Harnessing Computing Power in Origami-Inspired Systems: Design, Mechanism, and Embodied Perception through Physical Reservoir Computing

Abstract

This dissertation investigates how mechanical and soft robotic bodies can be used not only for motion and interaction, but also as physical substrates for computation and perception. The work is built on the framework of physical reservoir computing(PRC), in which a nonlinear dynamical system serves as a fixed computational kernel and only a simple readout layer is trained. Within this framework, the intrinsic dynamics of a mechanical body can project low-dimensional inputs into rich, high-dimensional transient responses that can be exploited for machine learning and sensing tasks with minimal digital computation. The central goal of this dissertation is to move PRC in mechanical and soft robotic systems beyond proof-of-concept demonstrations toward a task-aware design framework. To address this goal, the dissertation studies PRC across multiple platforms with increasing functional relevance. First, a folded Miura-ori origami structure is shown to support information-processing tasks including payload estimation, payload-position classification, input-pattern recognition, and multitasking, thereby establishing that mechanically rich bodies can serve as useful physical reservoirs for perception-oriented tasks. Second, a fabric pneumatic soft robotic arm is used to compare proprioceptive and exteroceptive perception, revealing that different task classes place different demands on reservoir dynamics, sensing richness, and training protocol. Third, an adaptive modular origami manipulator is repurposed as a reconfigurable physical reservoir to investigate how morphology, stiffness configuration, and excitation condition affect computing performance. These experiments show that there is no universally optimal reservoir design and that interpretable metrics such as spectral alignment and spatial correlation can link physical regime to task success. Fourth, a reduced-order dynamic model of the modular arm is developed to provide simulation-based mechanistic insight into how stiffness configuration, payload, actuation amplitude, and observation frequency shape the operating regime and nonlinear emulation capability of the reservoir. Finally, these principles are embodied in a soft modular arm with integrated bending strain gauges, demonstrating multi-modal sensing of posture, payload weight, and payload orientation through simple linear readout from body dynamics. Taken together, the results show that the computational capability of a physical reservoir is not a fixed property of morphology alone. Instead, useful embodied computation emerges from the interaction among body design, excitation, sensing modality, state construction, and task requirement. This dissertation therefore provides both experimental validation and design-oriented understanding for harnessing computing power in origami-inspired and soft robotic systems, and it establishes a pathway toward embodied intelligent machines with reduced sensing and computation complexity.