This study focuses on the locomotion capability improvement in a tendon-driven soft quadruped robot through an online adaptive learning approach. Leveraging the inverse kinematics model of the soft quadruped robot, we employ a central pattern generator to design a parametric gait pattern, and use Bayesian optimization (BO) to find the optimal parameters. Further, to address the challenges of modeling discrepancies, we implement a multi-fidelity BO approach, combining data from both simulation and physical experiments throughout training and optimization. This strategy enables the adaptive refinement of the gait pattern and ensures a smooth transition from simulation to real-world deployment for the controller. Compared to previous result using a fixed gait pattern, the multi-fidelity BO approach improves the robot’s average walking speed from 0.14 m/s to 0.214 m/s, an increase of 52.7%. Moreover, we integrate a computational task off-loading architecture by edge computing, which reduces the onboard computational and memory overhead, to improve real-time control performance and facilitate an effective online learning process. The proposed approach successfully achieves optimal walking gait design for physical deployment with high efficiency, effectively addressing challenges related to the reality gap in soft robotics.

4D printing (4DP) enables researchers and developers to fabricate soft robots in a more repeatable and integrated manner. The employment of smart materials also reinforces the performance and flexibility of the robot. However, compared with the traditional rigid robots, the 4D-printed soft robots normally have high nonlinearity and time-variant dynamics, which makes the modeling of these robots challenging. Furthermore, as an inherent disadvantage of all soft robots, they also suffer from low actuation precision and poor controllability. This chapter reviews the efforts in modeling and control of 4D-printed soft robots and then summarizes the available approaches. Then, this chapter will compare these approaches based on a 4D-printed actuator using temperature-stimulated shape memory polymers. It is demonstrated that the learning-based methods have great potential in performing better modeling and control of the 4D-printed robots compared with traditional alternatives.

Soft actuators have the advantages of compliance and adaptability when working with vulnerable objects, but the deformation shape of the soft actuators is difficult to measure or estimate. Soft sensors made of highly flexible and responsive materials are promising new approaches to the shape estimation of soft actuators, but suffer from highly nonlinear, hysteresis, and time-variant properties. A nonlinear and adaptive state observer is essential for the shape estimation from soft sensors. Current state estimation methods rely on complex nonlinear data-fitting models, and the robustness of the estimation methods is questionable. This study investigates the soft actuator dynamics and the soft sensor model as a stochastic process characterized by the Gaussian Process (GP) model. The unscented Kalman filter (UKF) is applied to the GP model for more reliable variance adjustment during the sequential state estimation process than conventional methods. In addition, a major limitation of the GP model is its computational complexity during online inference. To improve the real-time performance while guaranteeing accuracy, we introduce an edge server to decrease the onboard computational and memory overhead. The experiments showcase a significant improvement in estimation accuracy and real-time performance compared to baseline methods.

Digital twin (DT) models are computational models that can effectively represent different assets and processes in the manufacturing environment. Moreover, the DT models can support intelligent automation by integrating with the digital foundation and the data analytics provided by the cyber-physical system (CPS) in an industrial environment. To properly model a physical process, a DT model should be updated online to closely and timely model the underlying process and reduce modeling uncertainty in the CPS. However, most DT models are created offline and implemented online, which cannot be easily updated by using online data from heterogeneous product designs or manufacturing processes. This limitation arises from existing online learning methods, which are typically designed for identical structures, while real manufacturing CPS involves personalized designs and diverse processes. More importantly, there are limited samples for the same product design or manufacturing process due to manufacturing personalization, which slows down the online updating of DT models. In this article, the authors investigated online DT model updating based on data collected from different product designs and/or processes. The authors proposed a mutual active learning framework to identify informative samples from different designs or processes for online DT model updating. Specifically, by properly balancing the gradient-based features of the DT models and the similarity among these heterogeneous designs or processes, the proposed method can effectively query the most informative samples among heterogeneous processes to update the corresponding DT model in a timely manner. The advantages of the proposed method are illustrated by an engineering-driven statistical DT model for an additive manufacturing process (i.e., fused deposition modeling).

Soft actuators made from compliant materials are superior to conventional rigid robots in terms of flexibility, adaptability and safety. However, an inherent drawback of soft actuator is the low actuation precision. Implementing closed loop control is a possible solution, but the soft actuator shape can hardly be measured directly by commercially available sensors, which either are too stiff for integration or cause performance degradation of the actuator. Although 3D printing has been applied to print bendable sensors from conductive materials, they either have larger stiffness than the soft actuator or are made from specially designed materials that are difficult to reproduce. In this study, easily accessible commercial soft conductive material is applied to directly 3D print soft sensors on soft actuators. Different configurations of the printed sensors are studied to investigate how the sensor design affects the performance. The best sensor configuration is selected to provide shape feedback using its changing resistance during deformation. Compared with a commercial flexible bending sensor, the printed sensor has less influences on the soft actuator performance and enjoys higher shape estimation accuracy. Closed loop shape control of the actuator using feedback from the 3D printed sensor is then designed, implemented and compared with the control results using image feedback. A gripper consisting of three individually controlled soft actuators demonstrates the applications of the soft sensor.

The COVID-19 pandemic makes protective visors important for protecting people in close contacts. However, the production of visors cannot be increased greatly in a short time, especially at the beginning of the pandemic. The 3D printing community contributed largely in fabricating the visor frames using the rapid and adaptive manufacturing ability. While there are many open source designs of face visors for affordable 3D printers, all these designs fabricate mono-sized frames without considering diverse users’ dimensions. Here, a new method of visor post-processing technology enabled by closed loop controlled 4D printing is proposed. The new process can further deform the printed visor to any customized size for a more comfortable user experience. FEM analysis of the customized visor also shows consistent wearing experience in different circumstances compared with the old visor design. The fabrication precision and time cost of the method is studied experimentally. A case study regarding the reducing, reusing and recycling (3R) of customized visors in classrooms is proposed to enable the customized visors manufactured in a more sustainable way.

Sensors play a vital role in the industry transformation. Commercialsensors such as force sensors have limited options in shapes, stiffness, measuringranges, etc. Customized force sensors optimized for the production environmentcan greatly increase the integration workflow and avoid the trade-off in design freedomof using commercial sensors. 3D printing, as a rapid prototyping technology,offers great potential in fabricating force sensors customized to a specific application.However, most of the existing 3D printed force sensors are limited to onedirectionalsensing, while most of them use materials developed in-house. In thisstudy, a fully 3D printed force sensor using commercial conductive 3D printing materialsis presented. By utilizing the resistance change when under load, the sensorcan estimate the applied force in multiple directions. The resistive performance ofthe prototype 3D printed force sensor is first characterized and then validated in acase study.

Exploiting novel sensors and actuators made of flexible and smart materials becomes a new trend in robotics research. The studies on the design, production, and control of the new type of robots motivate the research fields of soft robots and 4D printed robots. 3D Printing (3DP) is an additive manufacturing technology that is widely used in printing flexible materials to fabricate soft robots. 4D Printing (4DP) combines 3DP technologies with smart materials to produce transformable devices. 4DP first prints structures with specifically designed responsive materials. When external stimuli such as temperature, voltage, or magnetic field are applied to the printed structure, it changes shape in a programmable way. The shape morphing property of 4DP makes it a novel approach to the actuators of robots.

The employment of these special materials empowers these new robots with better compliance and adaptability to the working environment. However, compared with the rigid counterparts, they also have complex dynamic properties such as substantial non-linearity and time-variance. These factors make the precise modeling and robust control of these new robots challenging and thus hinder their potential applications. Focusing on soft robotic systems enabled by 3DP and 4DP approaches, this dissertation studies both traditional and Machine Learning (ML)-based approaches to the modeling, perception, and control of soft, non-linear, and time-variant robotic systems. The main contributions of this dissertation are:

• The scheme of Closed-Loop (CL) controlled 4DP (CL4DP) using temperature stimulated Shape Memory Polymer (SMP) is designed and validated numerically and experimentally. The feedback control system increases the precision and robustness of the shape morphing process of 4D printed SMP. Applications of CL4DP are explored.
• Data-driven model identification methods are applied to learn the dynamic model of the shape morphing process of CL4DP and the learned model has good quality to support model-based control design. Model-free and adaptive Reinforcement Learning (RL) controllers are developed to deal with the non-linearity and time variance of 4D printed actuators. To improve the stability and quick adaptability, a concise basis function set is selected instead of blindly using Deep Neural Networks (DNNs).
• A quadruped robot enabled by soft actuators and its simulation model are developed. The computation efficiency and model accuracy of the simulator are studied and optimized by comparing different simulation methods such as Finite Element Method (FEM) and lumped parameter method.
• The optimal walking gait pattern of a soft-legged quadruped robot is found by grid parameter search and RL with a physics based simulation model. To speed up the RL training process, modeling tricks are used to reduce the simulation time of the model and curriculum learning is used to reduce the learning time.
• A soft sensor made by printable conductive materials and 3DP is designed and optimally calibrated to estimate the shape of a pneumatically driven soft actuator. The geometry of the soft sensor is optimally designed for the best linearity, hysteresis and drift properties. The online estimation is based on a linear regression model learned from experimental data.
• A pneumatically driven soft gripper is developed by 3DP, the printable soft sensor, and pole-placement control methods. The operation of the gripper does not require an external image feedback system to measure its shape, which is estimated by the integrated soft sensor. The position feedback by the soft sensor and the controller by the pole-placement method enable the soft gripper to perform complex tasks with high precision.

Användande av nya sensorer och aktuatorer av flexibla och smarta material har blivit en ny trend inom robotikforskning. Studier om design, produktion och styrning av den nya typen av robotar motiverar förskningen om mjuka robotar och 4D-printade robotar. 3D Printing (3DP) är en additiv tillverkningsteknik som används i stor utsträckning vid utskrift av flexibla material för att tillverka mjuka robotar. 4D Printing (4DP) kombinerar 3DP-teknik med smarta material för att producera transformerbara enheter. 4DP skriver först ut strukturer med specifikt designade responsiva material. När yttre stimuli som temperatur, spänning eller magnetfält appliceras på den utskrivna strukturen ändrar den form på ett programmerbart sätt. Omformningsegenskapen hos 4DP skapar ett nytt sätt att aktuera robotar.

Användningen av dessa specialmaterial ger dessa nya robotar bättre följsamhet och anpassningsförmåga till sin arbetsmiljö. Men jämfört med de stela motsvarigheterna har de också komplexa dynamiska egenskaper såsom betydande icke-linjäritet och tidsvarians. Dessa faktorer gör den nogranna modelleringen och robusta kontrollen av dessa nya robotar utmanande och hindrar därmed deras potentiella tillämpningar. Med fokus på mjuka robotsystem som möjliggörs av 3DP- och 4DP-metoder, studerar denna avhandling både traditionella och Machine Learning (ML)-baserade metoder för modellering, perception och kontroll av mjuka, icke-linjära och tidsvarierande robotsystem. De viktigaste bidragen från denna avhandling är:

• En metod för Closed-Loop (CL) reglerad 4DP (CL4DP) med temperaturstimulerad Shape Memory Polymer (SMP) har utvecklats och validerats både numeriskt och experimentellt. Reglersystemet ökar precisionen och robustheten i omformningsegenskapen för 4D-utskrivet SMP. Tillämpningar av CL4DP utforskas.
• Metoder för datadriven modellidentifiering tillämpas för att lära sig den dynamiska modellen av omformningsprocessen för CL4DP och den inlärda modellen är lämplig modellbaserad reglering. Modellfria och adaptiva Reinforcement Learning-regulatorer (RL) har utvecklats för att hantera icke-linjäriteten och tidsvariationen hos 4D-utskrivna aktuatorer. För att förbättra stabiliteten och snabb anpassningsförmåga väljs en mindre uppsättning basfunktioner istället för att blint använda Deep Neural Networks (DNN).
• En fyrbent robot med mjuka aktuatorer och dess simuleringsmodell utvecklas. Beräknings effektiviteten och modell noggrannheten hos simulatorn studeras och optimeras genom att jämföra olika simuleringsmetoder såsom Finite Element Method (FEM) och lumped parameter method.
• Det optimala gångmönstret för en mjukbent fyrbensrobot hittas genom rutnäts-sökning och RL med en fysikbaserad simuleringsmodell. För att effektivisera RL-tränings-processen används modelleringsknep för att minska simuleringstiden för modellen och curriculum learning används för att minska inlärningstiden.
• En mjuk sensor gjord av utskrivbara elektriskt ledande material och 3DP är designad och optimalt kalibrerad för att uppskatta formen på ett pneumatiskt driven mjuk aktuator. Den mjuka sensorns geometri är optimalt utformad för bästa linjäritet, hysteres och driftegenskaper. Online-uppskattningen är baserad på en linjär regressionsmodell som lärts från experimentella data.
• En pneumatiskt driven mjukgripare är utvecklad av 3DP, den utskrivbara mjuka sensorn och polplacerings reglering. Griparens funktion kräver inte ett externt bildåterkopplingssystem för att mäta dess form, vilket istället uppskattas av den integrerade mjuka sensorn. Positionsåterkopplingen från den mjuka sensorn och regulatorn genom polplaceringmetoden gör att den mjuka griparen kan utföra komplexa uppgifter med hög precision.

Using soft actuators as legs, soft quadruped robots have shown great potential in traversing unstructured and complex terrains and environments. However, unlike rigid robots whose gaits can be generated using foot pattern design and kinematic model of the rigid legs, the gait generation of soft quadruped robots remains challenging due to the high DoFs of the soft actuators and the uncertain deformations during their contact with the ground. This study is based on a quadruped robot using four Compressible Tendon-driven Soft Actuators (CTSAs) as the legs, with the actuator's compression motion being utilized to improve the walking performance of the robot. For the gait design, an inverse kinematics model considering the compression of the CTSA is developed and validated in simulation. Based on this model, walking gaits realizing different motion speeds and directions are generated. Closed loop direction and speed controllers are developed for increasing the robustness and precision of the robot walking. Simulation and experimental results show that omnidirectional locomotion and complex walking tasks can be realized by tuning the gait parameters and the motions are resistant to external disturbances.

Combining 3D printing and smart materials, 4D printing technologies enable the printed actuators to furtherchange their shapes or other properties after prototyping. However, the shape morphing of 4D printed actuatorssuffers from poor controllability and low precision. One of the main challenges is that the 4D printed actuatorsare hard to be modeled and it is difficult to develop an appropriate controller for them. In this study, variouspopular reinforcement learning (RL) methods are applied to address the problem of online and adaptive model-free control of 4D printed shape memory polymer (SMP). Their training efficiencies are compared and anadaptive LQR controller based on Q learning is developed to realize efficient online learning. The RL controllerachieves precise and quick shape control within 2 − −3 learning episodes and is adaptive to the changingproperties of SMP. The RL controller performance is then compared with a model-based LQR controller andshows high control precision and excellent adaptability to the varying control plant.