Integration of Local Neuromorphic Circuits in Soft Robotic Systems to Realize Learning and Memory

Organic electrochemical transistors (OECT) have shown to be promising components for neuromorphic computing. In particular, their ability to control stable conductance states over a large range, which can be used to emulate the synaptic weight in artificial neural networks (ANN); an important component in translating the network from digital to hardware, as to reduce operation power and to circumvent the von Neumann bottleneck. Simultaneously, the field of soft robotic systems has achieved inherently adaptive actuators by using compliant, soft materials, which makes them promising candidates to be used in non-conventional environments and for complex applications. However, they are typically pre-programmed and are subsequently not capable of adapting their behavior on longer time scales. Additionally, the weight and size of typical robotic learning systems do not make it suitable for integration in soft systems. Our goal is to integrate local neuromorphic circuits, as an alternative platform to achieve decentralized learning and memory in soft robotic systems for performance optimization and environment adaptability. Typical software-based ANN learning algorithms make use of backward propagation and gradient descent, the latter being particularly hard to achieve in hardware. A possible alternative route for the hardware-based ANN is to make use of a physical neural network (PNN), where the physical processes and local rules of the circuit are utilized to train the network.