Nonlinear adaptive neural controllers

Contemporary machine learning has been very successfully applied to processing static images and words in consumer applications, resulting in billions of dollars in recent acquisitions of machine learning companies by Microsoft, Amazon, Facebook, and Google. However, applications to dynamic information (e.g. movies, controlling robotics) has been less well-developed. In this project, will develop and apply a novel machine learning method to neural control system for a sophisticated robotic arm. We will use hierarchical optimal neural control, dynamic trajectory generation, and non-linear adaptive methods. These same algorithms lay the foundations for processing dynamic perceptual information as well. The methods allow for the generation of neural network controllers that require limited or no knowledge of the system, allow for one-shot learning, provide generalizable trajectory generation, exhibit online error correction, and provide a natural implementation in neuromorphic hardware. This will provide the company with a clear lead in state-of-the-art controllers for robotics applications.

Faculty Supervisor:

Dr. Bryan Tripp

Student:

Travis DeWolf

Partner:

Applied Brain Research

Discipline:

Engineering

Sector:

Life sciences

University:

University of Waterloo