Articulated Soft Robots

Soft robots can revolutionize several applications in confined spaces with high safety demands. Articulated soft robots (ASRs) feature a vertebrate-like structure with intrinsically compliant joints, making them a suitable design for snake-like manipulators with multiple degrees of freedom. Various challenges hinder the use of ASRs and are addressed in this thesis: (1) ASR designs should be modular, simplifying the construction of hyper-redundant robots with high dexterity. One major obstacle is the lack of open-source designs, which considerably limits accessibility, further development, comparability, and reproducibility in the research field. This thesis presents open-source designs for modular ASRs, serving as an experimental platform for robot-based evaluation or as a starting point for constructing hyper-redundant robots. (2) Operating such snake-like systems with many degrees of freedom should be intuitive. However, existing approaches primarily focus on remotely controlling the robot's locomotion into target areas. This thesis presents a unified strategy for intuitive telemanipulation of hyper-redundant ASRs, providing user-friendly reorientation possibilities in addition to locomotion. Such reorientation is crucial after reaching the target area, since both the view of a camera at the robot tip and the working area may be limited. (3) Modeling and controlling soft robots is challenging due to several nonlinearities. Learning-based strategies are promising and can help avoid potentially elaborate physics-based system modeling. However, black-box approaches are data-hungry and usually exhibit poor generalization. If model knowledge is available, physics-informed machine learning offers potential in providing fast surrogate models with high generalization. This work explores both black-box and hybrid learning for accurate model-based control, building on established approaches to feedforward control and model predictive control. Novel controllers for soft robots based on Gaussian processes, recurrent and physics-informed neural networks are presented. This dissertation adopts a holistic approach to researching methods and strategies for achieving progress towards employing ASRs. From modular design to intuitive telemanipulation and learning for control: Real-world applicability is demonstrated through several experiments with physical prototypes and a telemanipulation study. Furthermore, hard- and/or software were released as open source in all three research areas. This enables a low entry barrier for the entire research community to pursue further development, with one long-term goal: solving real-world problems using soft snake-like robots.