On Human Actions and Robot Skills

Abstract:

In this lecture we briefly discuss how human actions can be modeled so that a robot can not only recognize but also understand them, and how skills can/should be modeled for robot control. In particular, we very briefly review different approaches taken in computer vision and robotics. We touch briefly on concepts such as affordances, scene states, object-action complexes, action primitives, imitation learning, etc., and we identify and apply useful approaches from computer vision in robotics. We will discuss the use of parametric HMMs for modeling human movements, Latent Dirichlet Analysis (LDA) and classical GMMs for modeling robot movements. Finally, we will discuss our implementation on a mobile robot platform and the experiments that were carried out at Grundfos A/S in Denmark.
I will conclude the talk with an outlook on suitable approaches for continued developments and future research directions.