Inductive biases for interpretable robot learning and control
End-to-end learning has become ubiquitous in robotics research, and recent results show that complex tasks can be solved using vision-based control policies identified using reinforcement learning. However, if robots are to be deployed and trusted in safety critical, real-world environments, there is a need for verifiable and interpretable models. This talk will present recent results on extracting explicit programs from deep learning models, moving from model explanation to program synthesis using hybrid systems, and argue that computer-program-like control systems are more interpretable and generalisable than their end-to-end learning counterparts. I will also show that including more structure into end-to-end learning models not only allows for more interpretable models, but also provides improvements in extrapolation performance, and will discuss ways of including these inductive biases into the learning process, introducing switching density networks for switching controller identification, and an example using physics as inverse graphics. Finally, I will discuss ongoing work in the surgical robotics domain, where we are attempting to build safe surgical robot assistants with interpretable and predictable behaviour.
Michael Burke is a Research Associate at the Institute of Perception, Action and Behaviour at the University of Edinburgh. Prior to this, he led a team of 20 staff and students working in computer vision, machine learning and field robotics R&D at the Mobile Intelligent Autonomous Systems group at the Council for Scientific and Industrial Research (CSIR), South Africa, which he first joined in 2009. He holds a visiting lecturer position at the University of Witwatersrand and has a PhD in statistical signal processing from the University of Cambridge, a Masters of Science in electronic engineering from Stellenbosch University and a Bachelors in electronic engineering from the University of Pretoria. His research interests are in robot perception and learning.