Learning statical and dynamical behaviors in robotic metamaterials

Synthetic materials and robots typically exhibit pre-programmed behaviors. A fundamental challenge, however, is how to give these artificial systems the ability to learn desired behaviors. In this work, we use physical learning—a decentralized learning method that adjusts building blocks based on stimuli to enable physical systems to evolve and learn—to train both the static and dynamic responses of robotic metamaterials. This approach allows metamaterials to learn, forget, and relearn static shape changes. We can do so with multiple shapes and have multiple stable states. Crucially, by introducing non-reciprocal interactions between building blocks, our metamaterials can also learn non-reciprocal static shape changes, and dynamic limit cycles. We will show our system can emulate locomotion of living matter and achieve various demanded applications in a real-world experimental platform. This work is key for advancing the next generation of smart materials and robots.