izpis_h1_title_alt

POSPEŠENO UČENJE ROBOTOV ZA INTERAKCIJO Z OKOLJEM IN ČLOVEKOM NA PODLAGI SENZORIČNO-MOTORIČNEGA UČENJA
PETERNEL, LUKA (Author), Babič, Jan (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (10,68 MB)

Abstract
V prihodnje je premik robotov iz strukturiranih/nadzorovanih industrijskih in laboratorijskih okolij v človekov vsakdanjik neizogiben. Za uspešno vpeljavo robotov v naš vsakdanjik je treba rešiti dva osnovna problema, povezana z izvajanjem takih nalog. (1) Naloge, potrebe, orodja in okolje se razlikujejo med uporabniki. To zahteva pridobitev velike baze zelo specifičnih znanj, zato mora biti prenos znanja z uporabnika na robota enostaven, prilagodljiv in hiter. (2) Pri izvedbi večine vsakdanjih nalog se srečujemo s žičnimi interakcijami z nepredvidljivim in nestrukturiranim okoljem. Zato je pridobivanje modelov okolja zapleten postopek, kar oteži vodenje robota s klasičnimi pristopi. Za dobro delovanje robotov v takem okolju moramo najti alternativne načine učenja žičnih interakcij robota z okoljem. V prvem poglavju je najprej predstavljen uvod v tematiko s pregledom dosedanjega dela na tem področju. Nato so predstavljeni glavni cilji disertacije. Na koncu poglavja pa je povzetek vseh eksperimentov, ki so bili izvedeni v okviru disertacije. V drugem poglavju je najprej splošno predstavljena metoda učenja robotov z vključitvijo človekovega senzorično-motoričnega sistema v robotovo regulacijsko zanko. Ta temelji na sposobnosti človekovega senzorično-motoričnega učenja, ki mu omogoča prilagoditev na vodenje robota in kasnejši prenos pridobljenega znanja. Sledi opis metod za predstavitev strategije vodenja (znanja). Te obsegajo senzorično-motorični pare, trajektorije gibanja in adaptivne oscilatorje za opis stanja periodičnega gibanja. V zadnjem sklopu pa sta predstavljeni dve metodi strojenega učenja, ki sta bili uporabljeni za robotsko učenje pridobljenih strategij vodenja (Gaussov regresijski proces in lokalno utežena regresija). V tretjem poglavju je predstavljena metoda učenja humanoidnih robotov za primerne odzive telesa na žične interakcije s človekom in z okoljem. Pri tem je bila razvita metoda za pretvorbo senzoričnih informacij o stanju robotovega telesa v senzorično vzbujanje človeka. S tem je bila demonstratorju posredovana potrebna povratna informacija o stanju dinamike robotovega telesa med učnim postopkom. V okviru tega je bil razvit poseben haptični vmesnik, ki je izvajal silo na telo demonstratorja. V drugem delu poglavja je predstavljena metoda sprotnega učenja robota na podlagi delitve odgovornosti med trenutno naučeno strategijo vodenja robota in demonstratorjem. Ta omogoča postopen prenos odgovornosti vodenja z demonstratorja na robota in omogoča dodatno povratno informacije o stanju učenja. V zadnjem delu pa je predlagana metoda za združevanje demonstriranih strategij vodenja robotskega telesa z vodenjem gibanja roke na osnovi inverzne kinematike. V četrtem poglavju so predstavljene predlagane metode učenja robotske manipulacije z nestrukturiranim in nepredvidljivim okoljem. Te metode temeljijo na zmožnosti demonstratorja direktne modulacije in učenja impedance robotske roke. V ta namen so bile razvite metode, ki omogočajo demonstratorju vodenje togosti v realnem času. Pri tem smo reševali naloge, povezane z uporabo elementarnih orodij, s sodelovanjem robota z uporabnikom in sestavljanjem predmetov. Te lastnosti so ključne pri prihodnjem delovanju robotov v človekovem vsakdanjiku ali pri njihovi udeležbi pri raziskovanju vesolja, kjer so sredstva omejena. V petem poglavju je predstavljena metoda vodenja eksoskeletov. Ti mehanizmi obdajajo dele človeškega telesa in direktno pomagajo pri gibanju v sklepih. V okviru vpeljave robotov v vsakdanjik eksoskeleti predstavljajo komplement humanoidnim robotom, katerih namen je nuditi pomoč človeku na bolj posrednem nivoju. Predlagana metoda vodenja temelji na minimizaciji uporabnikove mišične aktivnosti prek adaptivnega učenja podpornih sklepnih navorov, ki jih izvaja eksoskelet. Glavna prednost metode je, da ne potrebuje modelov človeka in robota. Potrebni kompenzacijski navori se nenehno prilagajajo trenutnim pogojem. Metodo smo preizkusili z eksperimenti na več subjektih in pri tem analizirali medsebojno prilagajanje eksoskeleta ter uporabnika. V zadnjem poglavju sledi zaključek, v katerem so povzeti glavni prispevki disertacije k znanosti.

Language:Slovenian
Keywords:Človek v Regulacijski Zanki, Učenje Robotov, Fizična Interakcija, Adaptivno Vodenje Eksoskeletov, Haptični Vmesnik
Work type:Doctoral dissertation (mb31)
Organization:FE - Faculty of Electrical Engineering
Year:2015
Views:1741
Downloads:381
Metadata:XML RDF-CHPDL DC-XML DC-RDF
 
Average score:(0 votes)
Your score:Voting is allowed only to logged in users.
:
Share:AddThis
AddThis uses cookies that require your consent. Edit consent...

Secondary language

Language:English
Title:ACCELERATED ROBOTIC LEARNING FOR INTERACTION WITH ENVIRONMENT AND HUMAN BASED ON SENSORY-MOTOR LEARNING
Abstract:
It is inevitable that the robots will move from the structured and controlled environments, such as industry and laboratories, into our daily lives. To achieve the integration of robots into our daily lives, we have to solve two fundamental problems related to these tasks. (1) Each individual robot user has a different environment, needs, tasks and tools. This demands acquisition of a vast amount of very specific skills, thus the transfer of the skill from the user to the robot must be intuitive, adaptable and fast. (2) Many daily tasks require us to deal with physical interaction with unpredictable and unstructured environment. Because of this, the acquisition of models is very complex and makes the classical robot control difficult. To make robots successfully operate in such environment, it is necessary to find alternative ways to teach the robot how do deal with such interactions. In the first part of the first chapter, there is an introduction into the research field with an overview of the state-of-the-art. The second part explains the goals of the thesis. The last part contains an overview of the performed experiments. The second chapter presents the human-in-the-loop teaching method. The method is based on human sensorimotor learning ability that allows the demonstrator to first obtain the skill necessary for controlling the robot and then transfer that skill to the robot. This is followed by a presentation of methods to encode the control strategy (skill). These methods include: sensorimotor pairs, trajectory of motion and adaptive oscillators for describing the state of periodic motion. In the last part of this chapter, we present two machine learning methods that were utilised in our methods (Gaussian Process Regression and Locally Weighted Regression). The third chapter presents the proposed method for teaching humanoid robots how to deal with physical interaction of its body with human and environment. To this end, a method for converting robot sensory information into a human sensory stimulation was developed to give the demonstrator the necessary feedback about the robot body dynamics during the teaching process. In this scope, we developed a special haptic interface that exerted forces on the demonstrator's body. After that, we present a method for on-line robot learning where the control of the robot body is shared between the currently learnt strategy and human demonstrator. This enables a gradual transfer of the control responsibility from the demonstrator to the robot and offers an additional feedback about the state of the learning. In the last part, we propose a method that allows merging human-demonstrated posture-control skill with arm motion control based on inverse kinematics solution. The fourth chapter presents the proposed methods for teaching robot how to manipulate with unstructured and unpredictable environment. These methods were based on the ability of demonstrator to modulate and teach the impedance of robotic arm. We developed methods that allow the demonstrator to control the robot's stiffness in real-time. This approach was then used to solve tasks related to use of elementary tools, human-robot cooperation and part assembly. These tasks are crucial for future robot operation in human daily lives or in their participation in space exploration, where the available means are limited. The fifth chapter presents a method for exoskeleton control. These devices are made to enclose the human body parts and directly assist the motion in the joints. In the framework of integration of robots into the human daily lives, exoskeletons are a complement to the humanoid robots that are designed to provide assistance on a more indirect level. The proposed control method is based on minimisation of human muscle activity through adaptive learning of robot assistive joint torques. The main advantage of this method is that it does not require models of human and robot. Necessary compensation torques are adaptively derived according to the current conditions. The method was validated on multiple subjects and we analysed the human-robot co-adaptation. The last chapter recapitulates the main contributions of the dissertation and presents its conclusions.

Keywords:Human-in-the-Loop, Robot Learning, Physical Interaction, Adaptive Exoskeleton Control, Haptic Interface

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Comments

Leave comment

You have to log in to leave a comment.

Comments (0)
0 - 0 / 0
 
There are no comments!

Back