ePrints.FRI - University of Ljubljana, Faculty of Computer and Information Science

Robot learning and planning for pushing objects

Miha Troha (2010) Robot learning and planning for pushing objects. EngD thesis.

[img] PDF
Download (3832Kb)


    In this work we first examine the paradigm of learning by experimentation. We then detail certain known algorithms which enable learning by experimentation with a view to their implementation in a intelligent robot system. We direct particular focus toward the applicability of qualitative models. We first review Šuc’s QUIN algorithm, which can be used to induce quantitative models. We then proceed by examining the Qfilter algorithm, which utilizes qualitative models to enhance numerical predictions. Further, we discuss the basic concepts of planning. We first review the widely used STRIPS language and certain extensions of it. We continue with an examination of planning under uncertainty. Its high computational complexity propels us to develop a new approach to planning, called qualitative planning. This idea is, according to our knowledge, new and can thus be considered as the most valuable part of this work. We proceed with a definition of a real world problem which will serve as a base to test the previously described algorithms. The problem definition comprises a robot and an object. The main problem is defined as a shift of the object from its current to a desired location conducted by the robot. The solution consists of two steps. In the first step, models of pushing the box need to be constructed. The models are then applied to perform shifting tasks in the second step. In a detailed description of the first step, we first present our strategy for learning qualitative and numerical models of pushing the box. The strategy utilizes the already known QUIN and Qfilter algorithms to induce the models. The evaluation of the learned models is also afforded. We consider the qualitative models as intuitively understandable. In view of the usage of the models in planning, both have proven correct. We proceed with a detailed description of the second step, which is called planning and plan execution. This step is realized in two different ways. The first is qualitative planning and the second is called quantitative planning. The name of the latter stems from the fact that it does not consider qualitative models in the process of planning. We implemented both of them on our own, however, the theory for quantitative planning is already known. Finally, we give the results of qualitative and quantitative planning. The results indicate that qualitative planning is distinctively better.

    Item Type: Thesis (EngD thesis)
    Keywords: learning by experimentation, qualitative planning, planning under uncertainty, qualitative models, intelligent robot systems
    Number of Pages: 104
    Language of Content: Slovenian
    Mentor / Comentors:
    Name and SurnameIDFunction
    akad. prof. dr. Ivan Bratko77Mentor
    Link to COBISS: http://www.cobiss.si/scripts/cobiss?command=search&base=50070&select=(ID=00007767636)
    Institution: University of Ljubljana
    Department: Faculty of Computer and Information Science
    Item ID: 1094
    Date Deposited: 14 Jun 2010 10:08
    Last Modified: 13 Aug 2011 00:37
    URI: http://eprints.fri.uni-lj.si/id/eprint/1094

    Actions (login required)

    View Item