Speaker: Okan Koç, Sony AI
Title: Robust and Efficient Robot Learning in Highly Dynamic Tasks
Abstract: In this talk I will give an overview of the robot learning problems I tried to tackle during my PhD and afterwards. The research experiments span multiple robotic labs and environments, from Robot Table Tennis
in Tuebingen to the Humanoid upper-torso in ATR, Japan and the Amazon Scout robot (sadly no longer operational). After briefly introducing an optimization based trajectory generation approach to Robot Table Tennis, I will present a learning control algorithm that learns totrack these striking trajectories well using Iterative LearningControl. The proposed algorithm 'bayesILC' is model-based and adapts local linear models online while taking advantage of their covariances during the robust control updates. I will then talk about a
learning-from-demonstrations framework, where sparse movement primitive parameters can be extracted from demonstrations. An alternating optimization approach will be presented that iterates between a multi-task Elastic Net (for regression parameters) and nonlinear optimization (to adapt the features). The learned movement primitives are tested on a table tennis serve task. In the last part of my talk, I will summarize various research conducted at multiple locations, with my collaborators or at industry, and highlight some common issues faced in these diverse problem settings.
