Design, Control, and Motion Planning of Cable-Driven Flexible Tensegrity Robots

Tensegrity structures are an emergent type of soft-robotics that are compliant, lightweight, and impact-resilient. In collaboration with NASA Ames Research Center, research in the Berkeley Emergent Space Tensegrities Lab at UC Berkeley has largely focused on the design and control of these novel structures as potential surface exploration robots which could act as both landers and rovers. More recently, tensegrity robots have also been proposed for applications closer to home – working as disaster response and emergency co-robots to help first responders obtain situational awareness faster and safer. Constructed using isolated rigid bodies suspended in a tension network of elastic elements, tensegrity structures exhibit unique and advantageous mechanical properties for applications in uncertain and potentially hazardous environments, albeit at the cost of increased complexity for dynamic feedback control. In addressing these challenges, this work explores possible approaches for feedback control and state estimation for ground-based rolling locomotion with six-bar spherical tensegrities. In this dissertation, we explore problems pertaining to practical implementation – state estimation, modeling, motion planning, and optimal control of tensegrity robots under uncertainty. Leveraging the well-structured dynamics of Class-1 tensegrity robots, we implement and evaluate model-based Model Predictive Control and iterative local quadratic methods for tensegrity motion planning. Additionally, we consider alternative tensegrity topologies and actuator schema which may enable improved performance for task-specific objectives. Due to the many degrees of freedom and compliant nature of tensegrity structures, however, excessive state estimate errors may propagate catastrophically. To evaluate these effects, Bayesian state estimators are applied to tensegrity ground mobility in simulation, evaluating their performance under the additional constraints of low-cost sensors and potentially scarce and noisy sensor data. An imitation learning approach is introduced to achieve directed rolling motion using a contextual neural network policy, combining deep learning and optimal control for real-time feedback control of highly nonlinear tensegrity systems. Finally, a robust minimax control approach is proposed in order to address challenges which arise at the intersection and interaction of state estimation and trajectory optimization for flexible tensegrity robotics. Combined, these pragmatic research developments help advance the progression of this novel technology towards becoming a viable and more widely adopted robotics paradigm