Machine Learning for Soft Robot Sensing and Control: A Tutorial Study

Abstract

Developing feedback controllers for robots with embedded sensors is challenging and typically requires expert knowledge. As machine learning (ML) advances, the development of learning-based controllers has become more and more accessible, even to non-experts. This work presents the development of a tutorial to educate non-roboticists about MLbased sensing and control in cyber-physical systems using a soft robotic device. We demonstrated this by creating a recurrent neural network-based closed-loop force controller for a soft finger with embedded soft sensors. Our hypothesis is validated in a 2.5hour workshop session for students with no prior knowledge of robot control. This work serves as a tutorial for participants aiming to experience and perform a general benchmark for soft robot control tasks, with little or even no expertise in robotics.

Soft touchless sensors and touchless sensing for soft robots

“…Touchless technologies make use of different materials for sensing. These materials include soft alloys ( Zhukov, 2017 ), air ( Wang et al, 2022a ), light ( Wu et al, 2021 ), capacitance ( Ham et al, 2022 ), magnetism ( Dong et al, 2022 ), IR ( Muthuviswadharani et al, 2016 ), acoustic ( Kang et al, 2018 ) and virtual technologies ( Mattera et al, 2018 ). Improved fabrication techniques, resolution, and range in recent years largely expanded the application of sensors made of soft materials or noncontact media.…”

Soft robotic tactile perception of softer objects based on learning of spatiotemporal pressure patterns

“…The characteristic of the silicone layer resembles the natural dynamics of the human flesh, which in turn, deforms and dynamically changes the pressure distribution over the contact area depending on the property of the to-be-contacted object and the kind of contact with the object in the environment [23]. To measure the contact forces of the contact, six air chambers were cast at the tip of the anthropomorphic finger and connected to NXP MPXH6300AC6U pressure sensors via elastic hoses [24]. The air chamber would affect the stiffness of the skin where the finger is pressed, but would not affect the overall stiffness as the deformation is localized around each joint.…”

Reduced order modeling and model order reduction for continuum manipulators: an overview

“…However, it is not easy to capture commonly observed phenomena in soft robot real behavior such as fabrication inconsistencies, visco-hyper-elastic material properties, material creep, and hysteresis. Learnt solutions (when trained) are on the other hand real-time, accurate (within the training dataset domain), capable of handling complex structural geometries and nonlinearities, and easy to implement by not requiring significant knowledge of continuum robot theory ( Thuruthel et al, 2017 ; Thuruthel et al, 2018a ; Jolaei et al, 2020 ; Truby et al, 2020 ; Kim et al, 2021 ; Wang et al, 2021 ; Wang et al, 2022 ). However, the validity of a deep learning-based solution without a conservational law is limited to the richness and generality of the dataset used for learning, hence lacking robustness in presence of new conditions, contacts, and external loads which are not tested before (i.e., available in the dataset).…”