Hostname: page-component-7b7567c8c-wxvwp Total loading time: 0 Render date: 2026-11-16T02:21:06.439Z Has data issue: false hasContentIssue false

Selection of trajectory parameters for dynamic pouring tasks based on exploitation-driven updates of local metamodels

Published online by Cambridge University Press:  08 May 2017

Joshua D. Langsfeld
Affiliation:
Maryland Robotics Center, Institute for Systems Research, University of Maryland, College Park, MD, USA. E-mail: jdlangs@umd.edu
Krishnanand N. Kaipa
Affiliation:
Department of Mechanical and Aerospace Engineering, Old Dominion University, Norfolk, VA, USA. E-mail: kkaipa@odu.edu
Satyandra K. Gupta*
Affiliation:
Center for Advanced Manufacturing, Department of Aerospace and Mechanical Engineering, University of Southern California, Los Angeles, CA, USA
*
*Corresponding author. E-mail: guptask@usc.edu
Rights & Permissions [Opens in a new window]

Summary

Core share and HTML view are not available for this content. However, as you have access to this content, a full PDF is available via the 'Save PDF' action button.

We present an approach that allows a robot to generate trajectories to perform a set of instances of a task using few physical trials. Specifically, we address manipulation tasks which are highly challenging to simulate due to complex dynamics. Our approach allows a robot to create a model from initial exploratory experiments and subsequently improve it to find trajectory parameters to successfully perform a given task instance. First, in a model generation phase, local models are constructed in the vicinity of previously conducted experiments that explain both task function behavior and estimated divergence of the generated model from the true model when moving within the neighborhood of each experiment. Second, in an exploitation-driven updating phase, these generated models are used to guide parameter selection given a desired task outcome and the models are updated based on the actual outcome of the task execution. The local models are built within adaptively chosen neighborhoods, thereby allowing the algorithm to capture arbitrarily complex function landscapes. We first validate our approach by testing it on a synthetic non-linear function approximation problem, where we also analyze the benefit of the core approach features. We then show results with a physical robot performing a dynamic fluid pouring task. Real robot results reveal that the correct pouring parameters for a new pour volume can be learned quite rapidly, with a limited number of exploratory experiments.

Information

Type
Articles
Copyright
Copyright © Cambridge University Press 2017