, Z. Liu, Z. Wang, IEEE Rob. Autom. Lett. 2018, 3, L. Beccai, Adv. Mater. Technol. 2020, 5, 2000659.
3379. [33] H. Wang, M. Totaro, A. A. Blandin, L. Beccai, in 2019 2nd IEEE Inter.
[13] J. Zhou, Y. Chen, X. Chen, Z. Wang, Y. Li, Y. Liu, IEEE Rob. Autom. Conf. on Soft Robotics (RoboSoft), IEEE, Piscataway, NJ, April 2019, pp.
Lett. 2020, 5, 1867. 242–248.
[14] J. Zhou, X. Chen, U. Chang, J. T. Lu, C. C. Y. Leung, Y. Chen, Y. Hu, [34] L. Wang, Z. Wang, Soft Rob. 2020, 7, 198.
Z. Wang, IEEE Access 2019, 7, 101483. [35] Z. Ren, W. Hu, X. Dong, M. Sitti, Nat. Commun. 2019, 10, 1.

Adv. Intell. Syst. 2022, 2200081 2200081 (11 of 11) © 2022 The Authors. Advanced Intelligent Systems published by Wiley-VCH GmbH