You are here: Foswiki>RoboticsWeb Web>GSLAM (13 Jan 2014, zhangl15@LAB.CS.RPI.EDU)Edit Attach

The Application of Particle Filtering to Grasping Acquisition with Visual Occlusion and Tactile Sensing

Planar Grasping Testbed
Planar Grasping Testbed

Why is autonomous grasping and manipulation in unstructured environments still so hard for robots after 30+ years of research? For a robot to perform skilled grasping and manipulation, it has to have information that is hard to obtain with sufficient accuracy: the geometry of the object, estimates of physical quantities (such as weight and friction), and the position of the object and contacts.

The best scenario for a robot is that it can get an accurate physical model from a database, has a vision system that can track the object, and tactile sensors that can aid tracking when the fingers occlude the visual tracking features. Even in this case, the localization errors of the perception system and positioning errors allowed by the control system may not be small enough to ignore during the process of grasp acquisition or subsequent manipulation. These errors can cause the hand to bump the object accidentally when reaching or curling the fingers, possibly causing the object to slip and tumble out of the grasp. The problems are magnified when the physical quantities are known only roughly, or vary over space and time as friction parameters are known to do.

This paper is motivated by the idea that one can dramatically advance the state of the art in grasping and manipulation by designing algorithms that can estimate a physical model of a grasping system (composed of hand, object, and environment), while accurately tracking the object. We will refer to this problem as the G-SL(AM)^2 problem.The G-SL(AM)^2 problem is to autonomous robotic grasping, what the SLAM problem is to autonomous robotic mobility. The G stands for Grasping. SL(AM)^2 stands for: Simultaneous Localization, and Modeling, and Manipulation. The word\x93Modeling\x94implies that the robot will use its sensor systems (tactile, visual, and kinesthetic) to build and improve a model of the object.\x93Manipulation\x94implies that the robot will physically manipulate the object to help accomplish the modeling task.\x93Localization\x94implies that the robot will track the pose of the object during grasp acquisition and manipulation.\x93Simultaneous\x94implies that localization, modeling, and manipulation will all occur together \x96 in real time.

The G-SL(AM)^2 problem is filtering problem that presents special challenges peculiar to systems with intermittent contact: a nonsmooth dynamic model, a highdimensional state-space, and unknown contact friction parameters that vary unpredictably. To complicate things further, the dimension of the state-space varies in time as contacts form and break, and each such event effectively changes the structure of the dynamic model.

Our approach to the G-SL(AM)^2 problem is based on Bayesian filtering, in particular, particle filtering. A straight forward application would be to directly sample over the whole state space, which will be bound to fail in our problem, because in grasping, large portions of the state space are invalid. Specifically, particles should not be chosen that correspond to overlap between the geometric models of the bodies, nor should they correspond to contact forces that are not within their respective friction cones. Some may argue that a trial-and-error approach will help to recognize the correct samples, while we insist that such a method would lead to very few effective samples when contact is present, thereby causing poor estimation. To shed light on these issues, we present a case study of a particular scaled-down version of the G-SL(AM)^2 problem, using data from our planar grasp acquisition testbed shown in figure above. The main advantages of using this testbed for our initial study are the lower-dimensional state space and the smaller number of unknown model parameters. In this work, it is assumed that the geometric models and most parameters of the physical model are constant and known; only four friction parameters are assumed unknown.

Experiment Results

2d GSLAM Results


Li (Emma) Zhang and Jeffrey C. Trinkle. The Application of Particle Filtering to Grasping Acquisition with Visual Occlusion and Tactile Sensing. Technical Report 10-09, Department of Computer Science, Rensselaer Polytechnic Institute
author = {Li (Emma) Zhang and Jeffrey C. Trinkle},
institution = {Department of Computer Science, Rensselaer Polytechnic Institute},
title = {The Application of Particle Filtering to Grasping Acquisition with Visual Occlusion and Tactile Sensing},
year = {2010},
number = {10-09},


Please download all the videos in the attachment and put them in the same folder as the slides

-- LiZhang - 2011-03-01
Topic attachments
I Attachment Action Size Date Who Comment
G-SLAM.pptxpptx G-SLAM.pptx manage 1 MB 07 Jun 2011 - 20:59 UnknownUser Emma Zhang's Presentation Slide on G-SLAM at ICRA 2011
G-SLAM_ICRA_2011.pptxpptx G-SLAM_ICRA_2011.pptx manage 1 byte 09 Jun 2011 - 15:00 UnknownUser  
expSetup.jpgjpg expSetup.jpg manage 480 K 07 Jun 2011 - 21:09 UnknownUser  
new_demo1.aviavi new_demo1.avi manage 3 MB 01 Mar 2011 - 07:49 UnknownUser  
Topic revision: r8 - 13 Jan 2014, zhangl15@LAB.CS.RPI.EDU
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Foswiki? Send feedback