Abstract:

Despite the strong prevalence of robotic manipulators in the manufacturing world, achieving human-level manipulation competency, and in particular flexibility, remains an open problem in robotics. Mechanically, the rise of cheap and accurate manufacturing capabilities and sensors has allowed us to create robotic arms, grippers, and bases that could theoretically perform many of the same tasks that humans can. Unfortunately, we are limited by our lack of understanding of the underlying ”programming” of human movements and what allows a person to abstract their knowledge from completing one task to completing another similar (or not so similar) task. Research into Imitation Learning (IL) as a viable approach to manipulation has been around for almost 20 years, and its predecessor, Programming by Demonstration, was investigated over 30 years ago, but with the proliferation of Machine Learning techniques, particularly Deep Learning (DL), the power of IL has increased dramatically. These learning-based algorithms have several advantages over traditional approaches to manipulation, including increased flexibility and lower requirements for hand-designed kinematic or dynamic models, path planners, and controllers. Additionally, many of these algorithms have leveraged clever probabilistic inference methods or the representational power of DL to generalize a learned policy to a variety of related tasks. This talk will cover the state of the art in this domain, which is being referred to as Multi-task Imitation Learning, as well as into some of the details of my goals and current work in the field.

Bio:

Trevor Ablett is a first year PhD student in the STARS lab at the University of Toronto Institute for Aerospace Studies (UTIAS) in Toronto, Canada. He obtained his Bachelor of Engineering from McMaster University in Hamilton, Canada, and started a Master of Applied Science in Robotics in 2016 before transferring to a PhD after his first year. After starting with work in self-calibration of manipulators during his Master's and co-authoring an ICRA'2018 paper on his work, Trevor transitioned into the domain of applying learning approaches to manipulation tasks for his PhD. Prior to starting at the University of Toronto, Trevor was working in industry at an automation company in Canada where he programmed robots for warehouse autonomy and co-authored a patent on his work, and has also previously completed a Bachelor's Degree in Psychology.