Page 101 - Read Online

Page 22 of 35 Kulkarni et al. Soft Sci. 2025, 5, 12 https://dx.doi.org/10.20517/ss.2023.51

Table 3. Summary of soft robot control systems and their application examples
Control Description Examples of soft robot applications
systems
OL Feedback data is not required to make changes to the system. Depends Soft hydraulic actuator for highly repeatable results-
only on input to achieve the desired output. Prior knowledge of the advantageous for surgical applications [242] . Pneumatically
[257]
operating environment is required for more accurate control actuated soft robotic manipulators can inflate devices to
different shapes [257]
CL Feedback data from output is required to adjust the input signal to Drug delivery devices receive real-time feedback data from
achieve correct output. Feedback data can be obtained by sensors to foreign body responses to change the actuation regimen to
measure the current conditions of the environment and system. Prior deliver the correct amount of drug to the body [244] . SMA
knowledge of the operating environment is not required [257] spring bot can actuate/deform with sensors to obtain
feedback data [258]
MF Feedback data from sensors is used to adjust input signal to achieve Locomotion bot using MF to interact with the environment
correct output without the use of models of device/system [248] . Can be by learning the state transitions and optimized periodic
[260]
used for devices that have complex structures/actuation mechanisms control sequences . Pneumatic muscle actuator soft
and environments that cannot be described by mathematical robots can follow certain paths [261]
[259]
models
MB Modeling strategies including PCC [245] and FEM [247] to mathematically Environmentally adaptive bot achieves planar motion using
represent the device and its operation to modify the input signal to MB CL controllers for trajectory tracking and impedance
[262] [262]
achieve the correct output control . Uses PCC and FEM to model and simulate the
device. Piezoelectric inchworm for constrained
[263]
environments with MB motion control
PID Type of MF control system that uses error between actual and desired Bio-inspired underwater vehicles use four PID controllers
[264] [267]
output to modify input signal to achieve correct output . Uses for autonomous control . The pneumatic arm can reach
proportional, integral, and derivative factors that can be adjusted for certain destinations or follow trajectories [268]
[265] [266]
more accurate control . Variations include PI control system
AC Can be MB or MF, depending on the application [261] . Modified input A soft robotic arm that uses AC to control its position and
[253] [271]
depending on known parameter disturbances. MRAC uses the stiffness . Underwater soft robot that uses AC for visual
device model to compare the system’s actual and predicted output to servoing and prior knowledge of environmental impact for
[253] [269] [272]
adjust the control signal . Other types include direct AC and more accurate control
[270]
indirect AC
RL Can be MB or MF, depending on the application. Evaluated feedback of Motion control of the underwater robot by training a neural
output signal used to monitor system behavior and modify input signal network developed using a deep RL algorithm for a bot to
to maximize system reward and achieve correct output [273] . RL learn how to travel through unpredictable environments
approaches include value-based [274] , policy-based, or MB and can also [256] . A caterpillar bot uses RL to travel to different
be either active or passive-based [274] environments [275]
OL: Open-loop; CL: closed-loop; SMA: shape memory alloy; MF: model-free; MB: model-based; PCC: piecewise curvature models; FEM: finite
element method; PID: proportional-integral-derivative; PI: proportional-integral; AC: adaptive control; MRAC: model reference adaptive control;
RL: reinforcement learning.
Future considerations of the control of soft robotic devices include artificial intelligence (AI). The AI-driven
control systems and models may consider the nonlinear structure of soft robots and implement control that
is similar to biological organisms . An AI-driven control system may consist of a neural network with a
[276]
hierarchical structure of input variables and hidden nodes that are connected and weighted based on
importance . There are also designs to develop neural networks for robotics such as supervised
[277]
[278]
[279]
[281]
[280]
control , direct inverse control , neural AC , back-propagation utility , and adaptive critic
methods . These strategies allow the system to learn based on the environment and use case . Machine
[283]
[282]
learning algorithms can be used for the control of soft robots as the nonlinearity of soft robotic devices can
be modeled and integrated into a neural network . Machine learning models can also be developed to
[276]
[276]
model the interaction between different actuators with complex nonlinear structures . Therefore, the
development of AI-driven control systems and models may allow for more robust control of soft robotic
devices for extreme environments as they learn and adapt to uncertain conditions.
Future considerations for the design and use of soft robots
The application of traditional, rigid robots currently dominates in terms of availability. As soft robots
continue to develop, opportunities for commercial products are increasing. Currently, soft robots are
available for applications including grippers in manufacturing , sleep monitoring technology ,
[285]
[284]