Application of Particle Filtering to the G-SLAM Problem with Visual Occlusion and Tactile Sensing – Københavns Universitet

Application of Particle Filtering to the G-SLAM Problem with Visual Occlusion and Tactile Sensing

DIKU Talk by Jeff Trinkle, Professor of Computer Science, Directory of the CS Robotics Laboratory, Rensselaer Polytechnic Institute, Troy, NY 12180

Abstract

During a robot's attempt to reach toward and grasp an object, its hand will partially occlude its view of the object.  Then, as the robot attempts to complete the grasping process, control and modeling errors can cause accidental bumping of the object, and ultimately, grasp failure.  To avoid this problem, we develop a Bayesian filter that fuses vision and tactile data to simultaneously track the pose of the object and hand, while estimating important model parameters (e.g., friction coefficients) during the grasp acquisition.  Bayesign filters produce the best results with accurate mathematical models of the process being filtered.  In grasping, this suggests that the mathematical model must capture at least the two important physical phenomena: stick-slip friction and non-penetration constraints between the bodies.  These features make the model non-smooth and lead us to the choice of a particle filter.  In this talk, I will present the development of the particle filter and results obtained from applying it to a simplified planar grasping system.

BioSketch:

Jeff Trinkle received bachelor's degrees in Physics (1979) and Engineering Science and Mechanics (1979) from Ursinus College and Georgia Institute of Technology, respectively. In 1987, he received his PhD from the Department of Systems Engineering at the University of Pennsylvania, where he was a research assistant in the GRASP Laboratory. 

Since 1987, he has held faculty positions in the Department of Systems and Industrial Engineering at the University of Arizona and the Department of Computer Science at Texas A&M University.  From 1998 to 2003 he was a visiting research scientist at Sandia National Laboratories in Albuquerque New Mexico.  He moved to Rensselear Polytechnic Institute in Troy, New York, in 2003, where he served as Chair of the Computer Science Department until 2009.  He is now Professor of Computer Science, Director of the CS Robotics Lab, and Faculty Dean of Residential Commons.

Dr. Trinkle's primary research interests lie in the areas of robotic manipulation, multibody dynamics, and automated manufacturing.  Under the continuous support of the National Science Foundation since 1989, he has written many technical articles on theoretical issues underpinning the science of robotics and automation.  One of these articles was the first to develop a now-popular method for simulating multibody systems.  Variants of this method are key components of several physics engines for computer game development, for example, NVIDIA PhysX and the Bullet Physics Library.

Dr. Trinkle has served on the editorial boards of IEEE Transactions on Robotics and Automation, IEEE Transactions on Automation Science and Engineering, and Robotica.  He is the recipient of the 1985 IBM Graduate Research Fellowship, the 1989 Research Initiation Award from the National Science Foundation, the 1994 Texas A&M Center for Teaching Excellence Award, the 1998 Plank Company Faculty Fellowship, the 2004 Kayamori Best Automation Paper of the IEEE International Conference on Robotics and Automation, and the 2009 Humboldt Research Prize.  He spent the 2009-2010 academic year as a Humboldt Fellow at the Institute for Mechatronics and Robotics at the German Aerospace Center and the Institute for Applied Mechanics at Technical University of Munich.  In 2010 he became a fellow of the IEEE for his research contributions to robotic grasping and dexterous manipulation.

Scientific host: Kenny Erleben

DIKU Talks are open to all interested parties. All are welcome - participation is free.