Shape-controllable non-reciprocal robotic metamaterial

Adaptive materials learn to adapt their response to external stimuli through a decentralized training protocol. Recent work has proven the success of contrastive learning in training passive, in-equilibrium systems. However, the potential of this physical learning approach in active, out-of-equilibrium systems remains unexplored. Here, we adapt the contrastive learning protocol to an active system, our non-reciprocal robotic metamaterial, and show that it can actively learn and generate desired complex shapes. Crucially, the active nature allows our material to achieve non-reciprocal and bi-stable responses, both of which are not possible in passive materials. Our work is a first step towards learning in active physical systems, which are key to developing the next generation of smart materials and soft robots.