Department for Automation, Biocybernetics and Robotics





Seznam napovedanih predavanj / List of planned seminars

(povzetki so pod seznamom / abstracts under the list)

Archive:
2015/2016
2014/2015

2013 / 2014
2012/2013
sezona 2011/2012
sezona 2010/2011

Ni napovedanih predavanj. No planned seminars.
Pretekla predavanja / Past seminars - 2014/2015
05.06.2015: Aleš Ude
/ From IntellAct to ReconCell (A Reconfigurable robot workCell for fast set-up of automated assembly processes in SMEs)


In the recently concluded EU FP-7 project IntellAct (Intelligent observation and execution of Actions and manipulations), we created new technologies for facilitating the setting-up of automated assembly processes with robots. Success of IntellAct in industrial assembly applications had been demonstrated by unconstrained and unordered assembly of typical industrial benchmarks. Hence the developed technologies are now at a stage where they can be transferred to industry. ReconCell develops a new type of robot workcell, its required process infrastructures and the economic framework for automated robot assembly, especially designed for the needs of SMEs. SMEs would benefit from robotic automation, but often cannot use it due to set-up & maintenance complexity. This requires expert knowledge and time for configuration and programming, which is too costly for them. Robotic automation is, thus, normally economically infeasible for SMEs, especially for small batch sizes. ReconCell develops an easy to (re-)configure and (re-)program workcell, making robot solutions commercially viable even for small batch sizes (~1000 units) by reducing set-up & maintenance effort substantially. Specifically, the ReconCell System is based on a layered concept where we start together with the customer with business modelling the planned product assembly based on the ReconCell System to assess its economic viability and provide decisive Key Performance Indicators (KPIs). On approval, the next layer of the ReconCell System implements assembly with automated testing in simulation. Here we use reconfigurable hardware elements to design the required workcell layout and assembly processes. After verification, product assembly takes place in the real workcell under machine vision-based monitoring and novel force-based control of execution to assure product quality using the KPIs. We demonstrate the capabilities of the layered ReconCell System on three real use cases provided by the SMEs of our consortium and two more use cases, established through an open call.

11.05.2015: Nejc Likar
Implemantacija naloge potiskanja stola na realnem robotu Armar IIIb / Chair Pushing scenario implementation on Armar IIIb robot
Seminar bo predstavil delo, ki je bilo narejeno v sodelovanju z KIT in je sovpadalo s cilji projekta Xperience. Naloga potiskanja stola je zajemala reševanje problemov iz različnih področij, kot so lokalizacija stola, načrtovanje prijemanja, načrtovanje poti potiskanja ter kinematično oziroma dinamično modeliranje sistema. Naloga potiskanja stola je vsebovala tudi uporabo map delovnega prostora, katere so bile ključnega pomena pri reševanju problema načrtovanja poti. Celoten sistem je bil voden kot enovit sistem, kar je omogočalo razporeditev dela z izbiranjem vrednosti posameznih uteži utežnostne matrike.

The seminar talk will present the work which was done together with collaboration with KIT and which was included in the Xperience project. The chair pushing task included solving different problems such as chair localization, grasp planning, path planning, and kinematic (dynamic) modeling of the system. The chair pushing task also included usage of reachability maps, which were a vital component of the path planning problem. Whole system was controlled as a whole system, which enabled the work distribution between the robot joints during the chair pushing task, with selection of the individual weights of the weight matrix.

11.05.2015: Rok Vuga
/ DMPs with explicit encoding of speed profiles


In the seminar I will talk about recent development of dynamic movement primitives. We extended them to include additional factors which would enable adaptation of speed of motion without affecting its shape. We also developed a novel formulation in which the speed profile is encoded explicitly. I will present the background, the theory, and some experimens.

17.04.2015: Miha Deniša
Evalvacija podajnih elementarnih gibov / Evaluation of Complaint Movement Primitives
Seminar se bo lotil problema podajnega, a hkrati natančnega gibanja robota brez uporabe dinamičnega modela naloge. Prestavljen bo pristop za učenje novih podajnih gibov preko demonstracij izvedb naloge. Uporabljeni podajni elementarni gibi (angl.: Compliant Movement Primitives - CMPs) so sestavljeni iz dveh komponent. Trajektorijo pozicije, pridobljene preko človeških demonstracij, zapišemo kot dinamični elementarni gib (angl.: Dynamic Movement Primitives - DMPs). Pripadajočo trajektorijo navorov pa zapišemo kot linearno kombinacijo Gaussovih funkcij . Izračun novih, spremenjenim nalogam prilagojenih CMP-jev je mogoč z uporabo metod statističnega posploševanja. Večji del seminarja bo namenjen sami evalvaciji predlagane metode. Na grobo jo lahko razdelimo na štiri dele: 1) Evalvacija podajnih gibov je bila izvedena pri različnih nastavitvah togosti robota. Med izvedbo diskretne naloge smo opazovali napako sledenja. 2) Pri nepričakovanem trku robota z objektom smo med seboj primerjali različne togosti in pristope vodenja. 3) Ocenili smo natančnost sledenja diskretnih CMP-jev posplošenih glede na eno- in dvo-dimenzijske iskalne točke. 4) Posplošene periodične CMP-je smo uporabili za vrtenje ročaja in s tem spremenili višino nastavljivo mize. Za takšno nalogo bi bila pridobitev dinamičnega modela težavna in zamudna.

The talk will address the problem of achieving high robot compliance while maintaining low tracking errors without using explicit mathematical models of task dynamics. The proposed approach uses programing by demonstration to learn new task related compliant movement. The presented Compliant Movement Primitives (CMPs) are a combination of position trajectories, gained through human demonstration and encoded as Dynamic Movement Primitives, and corresponding torque trajectories encoded as a linear combination of radial basis functions. A set of example Compliant Movement Primitives is used with statistical generalization in order to execute previously unexplored tasks inside the training space. While the talk will present CMPs, the main focus of the talk will be on evaluation, which can be divided into four parts: 1) Compliant movement evaluation was done at various stiffness settings by observing tracking errors while executing a simple pick and place task using different control approaches, 2) Collision setup was used to compare robot behavior to unforeseen impacts while using different stiffness settings and control strategies, 3) Discrete CMPs and their generalization over one and two dimensional queries was evaluated w.r.t. tracking errors on a pick and place scenario, and 4) Periodic CMPs and generalized CMPs were evaluated using a hard-to-model task of raising a height-adjustable table by rotating its handle.

17.04.2015: Robert Bevec
Avtonomno zaznavanje in raziskovanje objektov z uporabo fovealnega vida / Autonomous Object Discovery and Exploration Integrated with Foveated Vision
Autonomous learning of visual appearance of unknown objects is a demanding task. We propose a method that integrates foveated vision on a humanoid robot with autonomous object discovery and explorative manipulation actions such as pushing, grasping, and in-hand rotation. The humanoid robot starts by searching for objects in a visual scene and generating hypotheses about which parts of the visual scene could constitute an object. The hypothetical objects are verified by applying pushing actions, where the existence of an object is considered confirmed if the visual features exhibit rigid body motion. In our previous work we showed that partial object models can be learnt by a sequential application of several robot pushes, which generates the views of object appearance from different viewpoints. However, with this approach it is not possible to guarantee that the object will be seen from all relevant viewpoints even after a large number of pushes have been carried out. Instead, we show that confirmed object hypotheses contain enough information to enable grasping and that object models can be acquired more effectively by sequentially rotating the object. We show the effectiveness of our new system by comparing object recognition results after the robot learns object models by two different approaches: 1. learning from images acquired by several pushes and 2. learning from images acquired by an initial push followed by several grasp-rotate-release action cycles.

Avtonomno učenje predstavitev neznanih predmetov je zahtevna naloga. Predlagamo metodo, ki poveže fovealni vid na robotski glavi z avtonomnim odkrivanjem objektov in raziskovalnimi potiski kot so potiskanje, prijemanje in vrtenje v roki. Humanoidni robot prične z odkrivanjem objektov, tako da ustvari hipoteze o obstoju objektov na podlagi vizualnih podatkov. Hipotezo preveri s potiskanjem potencialnega objekta in v primeru zaznanega togega premika potrdi njegov obstoj. V našem prejšnjem delu smo pokazali, da se lahko robot z zaporednimi potiski nauči predstavitve novih predmetov z različnih strani. S takih pristopom ni moč zagotoviti, da bo robot videl predmet z vseh relevantnih strani tudi po velikem številu potiskov. Z našo novo metodo pokažemo, da vsebuje potrjen predmet po prvem potisku dovolj informacij, da ga lahko robot poskusi prijeti. Robot se lahko nato bolj učinkovito nauči predstavitve predmeta z vseh strani, tako da ga v zaporednih korakih vrti. Učinkovitost predlagane metode prikažemo s primerjavo uspešnosti prepoznavanja predmetov po tem ko se robot nauči predstavitve z dvema različnima pristopoma: 1. učenje predstavitev iz slik po zaporednih potiskih in 2. učenje predstavitev iz slik zajetih po začetnem potisku in zaporednih nizih prijema, vrtenja in izpusta predmeta.

03.04.2015: Andrej Gams
Spreminjanje osnovnih gibov s povratno informacijo o sili: končni, periodični in kombinirani gibi / Modulating Dynamic Motion Primitives with Force Feedback Coupling Terms: Discrete, Periodic and Combined Discrete-Periodic Motions
Okvir dinamičnih generatorjev gibov , DMPjev, (včasih jim rečemo tudi osnovni gibi) vključuje lastnosti, ki omogočajo enostavno izvajanje robotskih trajektorij, npr. posredno odvisnost od časa, odziv na motnje in možnost spreminjanja trajektorij. Predlagali smo metodo prilagajanja končnih DMPjev na nivoju hitrosti z uporabo sklopitvenega člena na osnovi povratnozančne informacije o sili, s čimer smo se približali dejanskemu dinamičnemu obnašanju. Z učenjem sklopitvenega člena v naprej se naučimo direktne veje krmilne zanke, s čimer zmanjšamo ali izničimo potrebo po sprotnih popravah s povratnozančno informacijo. Na seminarju bom predstavil osnove tega pristopa za končne in periodične gibe, poudarek pa bo na kombiniranem periodičnem gibu, ki vljučuje začetni gib. Primer takega giba je hoja, kjer je prvi korak drugačen kot kasnejši. Tak kombiniran gib zahteva spremembe v kanoničnem sistemu DMPa ter v učenju informacije za direktno vejo krmilne sheme.

The framework of dynamic movement primitives (DMPs) contains many favorable properties for the execution of robotic trajectories, such as indirect dependence on time, response to perturbations, and the ability to easily modulate the given trajectories. We proposed a method of modulating discrete DMPs with coupling terms at the velocity level using force feedback, bridging the gap between kinematic and dynamic behavior. By learning the coupling term in advance, we essentially learned a feedforward component of the control loop, reducing on-line corrections using feedback. In the seminar I will present the basics of this approach for discrete and periodic tasks, with an emphasis on a combined discrete-periodic task. An example of such a task is walking, where the initial motion is not the same as subsequent periodic behavior. The combination of the discrete and the periodic behavior requires adaptations of both the DMP canonical system as well as of the learning of the feedforward components.

30.03.2015: Matjaž Omladič
Kako naj v tehniki razumemo pojem verjetnosti / How to understand the notion of probability in engineering
V predavanju bomo izhajali iz nekaterih osnovnih statistično verjetnostnih pojmov, ki se uporabljajo npr pri vzpodbujenem učenju v robotiki. Primer takega pojma je Kullback-Laiblerjeva divergenca, ki meri neke vrste oddaljenost med dvema verjetnostnima porazdelitvama (distribucijama). Ko minimiziramo to divergenco s tem obe porazdelitvi približujemo druga drugi. Pri tem pa računsko nikakor ni vseeno ali sta porazdelitvi diskretni ali zvezni, čeprav gre idejno za zelo soroden račun. V predavanju bomo predstavili pogled na verjetnost, ki nam olajša razumevanje teh pojmov. Pogled bomo pojasnjevali z zgledi.

Our talk is motivated by some notions in probability and statistics that have applications in reinforcement learning in robotics. An example of a notion like that is Kullback-Laibler divergence which is a measure of distance between two distributions. When we minimize this divergence we are coming closer to a match between the two distributions. There is a notable difference in computations between the case of two distributions being discrete and the case of continuous distributions. However the two computations are in principle quite similar. Our goal is to present a view on probability that facilitates the understanding of these concepts. We will give explaining examples.

06.03.2015: Aljaž Kramberger
Integracija in načrtovanje algoritma planiranja za namene robotskega sestavljanja različnih predmetov / Integration and development of planning algorithms for robot assembly of differently shaped parts
Sledeč seminar bo zavzemal dva sklopa. V prvem delu bo govora o robotskem sestavljanju različno oblikovanih delov v povezavi z algoritmom za planiranje akcij. Operacije sestavljanja objektov, so zelo uporaben proces v industriji in humanoidni robotiki. Pri kompleksnejših nalogah je potrebne nadzorni sistem, ki podaja informacijo kaj in kako je potrebno sestaviti. Program mora upoštevati vse možne akcije in vmes prilagoditi zaporedja sestavljanja, če pride do motenj med delovanjem robotskega sistema. V ta namen smo razvili komunikacijsko identifikacijski sistem med robotom in planerjem. Ta sistem identificira napake med izvajanjem zastavljene naloge robotskega mehanizma, ki bazirajo pretežno na zaznavanju sile. Če je naloga neuspešno izvedena, planer ponovno prilagodi zaporedje sestavljanja. V drugem delu seminarja bo govora o ROS – Matlab vmesniku. Predstavljen bo princip delovanja robotskega okolja ROS (Robot Operating System) in njegova struktura. Poudarek bo predvsem na integraciji Matlabovega okolja v ta sistem. Prikazani bodo primeri postavitve okolja in uporaba le tega.

The following seminar is divided in two parts. In the first part I will talk about robot assembly strategies in connection with a high level algorithm for planning actions. The problem of assembling differently shaped parts is well known in industrial and humanoid applications. For more complex tasks a planning system is needed, to provide the necessary information on how and what you need to assemble. The program must take into account all the possible actions and be capable to alter the sequence of assembly, if interference occurs during the robot execution of the planed task. To tackle this problem we developed a communication – identification system between the two separate systems. The identification algorithm identifies the errors that occur during the robot execution. The identification is don mainly according to the force perception of the robot. If the task is unsuccessful the planer has to readjust the assembly sequence. In the second part of the seminar the ROS – Matlab interface will be presented. A quick overview of the ROS (Robot Operating System) will be given. The main part will be the integration of the Matlab system in to the ROS environment. At the end some examples will be presented.

06.03.2015: Barry Ridge
Developing Software Tools for Action Data Tables in the ACAT Project


For the ACAT project, in order to effectively work with the core data structures defined by the project, those being Action Data Tables (ADTs), we found it both necessary and useful to build certain software tools with which to manipulate them. In this talk we outline the results of that development process so far, namely a command-line tool for the generation and editing of ADT XML files using ROS bag recordings of robot task interactions and a GUI editor that acts as an interface to the command-line tool. The GUI editor provides a means of both visualising the contents of a ROS bag recording and of annotating the contents in various ways, such that the annotations may be passed on to the command-line tool using an intuitive user interface.

23.01.2015: Jernej Čamernik
Vloga roke pri ohranjanju stalno motenega ravnotežja / Role of Hand Contact in Continually Challenged Postural Equilibrium


Humans dynamically control balance while standing by using various strategies that depend on individual, environmental and task-specific conditions. Leaning or holding to a support surface is one of these strategies. Contemporary studies include the additional hand contact in balancing situations but they are mostly describing reactive compensatory movements. It has been shown that reactive arm movements always occur in conditions where sudden a perturbation is applied and that they even occur in situations when compensatory movements are not necessary. Such unconditioned reactions mask the identification of motor control mechanisms that govern the motion of the arm. To overcome this limitation we developed a novel experimental paradigm where the balance of the subjects was continuously perturbed in different stability conditions by mild random perturbations that do not provoke reactive arm movements. We measured ten healthy male subjects who were standing on a force plate while being perturbed by a mild, random and continuous anteroposterior (AP) pulls at the waist. We measured the EMG activity of two lower leg, two trunk and two upper arm muscles. The AP shifts of subject's center of pressure (CoP) were measured using a force plate. All EMG and force plate data were analysed in a frequency domain and subjected to statistical analysis. Results show that the contribution of lower leg muscles was significantly reduced when the subjects were able to use the additional hand support. On the other hand, the activation of trunk flexors was higher when the subjects used the additional hand support and the activation of the trunk extensors remained unchanged. The CoP displacement in AP direction significantly decreased through the sessions regardless of the additional hand support which indicates motor adaptation in both conditions. Equivalently, the activity of most muscles gradually dropped through the sessions, where the biggest drop was just after the first session. The exception was the trunk flexor muscle, where the activity was substantially higher in the handle condition. The lowered activity of leg muscles in handle conditions suggests that the postural equilibrium was dominantly obtained by the use of hand muscles. The adaptation of muscle activity to the perturbation is immediate and tends to continue through the sessions which indicates possible optimization of muscle utilization. The CoP(AP) shifts were reduced in both conditions but remained more constant through sessions in the handle condition.

23.01.2015: Tadej Debevec
Hipoksija: Z zemlje v vesolje in nazaj / Hypoxia: From Earth to Space and back
Spremembe delnega tlaka kisika imajo lahko tako pozitivne, kot tudi negativne učinke na človeški organizem. Predavanje je namenjeno predstavitvi mojega dosedanjega raziskovalnega dela na področju učinkov hipoksije na športno sposobnost in zdravje v različnih okoliščinah. V prvem delu bom predstavil naše začetne raziskave na področju uporabe različnih hipoksičnih vadbenih modelov za izboljšanje športne sposobnosti in hematološke adaptacije. V drugem delu predavanja bo povdarek na preučevanju kombiniranih in ločenih učinkov hipoksije in (simulirane) mikrogravitacije, kot jih pričakujemo med bodočim bivanjem v vesolju, na prehranski status in telesno sestavo. Zaključni del bo sestavljen iz predstavitve nedavnih raziskav na področju kliničnih vidikov hipoksije, ki kažejo zanimive učinke hipoksije med daljšim obdobjem neaktivnosti na kontrolo glukoze in funkcioniranje metabolnega sistema.

Changes in partial pressure of inspired oxygen can have both, detrimental and beneficial effects on humans. In this talk I will outline my up-to-date research pertaining to effects of hypoxia on metabolic health and its utilization for performance enhancement. First I will present our early work focusing on diverse short intermittent hypoxic models aiming at augmenting athletic performance and hematological adaptations. The “space” part of the talk will focus on independent and additive effects of hypoxia and (simulated) microgravity, as envisaged during future space explorations, on nutritional status and body composition. Finally, back to Earth, I will present some very recent data showing interesting effects of hypoxia during periods of prolonged inactivity on glucose control and metabolic functioning

09.01.2015: Luka Peternel
Pristopi učenja robotov s človekom v kontrolni zanki za izvedbo periodičnih nalog / Human-in-the-loop robot learning approaches for periodic task execution
V seminarju bom predstavil metode učenja robotov s človekom v kontrolni zanki. Najprej bom predstavil metodo za vodenje eksoskeletov pri periodičnih nalogah. Osnovni princip te metode je adaptacija naučenih obnašanj navorov v sklepih robota na način, da se minimizira človekova mišična aktivnost. Nazadnje smo izvedli eksperimente na večjih subjektih, kjer nas je zanimala ocena metode glede na natančnost gibanja in hitrost adaptacije. Rezultati so pokazali, da je natančnost z uporabo metode primerljiva z natančnostjo, ki jo človek doseže brez metode. Poleg tega smo razvili vmesnik za učenja robotov s človekom v kontrolni zanki pri izvajanju naloge sestavljanja objektov. V tem pristopu človeku omogočimo spreminjanje robotove impedance v realnem času preko vmesnika, ki ga drži v roki

In the seminar I will present human-in-the-loop robot teaching methods that we recently developed. During my visit to ATR CNS labs in Kyoto Japan we developed a method for exoskeleton control in periodic tasks. The basic principle of the method is to adapt the learnt joint torque behaviour in a way to minimise the human muscle activity. In the recent visit, we performed additional experiments on multiple subjects to evaluate the methods performance in terms of motion accuracy and adaptation rate. The results showed that the accuracy using the method is comparable to the human motion without the method. In addition, we developed a novel interface for human-in-the-loop teaching of robot assembly tasks. In this approach the human teleoperator can control the robot impedance in real-time using hand-held interface.