Cloud Robotics for Data-Driven Robotic Manipulation Research at Scale

ABSTRACT:  In this talk, I will discuss the challenges and opportunities surrounding large scale experimentation for machine learning driven robotic manipulation research. While it is accepted wisdom that the performance of state of the art machine learning algorithms in fields like language modelling and computer vision scales directly with available input data, compute and model size, many open questions regarding these scaling laws remain when it comes to physical interactions of robotic systems with the world, where the complexity of data-driven learning may depend heavily on the physical characteristics of a given robotic manipulation task. I will discuss some of our research in this direction as well as a new open source robotic system we are developing which will allow large scale execution of robotic experiments in a massively parallelized manner using a Cloud Robotics approach.

BIO: Florian Pokorny is an Associate Professor of Machine Learning at the Division of Robotics, Perception and Learning here at KTH. His current research focuses on two main directions: 1) machine learning algorithms with a geometric or topological flavour or which utilize insights about geometry in the context of other methods such as deep learning and 2) machine learning methods that are tailored for robotic manipulation or motion planning and which incorporate available domain knowledge and information about physics and configuration space geometry in order to be data efficient. You can read more on his homepage: www.csc.kth.se/~fpokorny/