Physical learning in reprogrammable metamaterials for adaptation to unknown environments

Abstract: Reprogrammable mechanical metamaterials, composed of a lattice of discretely adaptive elements, are emerging as a promising platform for mechanical intelligence. To operate in unknown environments, such structures must go beyond passive responsiveness and embody traits of mechanical intelligence: sensing, computing, adaptation, and memory. However, current approaches fall short, as computation of the required adaptation in response to changes in environmental stimuli must be pre-computed ahead of operation. Here we present a physical learning approach that harnesses the structure's mechanics to perform computation and drive adaptation. The desired global deformation response of nonlinear metamaterials with adaptive stiffness is physically encoded as local strain targets across internal adaptive elements. The structure adapts by iteratively interacting with the environment and updating its stiffness distribution using a model-free algorithm. The resulting system demonstrates autonomous real-time adaptation (~seconds) to previously unknown loading conditions without pre-computation. Physical learning inherently accounts for manufacturing imperfections and is robust to sensor noise and structural damage. We also demonstrate scalability to complex metamaterial structures and different metamaterial architectures. By uniting sensing, computation, and actuation in a mechanical framework, this work makes key strides towards embodying the traits of mechanical intelligence into adaptive structures. We expect our approach to open pathways towards in-situ adaptation to unknown environment for applications in hypersonic flight, adaptive robotics, and exploration in extreme environments.