CS Events Monthly View
Qualifying ExamMotion Planning and System Identification for Reliable Robot Actions |
|
||
Friday, May 12, 2023, 03:00pm - 05:00pm |
|||
This talk will present a sequence of efforts on robot motion planning and modeling with the long-term objective of reliable robot operation via the lifelong adaptation of internal models used for planning and control.
In particular, I have worked on the sampling-based AO-RRT motion planner as part of a collaborative effort, which achieves probabilistic completeness and asymptotic optimality guarantees. I have also assisted in the integration of machine learning in motion planning to improve the computational efficiency of sampling-based planners. There are challenges, however, in the practical application of such planners on real robots due to modeling errors. The errors can be due to state observation noise as well as unmodeled or erroneously estimated parameters regarding the environment or the robot itself. I have looked into two possible directions to alleviate these issues.
The first direction explores feedback-based motion planning and control as an alternative to computing nominal trajectories. In particular, we showed that topological tools, such as Morse graphs, can be used for Region of Attraction (RoA) estimation in an explainable way. Morse Graphs can be leveraged for controller synthesis for robust feedback-based planning even in the presence of disturbances.
The second direction aims to reduce model noise. Given an unknown but uniform environment, we proposed to estimate friction parameters in a low-cost mobile robot to decrease the sim-to-real error. Furthermore, ongoing work considers parameters to be a non-uniform function of the environment. Inspired by the success of factor graphs for sensor fusion and state estimation, the approach computes friction maps to achieve environment parameter estimation. A friction map computes the friction potential that best explains the robot's behavior in a given environment. My future plans are to integrate such parameter estimation tools with motion planning methods to achieve reliable and adaptive robot behavior over long deployment periods.
Speaker: Edgar Alejandro Granados Osegueda
Location : SPR-403 (1 Spring Street, New Brunswick, NJ)
Committee:
Professor Kostas Bekris (Advisor)
Professor Abdeslam Boularias
Professor Mridul Aanjaneya
Professor Yipeng Huang
Event Type: Qualifying Exam
Abstract: See above
Organization:
Rutgers University
School of Arts & Sciences
Department of Computer Science
Contact Professor Kostas Bekris