Control Group and OHMC Seminar - Bernd Porr, University of Glasgow

Speaker: Bernd Porr, University of Glasgow 

Title: Adaptive closed loop control: from reflexes to planning

Abstract: Closed loop control usually focuses on the control loop itself. However, in this talk I start from the disturbance acting upon the control loop which ultimately motivates the control loop. In robotics disturbances are important as they represent a distinct feature in the environment the agent needs to act upon for example an obstacle but also a target location such as food. Learning to predict these reduces uncertainty and at the same time builds a model of the environment. I'm going to present different adaptive control strategies from ISO, ICO over CLDL to FCL for behaviour based robots and also biped walking. However, all these encode essentially only one behaviour. In our current research we argue that one can use the concept of disturbances for planning because ultimately they motivate transient control loops. Thus loops with a clearly defined lifetime. This concept can be used to reason about a situation the robot finds itself in where chains of temporary control loops can solve a planning problem and/or reach a target location in the presence of obstacles. All of the above have been implemented on real robots.

Short bio: Bernd Porr has his undergraduate Physics degree from the University of Bochum and his PhD from the University of Stirling in computational neuroscience. Since 2004 he has been at the University of Glasgow and is now senior lecturer at the School of Engineering. His early research was about adaptive learning for behaviour based robots. He then applied this to biped walking and invented the RunBot. He then worked on modelling and actual deep brain stimulation of the emotional system of the brain. He then moved on to directional hearing aids and audio perception. This gave rise to an app against Tinnitus, called Tinnitus Tailor for which he founded the company Glasgow Neuro. More recently he worked on approaches of how to use deep neuronal networks for adaptive learning in robotics but also to use it for adaptive filtering of noisy biomedical signals. His current research is in robotic planning and how to encode the environment in an efficient way on a real robot.

We strongly encourage you to attend the seminar in person. You can also connect via Zoom:

https://newnham.zoom.us/j/92544958528?pwd=YS9PcGRnbXBOcStBdStNb3E0SHN1UT09