Files in this item

FilesDescriptionFormat

application/pdf

application/pdfWALT-THESIS-2020.pdf (20MB)
(no description provided)PDF

Description

Title:Investigating pre-touch sensing to predict grip success in compliant grippers using machine learning techniques
Author(s):Walt, Benjamin Thomas
Advisor(s):Krishnan, Girish
Department / Program:Mechanical Sci & Engineering
Discipline:Mechanical Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:M.S.
Genre:Thesis
Subject(s):robotics
soft robotics
gripper
pre-touch sensing
machine learning
grip prediction
Abstract:This work explores the application of pre-touch sensing to a compliant gripper in order to navigate the last few centimeters while grasping fruit in an occluded, cluttered environment. Machine learning was used in conjunction with pre-touch sensors to provide qualitative feedback about the success of the gripper in picking the target fruit prior to contact. Three compliant grippers were each designed to pick a specific fruit (miracle berries, cherry tomatoes and small figs) without damaging them. These grippers were designed to be mounted on the hybrid soft-rigid arm of a mobile field robot. An IR reflectance, time of flight and color sensor were used as pre-touch sensors and arranged on the gripper in various combinations to explore the contribution of each sensor. The gripper-sensor system was trained by positioning it relative to a dummy fruit using a 6 DOF arm and gripping the target. Using the training data, five machine learning methods were explored: nearest neighbor, decision trees, support vector machines, multi-layer perceptrons and a naive Bayes classifier. The various sensor configuration-machine learning combinations were tested and evaluated based on their ability to predict grip success. Additional training was conducted to demonstrate the ability to differentiate fruit from foreign matter (e.g. leaves) that are in the gripper opening. Time of flight sensors using nearest neighbor and support vector machines along with the set of all three sensors using support vector machines and multi-layer perceptrons showed the highest prediction precision (= 90%) with the color sensor playing a key role in detecting foreign objects. The machine learning methods were similar in their ability to predict grip success with nearest neighbor showing the best overall results, while sensor ‘richness’ play an important role in differentiating the sensors with the three sensor combination showing the best results.
Issue Date:2020-07-22
Type:Thesis
URI:http://hdl.handle.net/2142/108527
Rights Information:Copyright 2020 Benjamin Thomas Walt
Date Available in IDEALS:2020-10-07
Date Deposited:2020-08


This item appears in the following Collection(s)

Item Statistics