RESEARCH ARTICLE

www.advmattechnol.de

3D-Printed Phase-Change Artificial Muscles with

1. Introduction
Currently, additive manufacturing is utilized to fabricate many different
actuators suited for soft robots. However, an effective controller paradigm is Additive manufacturing (AM) has
essential to benefit from the advantages of soft robots in terms of power sparked the attention of engineers
and scientists due to its exceptional
consumption, production costs, weight, and safety while operating near living flexibility and ability to print compli-
systems. In this work, an artificial muscle is additively manufactured with soft cated structures.[1] The 4D printing
silicone elastomer material capable of demonstrating several levels of fabrications are capable of altering
stiffness. The 3D-printed muscle is equipped with carbon fibers to receive a their structure in response to specific
stimulus signal and develop a programmable joint that can present different environmental signals like electric
fields, temperature, magnetic fields,
stiffnesses. A nonlinear controller is developed to autonomously control the
and light.[2] This feature of 4D printing
variable stiffness joint based on a reinforcement learning algorithm. The makes it increasingly popular to fabricate
controller exhibits a slight increase in settling time; however, it demonstrates soft actuators and robots.[3] Additionally,
a decrease in fluctuation amplitude by 33% and a substantial reduction in the use of 4D printing technology offers
power consumption by 41% in comparison to the optimized proportional advantages for addressing a number of
integral derivative controller. At the same time, it is adaptable to and reliable factors, including decreasing overall fab-
rication costs and assembly time.[4,5] In
in new conditions. The variable stiffness muscle is also used as a controllable the developing 4D printing technology,
mechanism to suppress the low frequency vibration. The study shows that the researchers employ a variety of materials,
muscle can successfully attenuate the vibration autonomously when it is including magnetic material, hydrogels,
increased. and shape memory polymers (SMP).[6,7]
Traditional manufacturing techniques
have so far been used to create a
significant number of robots for a range of industrial applications
while high energy consumption and material waste rates are
two of these technologies’ most noteworthy flaws.[8] Researchers,
M. Mohammadi, A. Z. Kouzani, A. Zolfagharian however, are investigating new techniques and processes to man-
School of Engineering ufacture soft robots and actuators to address the drawbacks of
Deakin University existing rigid robots.[9,10] In comparison to other manufacturing
Geelong, Victoria 3216, Australia processes, 4D printing could be seen as being noticeably advanta-
E-mail: [email protected]
geous in producing soft robots. One of the many routine jobs per-
M. Bodaghi
Department of Engineering
formed by robots is frequently manipulating objects in different
School of Science and Technology shapes around them.[11] The 3D-printed soft actuators frequently
Nottingham Trent University resemble human limbs with several degrees of freedom (DoF).[12]
Nottingham NG11 8NS, UK These robots can be labeled as rigid or flexible limbs. Robotic ma-
Y. Xiang nipulators with frames composed of steel or aluminum are com-
School of Information Technology mon because of their high tensile strength. On the contrary, these
Deakin University
Geelong, Victoria 3216, Australia robots will be inflexible and hefty.
Growing demand for portable, lightweight robotic arms as well
as developments in material science and robotics, have inspired
The ORCID identification number(s) for the author(s) of this article research on the design, fabrication, and manipulation of flexible
can be found under https://doi.org/10.1002/admt.202300199
robotic arms. Flexibility is provided by utilizing plastic, compli-
© 2023 The Authors. Advanced Materials Technologies published by ant mechanisms, or carbon-fiber frames, which significantly re-
Wiley-VCH GmbH. This is an open access article under the terms of the
Creative Commons Attribution-NonCommercial License, which permits duce the cost of production and the robot’s energy consumption.
use, distribution and reproduction in any medium, provided the original Additionally, soft robotic grippers are significantly more efficient
work is properly cited and is not used for commercial purposes. at handling fragile or irregularly shaped objects than traditional
DOI: 10.1002/admt.202300199 rigid grippers. The main limitations of such systems are their

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (1 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 1. a) The phase transitioning muscle in low/high-pressure conditions. b) Modified printer for silicone mandrel printing c) Silicone cylinder
printed on the mandrel (Step 1). Covering the silicone cylinder with extendable braided sleeve (Step 2). Printing silicone layers to cover the sleeve (Step
3). Removing the mandrel, inserting the carbon fiber, and filling the chamber with ethanol (Step 4).

limited range of gripping power and inability to execute diverse activated muscle by filling the muscle’s bladder with a liquid-to-
tasks in diverse surroundings. Researchers have proposed vari- gas phase transition liquid and reached 20% contraction.[24] A
able stiffness structures and specific designs to overcome these flexible phase transition muscle is 3D-printed and evaluated in
limitations.[13–15] One of the most popular soft actuators is pneu- this paper.
matic actuator, which can provide a high force-to-weight ratio by As a result of the robotic arm’s flexibility, there can be vari-
adjusting the pressure inside the actuator chamber.[16] In addi- ous levels of inaccuracy, such as more vibration before settling,
tion, this actuator can provide variable stiffness when the cham- especially close to the robotic arm’s end. Rigid manipulator con-
ber pressure is changed. Figure 1a shows a pneumatic cham- trol methods do not perform well with flexible manipulators. In
ber in a low and high-pressure situation, respectively. When the order to handle the time variant flexible robotic arms, unique al-
stretching force is applied to the muscle, it converts to the com- gorithms that are exact and efficient must be developed. The two
pressing force on the fluid inside the chamber. Therefore, the most widely utilized linear control approaches are proportional
required force to stretch the muscle to a specific amount varies integral derivative (PID) and linear quadratic regulator (LQR), yet
in different conditions. they are not preferred due to the nonlinearity and time variance
The additive-manufactured soft pneumatic actuators (SPAs) of the soft variable stiffness actuator.[25] Using full-state feedback,
can produce complex movements, including jumping, bend- Akyuz et al. proposed a PID control law to manage a robot arm
ing, and twisting, in response to the changing pressure in their with a flexible joint.[26] The PID control methodology can also
chamber.[17] In refs. [18–21], the authors develop a pneumatic be combined with various control techniques to develop a more
actuator that activates through high-pressure fluid provided via reliable controller and boost system efficiency.[27] The LQR is an-
pumping through AM to reduce manufacturing time and cost other significant and popular technique for controlling robotic
while providing various features, such as self-healing, precise arm systems. Since LQR seeks to optimize the quadratic cost, it
actuation, and size customization. The flexible joints can be has gained much popularity due to its simplicity, optimum per-
equipped with SPAs to provide multi-level variable stiffness joint formance, versatility, and analytical solution. However, in terms
autonomously. However, the integration of pump and pipes into of trajectory tracking, LQR and PID replies are equivalent.[28]
the SPAs systems often results in increased weight and bulk. Re- Nevertheless, the core controller for the industry still uses linear
searchers have proposed the implementation of phase transition approaches because of their simplicity.[29]
muscles (PTM) as a mean to eliminate the pump and produce the Other than PID and LQR, researchers investigated a variety
required pressure through phase transitions. Sangian et al. pro- of nonlinear techniques to develop a controller for flexible joint
posed an innovative phase change muscle filled with paraffin that robots, including fuzzy logic,[30–33] neural networks (NN),[34] and
could be activated thermally and can achieve 9% contraction.[22] neuro-fuzzy logic.[35] Additionally, flexible connections and joint
Another thermally activated muscle was proposed that is filled link mechanisms are controlled by the RL approaches.[36–39] RL is
with ammonia liquid with an embedded heating element to stim- a technique used in the control area to create adaptive controllers
ulate 17% contraction.[23] Mirvakili et al. introduced magnetically that could learn in real-time and become close to the desired

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (2 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

controller. A reward-penalty strategy is used by RL to respond bioprinter—CELLINK is modified to print the silicone chambers
to the input and feedback signal to control the system. The con- of the PTM (Figure 1b and Movie S1, Supporting Information).
troller modifies its behavior to maximize rewards. This approach First, two 16-gauge syringes are filled with Ecoflex 00–50 parts
makes it possible to design a type of adaptive controller with- A and B separately with thivex additive and loaded into the sy-
out having complete system dynamics information where the ringe pumps. The pump pushes the material through the vinyl
optimized controller can be developed by solving the Hamilton– tubes to the static mixer, which is fixed on the printer instead of
Jacobi–Bellman equations online.[38] the main nozzle. The static mixer allows input flows to be prop-
Despite the extensive research on 3D printing of various mate- erly mixed and releases the Ecoflex mixture at the output through
rials, including silicone, the 3D printing of silicone on a mandrel the dispensing needle with a 1.27 mm inner diameter. Thin lay-
is a novel area of research with its own unique set of challenges. ers of silicone are printed on the mandrel, while a stepper motor
One major challenge is the viscosity of the silicone; if it is too low, controls the mandrel. The needle was moved parallel to the man-
the silicone may become runny, while if it is too high, it requires drel axis at a fixed distance from the surface, while the mandrel
excessive pressure to be extruded. Additionally, controlling the is rotated at a constant speed to deposit multiple silicone layers
movement of the mandrel and tailoring the G-code to the man- (Figure 1c-Step 1). Then an expandable braided sleeve is manu-
drel system is also a challenge. Another gap in this field is the ally worn on top of the layers (Figure 1c-Step 2), and the printing
development of a proper controller for non-linear systems, such of the silicone is continued to cover the sleeve (Figure 1c-Step 3).
as soft joint robots. Deep reinforcement controllers (DRL) have The braided sleeve is able to convert the expansion force in the
been identified as a significant solution, however, they also come chamber to the contraction force.
with their own challenges, including the appropriate sizing of the 3D printing of the chamber offers several advantages over tra-
deep neural network (DNN), which must balance between satis- ditional manufacturing methods when creating a flexible phase
fying the complexity of the system and having the desired level of transitioning muscle. First, it decreases the cost of manufactur-
flexibility at the output. Furthermore, designing an effective re- ing by eliminating the need for casting molds and using less ma-
ward function is critical for a successful reinforcement controller. terial. Second, the structure of the muscle is customized using
This research aims to develop innovative mechanisms with 3D printing. For example, instead of uniform wall thickness, the
variable stiffness characteristics through additive manufacturing thickness of the wall is changed in specific parts to provide more
(AM) and address the control challenges associated with soft and complex movement instead of just linear contraction. This would
flexible structures. The unique aspect of this work lies in the com- typically require the use of various cast shapes for different move-
bination of three key elements: 3D printing the silicone cham- ments, which is both time-consuming and expensive to manufac-
ber on a mandrel, developing a reinforcement learning (RL) con- ture. Last, the expandable braided sleeve can be printed simulta-
troller for variable stiffness structures, and utilizing the structure neously with the silicone printing with a different pattern. This
in attenuating vibration. By utilizing 3D printing, the design of allows for the creation of intricate designs that would be difficult
the PTM is customizable, the cost of manufacturing is signifi- to achieve with traditional methods.
cantly reduced, and achieving different stiffness levels is easily After printing, the 3D-printed silicone chamber is filled with
attainable by changing the pressure inside the printed actuator ethanol, and the carbon fiber is inserted into the chamber (Fig-
chamber. The development of an RL controller is an innovative ure 1c-Step 4). The carbon fiber heating wire with 33 ohm m−1
approach to controlling compliant systems, which allows for im- resistance that can endure up to 200 °C. By integrating the carbon
proved adaptability and reliability. Additionally, this work aims fiber, the muscle is capable of being stimulated with low voltage
to evaluate the PTM’s capability to suppress low-frequency vibra- to evaporate ethanol and increase the chamber’s pressure. In the
tions similar to human limb tremors, which is a novel application last step, both sides of the chamber are sealed with a thicker layer
for this type of technology. Overall, this research presents a novel of silicone, 90% of the chamber is filled with ethanol, and 10 cm
approach to designing and controlling flexible manipulators and of the carbon fiber is inserted into a 7 cm length chamber. The
compliant systems, utilizing 3D printing and RL controllers to internal diameter of the muscle is 7 mm, the outer diameter is
achieve variable stiffness characteristics and suppress vibrations. 10 mm, the length is 70 mm, and ≈2.5 mL of ethanol is poured
into the muscle.
Various experimental processes are conducted to assess the
2. Results and Discussion performance of the additively manufactured artificial muscle. In
the first experiment, one end of the muscle is fixed, and the other
This paper offers a technique for utilizing the DRL method to de- end is relaxed (Figure 2a). Then a constant voltage is applied to
sign a controller to improve the precision of trajectory tracking in examine the maximum contraction of the muscle. In another ex-
a robotic arm with a 3D-printed variable stiffness joint. Addition- periment, a weight equal to 130 g is attached to the relaxed end
ally, the intention is to reduce power consumption, arm vibration, to evaluate the contraction of the muscle under an external force
and track errors. Furthermore, the variable stiffness structure is equal to 1.3 N (Figure 2b).
exploited to attenuate the limb’s tremor upon demand. The ex- In the first experiment, the relaxed muscle is to reach a 28.6%
perimental setup is assembled by a motor-driven rigid beam that contraction. This contraction is obtained in 56 s when the muscle
is equipped with two 3D-printed PTMs to provide flexibility to is activated by 5 V. In addition, the same contraction occurs in 39 s
the joint. The PTMs can express variable stiffness to increase while 8 V is applied to the muscle. The extension and contrac-
the range of flexibility of the joint. The PTMs chamber is filled tion of the muscle are repeatable without labor intervention. The
with ethanol and applying a low voltage to the integrated car- experiment is repeated ten times, and the muscle contraction is
bon fiber string can change the chamber’s stiffness. BIO X—3D recorded. Figure 2a shows the boundaries and the average value

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (3 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 2. a) Contraction experiment in relaxed condition. b) Loaded contraction. The experiment for both conditions is performed under 5 and 8 V.
c) Durability of five muscles in 10 days d) stiffness of the muscle.

of the contraction for this experiment. In the experiment with the sequently, 32 s is needed to reach 64 from 58 mm. For the full
130 g weight, because of the 1.3 N external force (Figure 2b), the recovery of the muscle from 53 to 70 mm, 110 s is required
contraction of the muscle decreases to 24%. The muscle reaches on average. This time depends on the environment’s tempera-
this contraction in 61 and 41 s for 5 and 8 V, respectively. ture and could be varied by changing the surrounding temper-
The muscle is not equipped with an active cooling system. ature, and this experiment is conducted at room temperature
Therefore, the recovery of the muscle to its initial length is slower (21 °C). Additionally, the contraction time is also dependent on
compared to its contraction. 33% of the recovery is achieved in environment temperature, and increasing environment temper-
18 s, meaning the muscle length reaches 58 from 53 mm. Sub- ature can decrease contraction time. Based on the carbon fiber

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (4 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

characteristics and the experiment’s time, the muscle’s power In the simulation environment (Supporting Information, Ap-
consumption can be analyzed. pendix 4), the DDPG and TD3 are set to be trained. The reward
The consumed energy of the muscle can be calculated based function is calculated at every step, and the accumulated reward
2
on W= VR t, while V is the applied voltage, R is the resistance of of these steps is calculated in every episode. The accumulated re-
the wire, and t is the time for contraction. The resistance of 10 cm ward is displayed in Figure 3 for both agents. The blue line shows
wire is 3.3 ohm; therefore, the required energy in the first exper- every episode’s reward, and to see the training trend in the train-
iment is 424 and 756 J for 5 and 8 V, respectively. Also, 462 and ing phase (Figure 3a,c), a moving average filter with a length of
795 J are needed in the second experiment to reach 24% of con- 50 is applied to the data shown in orange color. The trends in
traction while the 5 and 8 V are applied, respectively. The results Figure 3a,c are both upward trends that indicate the controllers
show that even though the contraction is obtained faster by ap- are improving their ability to control the system and acquire more
plying higher voltages, the amount of consumed energy is higher. rewards. In other words, the controllers follow the desired track
Therefore, the applied voltage can be chosen based on the appli- better and decrease the vibration of the beam at the same time.
cation of the muscle. If a higher speed of contraction is needed, The reward function is used on the PID controller to calculate
higher voltages must be applied. Nonetheless, if the muscle the accumulated reward for this controller to make a stop crite-
must be run on low-capacity batteries, the voltage can be de- rion. Therefore, the calculated amount is used to stop training
creased to reduce power consumption. However, the limitations when the RL controllers can collect more rewards in one episode
of carbon fiber in handling high voltages and currents must be than the PID controller. As shown in Figure 3, the accumulated
considered. reward sometimes drops because the agent is exploring various
A power regulator was utilized to power the controller (Ar- options to find the best way to control the system.
duino) and the PTM from a portable power source. This enables After training the RL agents (DDPG and TD3), the controllers
the precise control of the voltage applied to the muscle by control- apply to control the Quanser system to follow the step function,
ling the power switch (TIP120 transistor), allowing for a more ac- and the motor angle is shown in Figure S3a, Supporting Infor-
curate and consistent performance. The power consumption of mation (Appendix 1) compared to the response of the PID con-
the muscle is 500 J per cycle on average, and a 9 V alkaline bat- troller. DDPG and PID controllers are outperformed by the TD3
tery stores around 20 000 J. This amount of energy is sufficient controller in terms of the rise time (1.27, 1.24, and 1.21 s, re-
to power the muscle for 40 cycles. In practice, the battery pro- spectively). However, the PID settles more quickly (1.38 s) than
vides 30 cycles on average (Figure S1, Supporting Information). alternative controllers (DDPG: 1.44 s, TD3: 1.62 s).
Lithium-ion (Li-ion) batteries are better options because they are The manipulator fluctuation angle is shown in Figure S3b,
rechargeable and store a higher amount of energy with a similar Supporting Information (Appendix 1) for all the controllers.
weight. Another muscle parameter is the outer temperature of Although the RL controllers’ fluctuation persists longer, the
the muscle, which is important in wearable applications. During PID-controlled manipulator’s vibration has a higher peak
the experiment with the weight, the external side temperature (PID:0.18 rad, DDPG: 0.10 rad, TD3: 0.08 rad). Additionally, RL
of the muscles is recorded. The surface’s temperature increases algorithms require less action (applied voltage on the motor)
gradually, reaching 40 °C when the muscle has a 12 mm con- than the PID controller. The total of the squared voltage values
traction equal to 17% contraction. When the heating continues (that is directly proportional to the power consumption) for TD3
to get a higher contraction, the soft muscle finally obtains a 24% and DDPG is equal to 1 and 1.25, respectively, in terms of power
contraction while the surface temperature reaches 50 °C. Even consumption. The PID controller’s consumption is proportional
though this temperature is uncomfortable if it touches bare skin, to 1.97, which is nearly twice as much as the required power for
it could be utilized with a lower contraction rate or worn into fab- the TD3 controller.
rics to avoid direct touch. The joint stiffness is reduced to evaluate the controllers’ be-
In order to assess the stiffness of the muscle, tensile tests were havior under the new circumstances after investigating the effec-
performed on the PTM under varying states of contraction. The tiveness of various controllers on the system. The PID and TD3
results of these tests, depicted in Figure 2d, demonstrate the mus- controllers’ responses are displayed in Figure S4, Supporting In-
cle’s response in 5, 10, and 15 mm contractions. Higher levels of formation (Appendix 1). The controllers produce an acceptable
contraction could be achieved by increasing the pressure within outcome despite being adjusted to various conditions. In con-
the chamber, resulting in greater levels of stiffness. Notably, the trast to the PID controller, the TD3 has a better response time.
stiffness of the muscle at low levels of contraction (5 mm) is The RL agent overshoots by 8.2%, whereas the linear controller
110 N m−1 , while a stiffness of 700 N m−1 was measured in aver- overshoots by 11.8%. Additionally, TD3 has a rise time of 1.01 s,
age for the muscle exhibiting 20 mm contraction which means its which is faster than the PID controller’s 1.20 s. Additionally, the
stiffness increases 6.36 times. Table S1, Supporting Information manipulator’s oscillation boundary for both controllers is raised.
shows various variable stiffness materials compared to the PTM. The RL agent can learn and adjust to new circumstances re-
The durability of the PTM is a crucial factor that demands evalua- gardless of the outcome depicted in Figure S4, Supporting Infor-
tion owing to the silicone’s permeability. To this end, five muscles mation (Appendix 1). As a result, the agents were retrained, and
were produced, and the contraction experiment (unloaded) was Figure 3b,d depict the episode reward from the retraining phase.
conducted for 10 days. The contraction percentage was recorded In a few episodes, RL agents can adjust to the new system. The
during this period, as depicted in Figure 2c, which revealed a TD3 is evaluated after retraining in order to contrast with the PID
decreasing trend from 28% to 24%. Although the PTM demon- controller. The system’s reaction after using the step function is
strated substantial contraction, its permeability must be recon- seen in Figure 4a shows that the DRL controller performs better
sidered in future research. than the result in Figure S4, Supporting Information (Appendix

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (5 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 3. The accumulated reward of the agent during the training and retraining phase. a) Episode accumulated reward during training for the DDPG
agent. b) Retraining episode reward for the DDPG. c) Episode accumulated reward during training for the TD3. d) Retraining episode reward for the TD3.

Figure 4. PID and TD3 controllers’ performance after decreasing the stiffness of the joint and retraining the RL agent. a) Motor angle under TD3 and
PID controller to track the reference signal. b) Robot’s arm oscillation.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (6 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

1), while the PID controller remains unchanged. As a result, the is adjusted following the pressure change. In order to effectively
outcome is consistent with the previous experiment, but the RL control the system during this experiment, the TD3 algorithm’s
controller’s rising time is improved. The link settles more quickly settings are set to learn the new condition online.
than in the previous edition, which means that the retraining im- Figure S5, Supporting Information (Appendix 1) demonstrates
proved the link’s oscillation. that the retrained TD3 can control the system more accurately
To assess the muscle as a variable stiffness mechanism and than the PID controller because its parameters suit the new con-
the developed RL controller in an experiment, a setup, including dition. Compared to the PID controller, the TD3 controller is
five main parts is exploited. The setup includes 1) Rotary Servo quicker and has a shorter rise time (0.21 and 0.26 s, respectively).
Plant: Quanser SRV02 and Rotary Flexible Joint: Quanser ROT- Additionally, the PID controller’s overshoot is around 5%, and
FLEX Module. 2) Power Amplifier: Quanser VoltPaq-X1. 3) Mat- the RL agent has a 6.4% overshoot. In the new experiment, the
lab/Simulink software. 4) Data Acquisition Board: Quanser Q2- muscle’s temperature is reduced, therefore, the joint’s stiffness
USB and 5-power supply (Figure 5a). The spring is switched out is reduced compared to its original state, and it causes a more sig-
for the 3D-printed silicone muscle with a variable stiffness mech- nificant fluctuation of the joint under the supervision of the un-
anism that is capable of changing the stiffness directly by apply- changed PID controller (see Movie S3, Supporting Information).
ing voltage to the integrated carbon fiber. The experiment uses However, it still dissipates more quickly than the oscillation con-
a sampling rate of 40 samples per second (0.025 s). The incre- nected to the RL controller. Once more, when the RL controller is
mental encoder sensor, which is a part of the Quanser unit, may in control of the system, the link’s fluctuation magnitude is less
generate up to 4096 pulses each cycle and is used to calculate the severe. The voltage signal provided to the amplifier in the exper-
position of the rigid arm. The rotator and the rigid arm displace- iment served as the basis for calculating power usage. Based on
ment and angular velocity are provided to the RL algorithm as the the signal, the PID controller’s power consumption is 1.69 times
feedback signal. Critic and actor DNN have a similar architecture, more than the TD3 controllers.
including two hidden layers with 20 and 10 nodes each that can This study shows that the variable stiffness PTM gives our sys-
provide a suitable controlling signal for the system. tem several levels of stiffness without requiring human labor to
Figure 6a shows the motor angle as the controller tries to track take the structure apart and add new components. Despite its
the step function (see Movie S2, Supporting Information), in this benefits, there are still a few issues that could be covered in the
case rotating the arm by 45° (or 𝜋/4 rad) as part of the experiment. following works. The next phase in the 4D printing process in-
According to the graph, the rise time for the PID controller is volves co-printing the carbon fiber to decrease human interven-
0.22 s, which is almost equal to the rise time for the TD3. The tion in the fabrication process of the muscle. Another obstacle
rise time for the DDPG controller is 0.27 s, making it slower than that might need to be addressed in order to incorporate the vari-
the other controllers. The overshoot for DDPG controllers is neg- able stiffness structure into wearable technology is replacing the
ligible, but the TD3 and PID controllers show an overshoot of ethanol with materials with a lower boiling point temperature;
7% and 2%, respectively. therefore, the muscle could be activated at lower temperatures,
Figure 6b depicts the rigid arm’s oscillation. It shows that PID which leads to lower energy consumption as well. The improved
controller vibration in the joint will be dampened sooner than performance was attained by creating the DRL agent to control
the other approaches, comparable to the simulation. However, the variable stiffness joint. Without human involvement, the sys-
the RL controllers’ oscillation magnitude, particularly that gener- tem was able to adjust to new settings. The controller also con-
ated by the DDPG agent, is less than the PID’s response. Then, siderably reduces power usage, which is a huge benefit for using
the experiment is repeated with a different spring stiffness by it in portable devices.
applying the stimulus signal through the carbon fiber to assess By combining the adaptive controller with the time variable
the effectiveness of the controller under various conditions. The system, the output may be used in a broad range of applications.
soft muscle’s temperature is changed, and then the pressure in- In order to gain higher efficiency, most wearable devices, includ-
side the silicone’s chamber is altered. The stiffness of the muscle ing prosthetics, rehabilitation equipment, and assistant systems,

Figure 5. The experimental setup 1) Laptop/Matlab software. 2) Data Acquisition Board. 3) Power Amplifier. 4) Flexible joint Quanser unit equipped
with the PTM. 5) Power supply.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (7 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 6. a) Motor’s angle response to step function controlled under linear and nonlinear controllers in the experiment. b) Manipulator’s endpoint
fluctuation angle in the step response under linear and nonlinear controllers in the experiment.

must include a variable stiffness mechanism. The individuals making it more comfortable and user-friendly. Second, the de-
who are wearing the gadget want their organs to show differ- sign is more general and can be customized for specific tremors
ent features in various situations; therefore, the gadget must be a in different parts of the upper limb, including flexion–extension,
time-variant structure. Another issue is the amount of power that radial–ulnar, or pronation–supination motion.[40] In individuals
wearable technology consumes.[40] In comparison to a conven- with flexion–extension tremor, for instance, the muscles associ-
tional controller, the device might run longer with the designed ated with the tremor are located between the forearm and hand
nonlinear controller suggested in this research while using the on the dorsal and palm sides, and the artificial muscle is placed
same quantity of attached battery. parallel to that. When the tremor is detected, the controller stimu-
An application of the PTM component with variable stiffness lates the artificial muscle to achieve higher stiffness, which dissi-
in human limb orthosis is shown in Figure 7 to suppress upper pates the oscillating movement of the muscle similar to ref. [41].
limb tremors. The variable stiffness glove can be actuated by ap- Although the mechanism may affect voluntary motion, the vol-
plying electric power to the silicone muscles, and RL agents con- untary force is typically higher than the tremor forces. Therefore,
trol their temperature while the artificial soft muscles are placed the individual can still perform their task while the tremor force
alongside the limb muscle. The gadget is more user-friendly be- is dissipating. Regarding the voltage control of the PTM, the con-
cause of its lightweight and softness. While the upper limb starts troller obtains feedback from the tremors and the PTM temper-
vibrating, the temperature is raised to increase stiffness in the ature utilizing sensors. When a tremor occurs, the controller ad-
structure. But increasing the stiffness, the passive force of the justs the voltage on the muscle via an adjustable voltage booster
artificial muscle will decrease the tremor of the arm. A signif- to increase its stiffness. When the tremor subsides to a reason-
icant problem with wearable technologies is their weight.[40] A able level, the controller maintains the pressure by controlling
lightweight wearable gadget can be achieved by reducing the de- the temperature.
vice’s power consumption, which also reduces the bulk of the On the same setup, a different experiment is done to test
battery. Additionally, the design of the PTM offers a distinct ad- how well the muscle stops the low frequency vibration similar
vantage over pneumatic muscles in terms of weight and volume to human arm tremor. In this experiment, the silicone muscle
reduction. This is due to the absence of the need for fluid pump is added to the springs in the parallel position instead of replac-
connections. ing them on the Quanser flexible joint unit. Upper limb tremors
Compared to traditional tremor suppression gloves, this glove can occur at various frequencies and amplitudes. However, the
has several advantages. First, it weighs less and is more flexible, most common frequency is 4 Hz. Therefore, in the experiments,

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (8 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 7. Variable stiffness PTM glove to suppress tremors in forearm and hand.

various sinusoidal signals with 4 Hz frequency and two different cause of its flexible and friendly structure while it has a small vol-
amplitudes are applied to the motor to generate low frequency ume and being lightweight. Nevertheless, 3D-printed soft robots
vibration to mimic human arm tremor. When the motor’s gear require further development before they can be used as ubiqui-
starts to move, the muscles are in an idle state. As a result, the tous technology. Exploration of the field of AM and soft robotics,
rigid arm can move in a semi-sinusoidal motion with the pri- which is viewed as a breakthrough technique, will help disclose
mary frequency equal to the motor (4 Hz). The artificial muscles new possibilities in various industries in the coming years. Com-
are activated to suppress the vibration in the rigid arm. The acti- pared to rigid actuators, soft actuators are more sophisticated due
vated muscle makes the joint stiffer, which causes a decrease in to their high degrees of freedom. In order to use soft robotics in
the vibration of the Quanser unit arm, as shown in Figure 8 and a wide range of applications, new controlling paradigms need to
Figures S6 and S7, Supporting Information (Appendix 1). be investigated together with the soft robot’s development.
Figure 8a,b show that when the muscle is added to the flexible
joint with 15° vibration amplitude, even in idle mode, it affects 3. Conclusion
the vibration amplitude. Nevertheless, compared to the suppres-
sion in the active mode is less than half. The result shows 21.75% This article discussed the manufacturing of flexible silicone mus-
and 47.32% suppression for idle and active mode, respectively. cle through additive manufacturing and evaluated its features, in-
Figure S6a,b, Supporting Information (Appendix 1) depict the cluding contraction, repeatability, recovery time, and power con-
repeatability of the muscle. The difference between various tri- sumption. The 3D-pritned silicone muscle was integrated with
als is negligible, with an equal amount of vibration attenuating. carbon fiber string; therefore, it can be activated with a low-
Figure S6c,d, Supporting Information (Appendix 1) illustrate the voltage signal as the stimulus signal. The variable stiffness char-
idle mode of the muscle before activating and after cooling down acteristic of the muscle was introduced and utilized to develop
in the 120 s. The lack of an active cooling system delays the a variable stiffness joint robot. The variable stiffness joint robot
cooling phase and takes longer compared to its activating phase. was controlled using the developed RL method using the reward
Despite that, the pictures represent the same suppression per- function optimization concept. The experimental apparatus for
centage in both trials. assessing the controller was a rigid arm with a flexible joint based
The experiments are repeated for a vibration with a higher am- on the 3D-printed variable stiffness muscle.
plitude equal to 60°, shown in Figure 8c,d illustrate the compari- When the system’s features varied, the controller learnt to
son of the effect of the muscle in active and inactive mode. Com- adapt to new conditions online. The RL controller’s rise time
pared to the vibration with lower amplitude, the muscle performs changes from 0.26 to 0.27 s (after retraining), but the identical
better with more severe vibration, and the suppression in the idle parameter for RL raises from 0.26 to 0.33 s when the stiffness
mode is much smaller than in the active mode. It can reduce the of the joint changes. Additionally, compared to its linear version,
vibration by 14.75% in idle mode, while 81.88% of the vibration the RL’s oscillation and power consumption are reduced by 56%
is suppressed in activated mode. Figure S7a,b, Supporting Infor- and 58%, respectively. Also, the developed muscle was used in an
mation (Appendix 1) represent the performance of the muscle in experiment to suppress vibrations similar to tremors in the hu-
inactive mode before activating and after cooling down in 120 s man upper limb. By increasing the stiffness of the muscle when
with a minor difference. the vibration occurs, the vibration’s amplitude is decreased up to
Based on the result of suppression vibration, the muscle can 81.88%. Also, its repeatability and recovery in suppressing vibra-
be a great candidate for wearable devices to suppress tremor be- tion were assessed.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (9 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

Figure 8. Suppressing tremor with 15° and 60° amplitude by silicone muscle. a) Comparison of the 15° vibration in various situations in the time
domain: 1) Without the muscle. 2) Muscle is added in the idle mode. 3) Muscle is activated. b) The same comparison of the vibration suppression in
the frequency domain. c) Comparison of the 60° vibration in various situations in the time domain: 1) Without the muscle. 2) Muscle is added in the
idle mode. 3) Muscle is activated. d) The same comparison of the vibration suppression in the frequency domain.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (10 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

The combination of 4D printing of soft actuators with reliable by changes in the length of the muscle to evaluate the thermal insulation of
and adaptable controllers empowered by RL algorithms offers a the muscle versus the contraction percentage during its activating phase.
promising outcome with a low energy consumption rate for fu- Controllability of the Soft Joint Equipped with the PTM and Low Frequency
Vibration Suspension: This characteristic was measured using a setup
ture cutting-edge and creative applications.
where the main component was the Quanser flexible joint that included a
free rigid arm connected to two identical springs to provide flexibility. To
measure the controllability of PTM the springs were replaced with the soft
muscle and to assess low frequency vibration attenuation the PTM was
4. Experimental Section added to the spring in parallel mode. The module can be derived from
Matlab Simulink, but the signal must be amplified where the Quanser
Materials: 1) Ecoflex 00–30: Platinum-catalyzed silicone material with VoltPAQ-X1 Amplifier was exploited to amplify the driving signal. Also,
part A and part B. Platinum catalyst and silicones were included in com- for data acquisition from the flexible joint module the Quanser Q8-USB
ponent A, whereas component B held the cure inhibitor, cross-linker, and Data Acquisition devise was utilized. An RL-based controller was devel-
polymer. 2) THI-VEX: a silicone thickener that is added to Ecoflex silicones oped (in MATLAB/SIMULINK on a common computer with the Intel Core
in a weight-based manner. 3) Carbon fiber: the Hefei Minco Heating Ca- i5-10310U Processor and 16GB RAM) in this research to control the flex-
ble Co carbon fiber heating wire with 33 ohm m−1 resistance and 3 mm ible joint and while the muscle stiffness was changed, the controller was
thickness contains 12 000 threads that can endure up to 200 °C. 4) Expand- trained online to adopt to new condition. In the vibration mitigation test,
able sleeve: expandable sleeving made from PET (polyester) monofilament while the vibration occurred the stiffness of the muscle was increased to
yarns. 5) Ethanol: volatile and colorless liquid (CH3 –CH2 –OH). suppress the vibration.
Fabrication of Soft Actuator: The silicone was prepared by mixing
1A:1B by volume. It had a low viscosity to ensure easy mixing and de-
gassing. In the printing processes, higher viscosity was required that was
achieved by adding 2% thivex by volume. Therefore, 10 mL of part A and Supporting Information
part B were added to two syringes separately. 0.2 mL of thivex was added
Supporting Information is available from the Wiley Online Library or from
to each syringe to increase the viscosity. The syringes were loaded into the
the author.
syringe pumps to push through the static mixer. The dispensing speed of
the silicone was set to 10 mL s−1 . The needle inner diameter was 1.27 mm,
and the nozzle speed is 13 mm s−1 . Further aid in curing was provided by a
24 W ceramic heater after printing the silicone. The cured printed silicone Acknowledgements
was covered with expandable sleeve and the same procedure was contin-
ued until the sleeve was covered with a thin layer of the silicone. The curing The authors would like to thank the School of Engineering, the Faculty
of the silicone was assisted by using the heater again. The printed silicone of Science, Engineering and Built Environment, and Deakin University for
was unloaded from the mandrel and the carbon fiber with the length of financial support.
10 mm was inserted into the printed structure. Both sides of the structure Open access publishing facilitated by Deakin University, as part of the
were covered with silicone to enclose the chamber, then 2.5 mm ethanol Wiley - Deakin University agreement via the Council of Australian Univer-
was injected into the inner chamber. sity Librarians.
Contraction Characterization: The contraction of the artificial muscle
was evaluated through two experiments. In the first experiment, one end
of the muscle was fixed while the other end was relaxed, and a constant Conflict of Interest
voltage was applied to measure the maximum contraction. POWERTECH
0–30Vdc 5A regulated power supply was used to apply voltage on the mus- The authors declare no conflict of interest.
cle and a Keji Stainless Steel Ruler 30 cm was used to measure the size of
the muscle. Another experiment measured the muscle’s ability to contract
in response to an external force by attaching a 130 g weight to the relaxed Data Availability Statement
end. In both experiments, 5 and 8 V were applied to the muscle and these
experiments were repeated to check the muscles performances. The data that support the findings of this study are available from the cor-
Muscle Recovery: The muscle recovery was performed in the room responding author upon reasonable request.
temperature 21 ○ C. To conduct the experiment the PTM was activated,
then the power supply (POWERTECH 0–30Vdc 5A) was turned off. The
length of the PTM was measured by utilizing a ruler (Keji Stainless Steel
Ruler). The time after disconnecting the power supply until the muscle Keywords
reached its initial length (70 mm) is the recovery time. 4D printing, actuators, autonomy, controls, phase changes, silicone
Power Consumption: The power consumption of the PTM to activate
was an important feature that was measured based on the applied voltage
and the characteristic of the carbon fiber provided by The Hefei Minco Received: February 10, 2023
Heating Cable Co. the amount of consumed energy was calculated based Revised: April 12, 2023
2 Published online: May 4, 2023
on the formula VR t where V is the constant voltage applied on both ends of
the muscle and R is the electrical resistance of the muscle, and t represents
the time period from turning on the power source until the muscle reaches
its highest contraction. This experiment was repeated in various condition
including applying different voltage to discuss the energy and activating [1] S. K. Sood, K. S. Rawat, G. Sharma, IEEE Trans. Eng. Manag 2022,
time required for activating the muscle. https://doi.org/10.1109/TEM.2021.3134128.
Thermal Insulation of the Muscle: Another key aspect of the muscle fea- [2] A. Zolfagharian, A. Kaynak, M. Bodaghi, A. Z. Kouzani, S. Gharaie, S.
tures was its external temperature, which was relevant in wearable devices. Nahavandi, Appl. Sci. 2020, 10, 3020.
During contraction on recovery of the muscle the temperature of the mus- [3] J. Choi, O.-C. Kwon, W. Jo, H. J. Lee, M.-W. Moon, 3D Print. Addit.
cle’s exterior was recorded. For this purpose, the contact thermometer in- Manuf. 2015, 2, 159.
tegrated in HOTPLATE & MAGNETIC STIRRER PRO—ISG was used. The [4] X. Huang, M. Panahi-Sarmad, K. Dong, R. Li, T. Chen, X. Xiao, Com-
thermometer was attached to the PTM and the temperature was recorded posites, Part A 2021, 147, 106444.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (11 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH
 2365709x, 2023, 23, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/admt.202300199 by Indian Institute Of Technology Kanpur, Wiley Online Library on [16/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.advancedsciencenews.com www.advmattechnol.de

[5] A. Sharma, A. Rai, Mater. Today Proc. 2022, https://doi.org/10.1016/ [24] S. M. Mirvakili, D. Sim, I. W. Hunter, R. Langer, Sci. Rob. 2020, 5,
j.matpr.2022.03.679. eaaz4239.
[6] S. Mallakpour, F. Tabesh, C. M. Hussain, Adv. Colloid Interface Sci. [25] C. Tawk, G. Alici, in Smart Materials in Additive Manufacturing, Volume
2021, 294, 102482. 2: 4D Printing Mechanics, Modeling, and Advanced Engineering Appli-
[7] F. Momeni, X. Liu, J. Ni, Mater. Des. 2017, 122, 42. cations, Elsevier, New York 2022, pp. 103–140.
[8] S. Y. Hann, H. Cui, M. Nowicki, L. G. Zhang, Addit. Manuf. 2020, 36, [26] I. H. Akyuz, E. Yolacan, H. M. Ertunc, Z. Bingul, in 2011 IEEE Int. Conf.
101567. on Mechatronics, IEEE, Piscataway, NJ, 2011 pp. 409–414.
[9] A. Zolfagharian, A. Kaynak, A. Kouzani, Mater. Des. 2020, 188, 108411. [27] M. Ahmad, M. Suid, M. Ramli, M. Zawawi, R. R. Ismail, in 2010 6th
[10] M. N. I. Shiblee, K. Ahmed, M. Kawakami, H. Furukawa, Adv. Mater. Int. Colloquium on Signal Processing & its Applications, IEEE, Piscat-
Technol. 2019, 4, 1900071. away, NJ 2010, pp. 1–5.
[11] S. A. Ajwad, J. Iqbal, presented at Int. Conf. on Engineering and Emerg- [28] M. A. Ahmad, in 2008 Second UKSIM European Symp. on Computer
ing Technologies (ICEET), Lahore, March 2015. Modeling and Simulation, IEEE, Piscataway, NJ, 2008, pp. 40–45.
[12] S. Ajwad, M. Ullah, B. Khelifa, J. Iqbal, J. Balk. Tribol. Assoc. 2014, 20, [29] C. Chitu, J. Lackner, M. Horn, H. Waser, M. Kohlböck, in Proc. of the
499. Joint INDS’11 & ISTET’11, IEEE, Austria 2011, pp. 1–5.
[13] M. Runciman, J. Avery, M. Zhao, A. Darzi, G. P. Mylonas, Front. Rob. [30] J. Ju, Y. Zhao, C. Zhang, Y. Liu, Algorithms 2018, 11, 189.
AI 2020, 6, 141. [31] S. M. Ahmadi, M. M. Fateh, Trans. Inst. Meas. Control 2019, 41, 4023.
[14] M. Brancadoro, M. Manti, F. Grani, S. Tognarelli, A. Menciassi, M. [32] H. Ahmadian, H. A. Talebi, I. Sharifi, in 2020 28th Iranian Conference
Cianchetti, Front. Rob. AI 2019, 6, 12. on Electrical Engineering (ICEE), IEEE, Piscataway, NJ 2020, pp. 1–6.
[15] J. Zhou, Y. Chen, Y. Hu, Z. Wang, Y. Li, G. Gu, Y. Liu, Soft Rob. 2020, [33] P. Sarkhel, N. Banerjee, N. B. Hui, SN Appl. Sci. 2020, 2, 1124.
7, 743. [34] W. He, Z. Yan, Y. Sun, Y. Ou, C. Sun, IEEE Trans. Neural Networks
[16] A. Deaconescu, T. Deaconescu, Appl. Sci. 2022, 12, 6459. Learn. Syst. 2018, 29, 5993.
[17] A. Zolfagharian, M. P. Mahmud, S. Gharaie, M. Bodaghi, A. Z. [35] B. Subudhi, A. S. Morris, Appl. Soft Comput. 2009, 9, 149.
Kouzani, A. Kaynak, Virtual Phys. Prototyping 2020, 15, 373. [36] S. K. Pradhan, B. Subudhi, IEEE Trans. Autom. Sci. Eng. 2012, 9, 237.
[18] W. Liao, Z. Yang, Mater. Horiz. 2023, 10, 576. [37] T. Long, E. Li, Y. Hu, L. Yang, J. Fan, Z. Liang, R. Guo, IEEE Trans.
[19] C.-H. Li, H.-Q. Wang, Z.-Y. Huang, D.-W. Yue, F.-Z. Wang, Mater. Neural Networks Learn. Syst. 2020, 32, 841.
Horiz. 2023. [38] D. Sendrescu, G. Bujgoi, D. Chintescu, in 2020 21th Int. Carpathian
[20] M.-j. Kim, B.-g. Kim, J.-s. Koh, H. Yi, Autom Constr 2023, 145, 104660. Control Conf. (ICCC), IEEE, Piscataway, NJ 2020, pp. 1–5.
[21] C. De Pascali, G. A. Naselli, S. Palagi, R. B. Scharff, B. Mazzolai, Sci. [39] D. Pavlichenko, S. Behnke, arXiv:2203.07051 2022.
Rob. 2022, 7, eabn4155. [40] M. Mohammadi, A. Zolfagharian, M. Bodaghi, Y. Xiang, A. Z.
[22] D. Sangian, S. Naficy, G. M. Spinks, J. Intell. Mater. Syst. Struct. 2016, Kouzani, Bio-Des. Manuf. 2022, 5, 786.
27, 2508. [41] Y. S. Narang, B. Aktaş, S. Ornellas, J. J. Vlassak, R. D. Howe, Soft Rob.
[23] E. Sun, T. Wang, Smart Mater. Struct. 2019, 28, 127001. 2020, 7, 724.

Adv. Mater. Technol. 2023, 8, 2300199 2300199 (12 of 12) © 2023 The Authors. Advanced Materials Technologies published by Wiley-VCH GmbH