Physical Reservoir Computing With Compliant Fiber Networks

Abstract Inspired by the webs of orb-weaving spiders, we explore the potential of compliant fiber networks to serve as mechanically intelligent physical reservoir computers. By exploiting the nonlinear dynamics inherent in the mechanical deformations of networks of connected fibers, we perform complex computational tasks without the need for conventional neural network architectures. Using Cosserat rod-based models, we simulate the behavior of these fiber networks, evaluating their computational abilities through the metrics of nonlinear capacity and memory capacity. We find that increasing the number of fibers enhances the nonlinear computational and memory recall capacity of the network, improving its computational performance. To validate these findings, we construct a physical prototype of the fiber network and experimentally demonstrate its ability to nonlinearly process input signals and recall previous inputs. Results suggest that compliant fiber networks can serve as effective physical reservoirs for mechano-intelligent computation, offering a low-cost, scalable, and manufacturable platform for applications in robotics, autonomous systems, and structural monitoring.