Explainable Machine Learning for Robotics

Dr. Joseph Campbell

Carnegie Mellon University

Rapid advances in machine learning have endowed robots with an increased capacity for autonomous operation. However, state-of-the-art models, e.g. deep neural networks, often contain opaque underlying representations that make it difficult to understand how and why these models make decisions. This is problematic, particularly when model decisions don’t align with human expectations, as transparent decision-making is needed to ascertain if a decision is based on sound reasoning and can be trusted. I aim to bridge this gap by developing performant machine learning models which allow robots to explain their actions to human users.

In this talk, I will first discuss principled approaches for learning transparent models which effectively balance accuracy and explainability. Next, I will introduce algorithms which leverage model explanations to improve performance, through both external intervention and model self-correction. Finally, I will show recent results in which these methods enable transparent and practical human-robot interaction. Through engaging in challenging tasks such as hugging, cooperative manipulation, and catching dynamic objects, this work represents a meaningful step towards robots that can seamlessly and transparently operate alongside humans.

 Joseph Campbell is a Postdoctoral Fellow in the Robotics Institute at Carnegie Mellon University, working with Katia Sycara. He is interested in developing smarter robots that can safely operate with and around humans. His research bridges machine learning and robotics, with a focus on developing explainable machine learning models and methods that allow robots to operate with full transparency. Before joining CMU, Joseph earned his PhD from Arizona State University under Heni Ben Amor and was a visiting researcher at the National University of Singapore and Osaka University. His work has been supported by two NSF EAPSI Fellowships and a Dean’s Fellowship from ASU.

https://columbiauniversity.zoom.us/j/93852195221?pwd=aThsWFg0KzdzM2lhTVR4VTVROGNWQT09
Meeting ID: 938 5219 5221
Passcode: 649858