Closing the Control Loop with Time-Variant Embedded Soft Sensors and Recurrent Neural Networks

Sensing, Actuating, and Interacting Through Passive Body Dynamics: A Framework for Soft Robotic Hand Design

Robotic hands have long strived to reach the performance of human hands. The physical complexity and extraordinary capabilities of the human hand, in terms of sensing, actuation, and cognitive abilities, make achieving this goal challenging. At the heart of the physical structure of the hand is its' passive behaviors. Seen most clearly in soft robotic hands, these behaviors influence and affect the mechanical, sensing, and control functionalities. With this perspective, we present a framework through which passivity in robot hands can be understood, by concretely identifying the role of passivity in the design, fabrication, and control of soft hands. In this framework we focus on the interactions between the physical hand and the: environment, internal actuation, sensor morphology, and wrist control. Taking these surrounding systems away, we are left with a passive soft hand whose behaviors emerge from external interactions. Inspired by the human hand, we define the role of these four key interacting pillars and review how state-of-the art robot hands utilize these four elements to aid functionality. We show how these pillars promote hybrid soft-rigid hands with rich behaviors, providing benefits in terms of the increased adaptability to uncertain environments, improved scalability and reduction in the cost of actuation, sensing, and control. This review provides a conceptual framework for approaching hand design and analysis through consideration of the passive behaviors. This highlights not only the advances that can be made by approaching the problem in this way but also the outstanding challenges that stem from this outlook.

Data-driven method for damage localization on soft robotic grippers based on motion dynamics

Damage detection is one of the critical challenges in operating soft robots in an industrial setting. In repetitive tasks, even a small cut or fatigue can propagate to large damage ceasing the complete operation process. Although research has shown that damage detection can be performed through an embedded sensor network, this approach leads to complicated sensorized systems with additional wiring and equipment, made using complex fabrication processes and often compromising the flexibility of the soft robotic body. Alternatively, in this paper, we proposed a non-invasive approach for damage detection and localization on soft grippers. The essential idea is to track changes in non-linear dynamics of a gripper due to possible damage, where minor changes in material and morphology lead to large differences in the force and torque feedback over time. To test this concept, we developed a classification model based on a bidirectional long short-time memory (biLSTM) network that discovers patterns of dynamics changes in force and torque signals measured at the mounting point. To evaluate this model, we employed a two-fingered Fin Ray gripper and collected data for 43 damage configurations. The experimental results show nearly perfect damage detection accuracy and 97% of its localization. We have also tested the effect of the gripper orientation and the length of time-series data. By shaking the gripper with an optimal roll angle, the localization accuracy can exceed 95% and increase further with additional gripper orientations. The results also show that two periods of the gripper oscillation, i.e., roughly 50 data points, are enough to achieve a reasonable level of damage localization.