Self-learning Mechanical Circuits
ABSTRACT
Self-learning materials, which use their intrinsic dynamics to learn information about their environments, represent an emerging frontier in engineering. Expanding the scope of self-learning in materials requires building dynamical systems which execute specific learning tasks. Here we provide a framework for constructing self-learning materials by first introducing an elastic learning unit, which we then use to build more complex self-learning elastic networks. Our basic learning unit is a directed spring with dynamically varying stiffness whose update dynamics mimics artificial neural networks. Exploiting this mapping, we exhibit mechanical networks which learn hidden patterns within their environment. Our experimental and theoretical results demonstrate that neural network dynamics embedded in elastic materials can solve learning tasks.
Recently, we reported the invention of self-learning mechanical circuits, which interact with their environment mechanically, and continually update their internal state as the environment changes. We define mechanical circuits to be composed of adaptive springs connected in network configurations, which learn by updating their own stiffness in response to various inputs. To mimic neural network-like computations, we additionally engineer these springs to be directed, so that their stiffness update depends upon which end of the spring experiences a greater force. To our knowledge, this construction is the first realization of an adaptive directed spring (ADS), and constitutes the fundamental unit of our mechanical circuits. By combining dynamical systems theory with an experimental realization of an ADS, we demonstrate that ADS networks can integrate information about a changing environment and modulate their behavior. Crucially, this learning behavior is embodied in the mechanical domain itself, allowing for environmental interaction unmediated by a separate sensory layer. We present a combinatorial design framework and a library of mechanical circuits that exhibit a broad range of self-learning behavior.
BIG QUESTION
“How can we build a tangible model to study self-learning materials?”
Project Status
Active
Project funded through HAI Stanford
