Learning Multiple Behaviors through Viscoelastic Adaptation in Soft Robotic Materials

Learning is often associated with brains and neural circuits, yet many natural systems without a nervous system - ranging from unicellular organisms to soft-bodied animals - exhibit adaptive behavior. Inspired by this idea, we explore physical learning in robotic systems, where the dynamics emerges from the system’s own local couplings rather than from an external controller. We use a dynamical learning rule for mechanical model systems composed of coupled hinges. Learning occurs through delayed feedback during a training phase and is later retrieved as an emergent collective behavior. In this framework, the system learns effective elastic couplings, including potentially non-reciprocal ones, but seems limited to encoding a single emergent behavior. To overcome this constraint, we extend the expressivity of the system by allowing it to learn not only its elastic but also its viscous response. By introducing a multi–time–scale–dependent learning rule, the system can encode and retrieve multiple dynamical behaviors. This approach establishes a route toward the design of autonomous, adaptive, and programmable soft robotic materials capable of learning without a centralized computer.