Upravljanje podatnom robotskom hvataljkom bazirano na učenju

master's thesis

• Title (english): Soft sensing robot finger control based on model learning
• Author: Ivona Krajačić
• Mentor: Matko Orsag (mentor)
• Committee member: Matko Orsag (predsjednik povjerenstva)
• Committee member: Stjepan Bogdan (član povjerenstva)
• Committee member: Zdenko Kovačić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering

Abstract

Cilj ovog rada bio je izrada senzora dodira čija je glavna značajka gumena opna. Deformmacija opne je najprije simulirana u platformi Unreal Engine, te su u tu svrhu kreirana dva modela opne, s točkastim i rešetkastim uzorkom. Deformacija opne nastale u kontaktu s testnim prizmatičnim tijelom pod određenim kutom i dubinom prodiranja, analizirane su metodama dubokog i strojnog učenja. Pozicije centroida točkaste opne služile su kao ulazni primjeri multivarijante višestruke linearne regresije, dok su slike s oba tipa opne bile ulazni primjeri potpuno povezane duboke konvolucijske mreže s dva regresijska izlaza.
Što se tiče fizičke realizacije senzora, gumena opna se nalazi na vrhu kučišta čija je gornja polovica ispunjena gelom. Gel je od ostatka kučišta odvojen pleksiglasom ispod kojeg se nalazi Rasberry Camera V2 i LED diode koje osvjetljuju unutrašnjost senzora. Kamera je flat kabelom spojena na Rasberry Pi računalo na kojem je instaliran operacijski sustav Ubuntu, a komunikacija s kamerom ostvarena je uz pomoć ROS paketa raspicam_node. Prikupljanje podataka za učenje ostvareno je korištenjem robotskog maipulatora Franka. Slike su predstavljale ulazni skup primjera za učenje potpuno povezane konvolucijske duboke mreže s dva regresijska izlaza: vanjska sila i dubina prodiranja.

Abstract (english)

The aim of this thesis was to make a touch sensor whose main feature is a rubber membrane. Membrane deformation was first simulated in the Unreal Engine platform, and two membrane models were created for this purpose, with a point and grid patterns. The deformation of the membrane formed in contact with the test prismatic body at a certain angle and depth of penetration were analyzed by deep and machine learning methods. The positions of the centroids of the point membrane served as input examples of the multivariate multiple linear regression, while the images from both types of membrane were input examples of a fully connected deep convolutional network with two regression outputs.
As for the physical realization of the sensor, the rubber membrane is located on top of the sensor box whose upper half is filled with gel. The gel is separated from the rest of the case by plexiglass below which is a Rasberry Camera V2 and LEDs that illuminate the inside of the sensor. The camera is connected with flat cable to a Rasberry Pi computer on which the Ubuntu operating system is installed, and communication with camera is achieved with the help of the ROS package raspicam_node. Data collection for learning was accomplished using the robotic maipulator Franka. The images represented an input set of examples for learning a fully connected convolutional deep network with two regression outputs: external force and depth of penetration.

• Keywords: podatno upravljanje; strojno učenje; duboko učenje
• Keywords (english): soft sensing; machine learning; deep learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:695741
• Study programme: Title: Electrical Engineering and Information Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije (magistar/magistra inženjer/inženjerka elektrotehnike i informacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-15 10:35:10