Data-driven Control Approaches for Modular Soft Robots

Compared to rigid robots, soft robots possess higher degrees of freedom and hence are able to achieve some tasks challenging to their rigid counterparts. Thanks to their flexibility, soft robots can be applied in constrained environment exploration. In addition, the low stiffness endows soft robotics with safety during interaction with external objects, environments, and humans. Due to these advantages, soft robots have been widely leveraged in medical and rehabilitation applications, environment exploration, and industrial grippers. In various soft robots, modular soft robots have demonstrated their unique benefits by performing sophisticated deformations. The multiple modules can deform independently and hence exhibit higher degrees of freedom than single-module robots. Therefore, this kind of soft robot has promising potential in many areas, such as serving as a surgical tool in the medical area. However, this modular structure also induces challenges in sensing and control. The soft material or flexible structure of soft robots leads to nonlinearity, hysteresis, and time delay in soft robot motion. In addition to these difficulties, modular soft robots further introduce the modular structure, accumulating errors along the module sequence. Also, modular soft robot sensing also meets challenges. To achieve high accuracy control, feedback control is necessary. External sensing, such as optical or electromagnetic tracking, constrains the modular soft robots inside the limited sensing area, while internal sensing, such as IMU or encoder, always suffers from sensing errors caused by drift or cable slack. Based on this background, it is necessary to propose control strategies specifically for modular soft robots to achieve the practical application of modular soft robots. To deal with the challenges mentioned above, we propose a series of strategies, mainly leveraging neural networks. The nonlinear activation functions can be applied to deal with the nonlinearity of soft robot motion. The network structure is good at estimating mapping between robot state space, configuration space, and actuation space without the knowledge of the physical model. In different categories of neural networks, recurrent neural networks can deal with sequential problems like time sequence and module sequence in space. Therefore, neural networks are effective tools for modular soft robot control. This thesis introduces some data-driven control strategies for single-module and modular soft robots. In Chapter 1, we briefly introduce the motivation of this study related to soft robot control. In Chapter 2, we review the approaches applied to soft robot modeling and control, including the Jacoian approach, physical approach, statistical approach, neural network, and reinforcement learning. Chapter 3 focuses on single-module soft robot controllers for interchangeability and robot adaptability. In Chapter 4, we introduce the configuration controller and planning strategy for modular soft robots. Finally, in Chapter 5, we summarize our work and propose some possible directions for future work.