Faculty Mentor
Dr. Mark Dean
Department (e.g. History, Chemistry, Finance, etc.)
Electrical & Computer Engineering
College (e.g. College of Engineering, College of Arts & Sciences, Haslam College of Business, etc.)
Tickle College of Engineering
Location
Knoxville, TN
Event Website
https://eureca.utk.edu/
Year
2019
Abstract
We discuss the development of a self-adjusted balancing robot (SABR) using a neuromorphic computing framework for control. Implementations of two-wheeled balancing robots have been achieved using traditional algorithms, often in the form of proportional-integral-derivative (PID) control. We aim to achieve the same task using a neuromorphic architecture, which offers potential for higher power efficiency than conventional processing techniques. We utilize evolutionary optimization (EO) and the second iteration of Dynamic Adaptive Neural Network Arrays (DANNA2) developed by the Laboratory of Tennesseans Exploring Neural Networks (TENNLab). For the purpose of comparison, a traditional balancing robot was first designed using PID control; the neuromorphic implementation was then designed. This work demonstrates the simplicity and flexibility of DANNA2's neural network architecture, as this framework can be simulated on a simple computing platform. As a proof-of-concept, a trained network was able to balance the physical system by simulating the network on a Raspberry Pi. We further discuss possible improvements to the system and future work implementing the system on an FPGA.
SABR: Development of a Neuromorphic Balancing Robot
Knoxville, TN
We discuss the development of a self-adjusted balancing robot (SABR) using a neuromorphic computing framework for control. Implementations of two-wheeled balancing robots have been achieved using traditional algorithms, often in the form of proportional-integral-derivative (PID) control. We aim to achieve the same task using a neuromorphic architecture, which offers potential for higher power efficiency than conventional processing techniques. We utilize evolutionary optimization (EO) and the second iteration of Dynamic Adaptive Neural Network Arrays (DANNA2) developed by the Laboratory of Tennesseans Exploring Neural Networks (TENNLab). For the purpose of comparison, a traditional balancing robot was first designed using PID control; the neuromorphic implementation was then designed. This work demonstrates the simplicity and flexibility of DANNA2's neural network architecture, as this framework can be simulated on a simple computing platform. As a proof-of-concept, a trained network was able to balance the physical system by simulating the network on a Raspberry Pi. We further discuss possible improvements to the system and future work implementing the system on an FPGA.
https://trace.tennessee.edu/utk_eureca/2019/engineering/10