Date of Award
Doctor of Philosophy (PhD)
Robotic materials are a novel class of materials that tightly integrate sensing, computing, and actuation into an engineered material or composite to allow the behavior of the material to be defined algorithmically. Robotic materials are constructed using an embedded network of computing nodes based on small, inexpensive microcontrollers. Examples of such materials include morphable airfoils which change shape in response to flight conditions or mission parameters, robotic skins with rich tactile sensing capabilities that recognize texture or touch gestures, clothing with tightly integrated sensing to assist with or augment the wearer's perception of the environment, or materials with dynamic camouflage capabilities.
In this thesis, I develop a framework for in-material processing which tightly couples modularized deep neural networks and high-bandwidth sensors using a network of embedded, material-scale components. This framework enables materials to learn multiple desired responses to stimuli, avoiding the need for accurate modeling of the dynamics of the material and stimuli.
I utilize a modular neural network design consisting of convolutional (CNN) and long short-term memory (LSTM) layers implemented in each node in the material as a computational approach for robotic materials. This network architecture allows for nodes in the material to process local sensor values, maintain local state information, and communicate with nodes in a local neighborhood in the materials. A multiobjective optimization approach is employed to automatically design the neural network architectures which maximizes the performance of the network while ensuring hardware budgets, such as memory requirements, are maintained. A communication network design is also developed to allow network modules to learn a communication protocol that limits communication to a desired rate, ensuring in-network bandwidth constraints are maintained.
I demonstrate the suitability of this computational model for robotic materials using examples in several domains. An RF-based e-textile gesture input device capable of distinguishing between user control gestures is used to control arbitrary external devices. A tire with embedded piezoelectric sensing capabilites for use in high-performance autonomous vehicles performs state-of-the-art identification of terrains driven on. Two robotic skins are presented---one which is capable of detecting and localizing contact, and identifying the texture of the contacting objects; and a second which assists with avoiding collisions with obstacles and identifies affective touch gestures performed by a human collaborator. Finally, a distributed approach to human activity recognition is presented whose activity identification performance is comparable to a centralized approach, but can be implemented on hardware designed for wearable applications, as opposed to a GPU-enabled device. The examples shown demonstrate that robotic materials can perform significant in-material processing; are loosely coupled from a host system, communicating a minimal number of low-bandwidth events to the host; and can exhibit multifunctional behavior that is analyzed for safety or performance considerations.
Hughes, Dana, "In-Material Processing of High Bandwidth Sensor Measurements Using Modular Neural Networks" (2018). Computer Science Graduate Theses & Dissertations. 176.