Robotics: Science and Systems IV

Learning to Manipulate Articulated Objects in Unstructured Environments Using a Grounded Relational Representation

Dov Katz, Yuri Pyuro, Oliver Brock

Abstract: To successfully perform complex manipulation tasks in unstructured environments, a robot must be able to obtain information about the objects it manipulates. We propose an approach for the incremental acquisition of task-specific information through interactions with objects. The application of relational reinforcement learning enables a robot to continuously improve its performance in this model acquisition task. We demonstrate the effectiveness of the approach for the task of manipulating articulated objects. For these initially unknown objects, the robot has to determine an accurate kinematic representation. Our results show that the robot improves its model acquisition ability with increased experience. Experience acquired for a particular articulated object transfers to model acquisition for another object and results in significant speed-up of the model acquisition process. The effectiveness of our approach stems from the fact that the representation used for relational reinforcement learning is grounded in the task-specific perceptual and actuation capabilities of the physical robot. This grounding partitions the state space based on task-specific, real-world structure, rendering the learning problem tractable.

Download:

Bibtex:

@INPROCEEDINGS{Katz-RSS08,
    AUTHOR    = {Dov Katz, Yuri Pyuro, Oliver Brock},
    TITLE     = {Learning to Manipulate Articulated Objects in Unstructured Environments Using a Grounded Relational Representation},
    BOOKTITLE = {Proceedings of Robotics: Science and Systems IV},
    YEAR      = {2008},
    ADDRESS   = {Zurich, Switzerland},
    MONTH     = {June},
    DOI       = {10.15607/RSS.2008.IV.033} 
}