Continuous State POMDPs for Object Manipulation Tasks

Abstract

My research focus is on using continuous state partially observable Markov decision processes (POMDPs) to perform object manipulation tasks using a robotic arm. During object manipulation, object dynamics can be ex-tremely complex, non-linear and challenging to specify. To avoid modeling the full complexity of possible dy-namics, I instead use a model which switches between a discrete number of simple dynamics models. By learn-ing these models and extending Porta’s continuous state POMDP framework (Porta et al. 2006) to incorporate this switching dynamics model, we hope to handle tasks that involve absolute and relative dynamics within a sin-gle framework. This dynamics model may be applica-ble not only to object manipulation tasks, but also to a number of other problems, such as robot navigation. By using an explicit model of uncertainty, I hope to cre-ate solutions to object manipulation tasks that more ro-bustly handle the noisy sensory information received by physical robots.