Logo: University of Southern California

Events Calendar


  • PhD Defense - Christian Potthast

    Thu, Apr 28, 2016 @ 10:00 AM - 12:00 PM

    Thomas Lord Department of Computer Science

    University Calendar


    PhD Defense - Christian Potthast
    Thursday, April 28, 2016 @ 10:00 am - 12:00 pm
    Computer Science

    Title: Information Theoretical Action Selection

    Location: RTH 406

    Time: 10:00am - 12:00pm , April 28th, 2016

    PhD Candidate: Christian Potthast

    Committee members:

    Prof. Gaurav S. Sukhatme (Chair)
    Prof. Stefan Schaal
    Prof. Sandeep K. Gupta

    Abstract:

    For robots to become one day fully autonomous and assist us in our daily life's, they need to be able to self reliantly acquire information about their environment. Challenges arise from limited energy budget to operate the robot, occlusion as well as uncertainty in data captured by noisy sensor. To cope with such challenges, the robot needs to be able to rely on a system that enables him to capture efficiently information and stay well within its constraints. Furthermore, information acquisition should be reactive to sensor measurement, incorporate uncertainty and tradeoff information gain with energy usage.

    In my thesis we look at the realization of such systems using well established information theoretical quantities to formulate a framework as general and versatile as possible. Specifically, we look at the task of defining objective functions that enable us to tradeoff information with acquisition cost, enabling the robot to gather as much useful information as possible, but at the same time keep energy consumption to a minimum. We address this challenge for a variety of different tasks.

    First, we look at the problem of 3d data acquisition which is of outmost importance to a robotic system since the robot needs to know the environment it is operating in. In this work I propose a framework that enables the robot to quickly acquire information by sequentially choosing next observation positions that maximize information. Next, we look at adaptive action selection in the context of object recognition on robots with limited operating capabilities. I propose an information-theoretic framework that combines and unifies two common techniques: view planning for resolving ambiguities and occlusions and online feature selection for reducing computational costs. Concretely, this framework adaptively chooses two strategies: utilize simple-to-compute features that are the most informative for the recognition task or move to new viewpoints that optimally reduce the expected uncertainties on the identity of the object. Lastly, I present an online trajectory optimization approach that optimizes a trajectory such that object recognition performance is improved. With the idea in mind that the robot needs to make progress towards a goal a cost function is formulated formulated the that allows the robot to improve recognition performance, reduces information acquisition time while simultaneously moving towards the goal point.

    Location: Ronald Tutor Hall of Engineering (RTH) - 406

    Audiences: Everyone Is Invited

    Contact: Lizsl De Leon

    Add to Google CalendarDownload ICS File for OutlookDownload iCal File

Return to Calendar