We are developing an AI agent capable of solving such physical reasoning tasks. If you would like to know more about this thesis/project opportunity, please contact:
Dr. Andrew Melnik <andrew.melnik(at)uni-bielefeld.de>
This thesis / project will focus on practical applications of state of the art style generative adversarial network models (StyleGAN) [see References here].
If you would like to know more about this thesis / project opportunity, please contact: Dr. Andrew Melnik <andrew.melnik(at)uni-bielefeld.de>
This challenge tackles a key problem in the transportation world: How to efficiently manage dense traffic on complex railway networks? This is a real-world problem faced by many transportation and logistics companies around the world such as the Swiss Federal Railways and Deutsche Bahn. We are developing an AI agent capable of solving such problem.
If you would like to know more about this thesis/project opportunity, please contact:
Dr. Andrew Melnik <andrew.melnik(at)uni-bielefeld.de>
This thesis / project will focus on practical applications of state of the art in computer vision[1][2] to detect objects and segment images [3][4][5][6].
In this project / thesis, you are tasked with developing a controller which enables a physiologically-based human model to navigate a complex obstacle course as quickly as possible [1][2][3][4]. You are provided with a human musculoskeletal model and a physics-based simulation environment where you can synthesize physically and physiologically accurate motions.
In this project / thesis, you will explore recent advances in deep learning for manipulating objects with robotic hands in simulated environments [1][2]. You will try out solutions that cleverly combine deep reinforcement learning, supervised learning, and engineering.
In this thesis / project, you will have the opportunity to further develop an Artificial Intelligenceagent capable of performing complex purposeful actions in MinecraftorDoomenvironments.
If you would like to know more about this thesis / project opportunity, please watch this video and contact:
Motivation is the key to successful learning, in particular practicing to play a music hit has a potential to increase motivation and by doing so to accelerate learning.
The goal of this project is to implement a NN music hit generator and integrate it into existing intelligent tutoring software.
Playing a musical instrument well is a skill that many people are pursuing over years. What makes it particularly difficult is the absence of a teacher to provide feedback and encouragement. While practicing, we commonly neither choose the best practice strategies, nor are we capable of self-analyzing our performance.
Recommender systems are being ubiquitously used to optimally couple users and products. The goal of this project is to employ standard methods used to produce recommendations [1,4]
Finding an optimal learning strategy not only for machines but also for humans is central to a large research effort dedicated to development of ITS. Creating a computational environment that enables human learners to gain full control of their learning efficiency is
Virtual try on of clothes with deep learning and image segmentation using just a 2d image. To create a shopping environment close to reality, virtual try-on technology has attracted a lot of interests recently by delivering product information similar to that obtained from direct product examination. It allows users to experience themselves wearing different clothes without efforts of changing them physically.
The goal of this project is to develop a modelling approach to enable a human to learn a complex motor skill quickly. Learning of motor skills such as piano playing involves mastering of multiple subskills, such as reading music score, simultaneously controlling fingers and hands.
We are developing an AI agent for drone control capable of solving SLAM-related problems [1] in environments with challenging features such as changing light conditions, low illumination, adverse weather, and dynamic objects. If you would like to know more about this thesis/project opportunity, please contact: Dr. Andrew Melnik <andrew.melnik(at)uni-bielefeld.de>
Under the framework of Deutsche Forschungsgemeinschaft (DFG) project—DEXMAN (https://gepris.dfg.de/gepris/projekt/410916101?language=en) , we are offering a master thesis topic “robotic dexterous manipulation learning by demonstration”.
Learning of a new skill is a process that is commonly accompanied by an expert. Inspired by the MAML [1] the goal of this thesis is to implement a meta-learner to accelerate human learning. We will develop a tool that teaches a novice to play piano, guided by the dexmo exoskeleton.
We have developed iObject, an intelligent object, that can measure its pose in space as well as the force/pressure profiles of human or robot hands grasping it. Using this object, several theses topics become possible:
Sense of touch, or haptic perception is the most robust and ubiquitous human sense. The goal of this project is to endow a robot with a simple and efficient method to haptically interact with its 3D environment. To this end, the project will employ both the data recorded from a human performing a task haptically, and robots' own haptic control primitives.
Vision-Based robot control typically requires a large amount of manual tuning and feature selection (edge, color, SIFT features, etc.). Modern (GPU-driven) computing power and more robust numerical methods, nowadays allow for training of deep neural networks that extract suitable features for a given task automatically and that are able to learn control tasks in an end-to-end fashion, i.e.
The goal of this thesis is to combine a frame-based scene segmentation approach (based on depth and color) and a tracking-based approach (based on coherent motion of SIFT features) to yield more robust tracking results.
Es soll für ein robotisches Greifsystem eine Anbindung an ein Brain-Computer Interface erstellt werden, die, basierend auf der jeweiligen Szene, verschiedene Auswahlmöglichkeiten anbietet.
Im Rahmen eines EU-Projektes ist ein Smartphone-basiertes System zur Unterstützung der Diagnose und Therapie von Patienten mit bipolaren Störungen (manisch-depressive) entwickelt worden.
Bei der Messung von EEG Daten, d.h. elektrophysiologischer Daten des Gehirns, versucht man jegliche Bewegung des Nutzers zu vermeiden, da die Muskelaktivitaet das eigentliche kortikale Signal ueberlagert. Zwar gibt es algorithmische Verfahren, um diese s.g.
In vielen Untersuchungsszenarien werden Augen- und Körperbewegungen von Probanden simultan erfaßt, oft liegt jedoch kein gemeinsames raum-zeitliches Bezugssystem zugrunde. Ziele der Arbeit sind die Überführung der verschiedenen Daten in ein solches System und die integrative Datenvisualisierung.
Komplexe Gelenkwinkeltrajektorien (z.B. Tennis-Schlag) erfordern eine Repräsentation der Trajektorie, die über die Interpolation zwischen ein paar Stützstellen hinaus geht, und zudem adaptiv auf neue Ziele (z.B.
Die Entwicklung von Brain-Computer Interfaces (Abk. BCI, Systeme, bei denen eine Maschine, z.B. ein Computerspiel oder ein Roboter mit "Gedanken" gesteuert wird) hat in den letzten Jahren grosse Fortschritte gemacht.
Sampling-basierte Planungsverfahren werden eingesetzt, um die Bewegungen komplexer Roboter in komplizierten Umgebungen mit vielen Hindernissen zu planen. Seit der Entwicklung dieser Technik von Kavraki (1996) und LaValle (1998) sind eine Vielzahl von Varianten entstanden.
Wenn Roboter in einem menschlichen Umfeld agieren, müssen sie in der Lage sein, auf plötzliche Änderungen der Umgebung zu reagieren. So können Hindernisse hinzukommen oder verschwinden oder sich auch kontinuierlich über die Zeit bewegen.
Für zwei Kuka Leichtbauroboter, die jeweils über 7 Freiheitsgrade verfügen, sollen in Simulation verschiedene Anordnungen/Abstände miteinander verglichen werden. Ziel ist es, den gemeinsamen (tatsächlich auch erreichbaren) Arbeitsraum der beiden Arme zu maximieren bei gleichzeitiger Minimierung der Kollisionsgefahr der Arme miteinander.
Datenmodelle dienen im wesentlichen dem Ziel, die zugrunde liegende Strukturen und Abhängigkeiten in beobachteten Daten zu erkennen und darzustellen. Eine ziel-orientierte manuelle Bewegung (z.B. Hand nähert sich einem Objekt, Hand weicht zurück, Hand greift zu usw.) kann durch eine mehrdimensionale Zeitreihe von Beobachtungen dargestellt werden.
Menschen besitzten die Fähigkeit, einfache und komplexe zielorientierte Bewegungen eines Gegenüber zu unterscheiden und zu klassifizieren (z.B. Hand nähert sich an, Hand weicht zurück, Hand greift zu usw.). Die Erkennung dieser Interaktionsmuster ist ein vielschichtiger Vorgang im menschlichen Gehirn, der zum grossen Teil noch unerforscht ist.
Finding an optimal learning strategy not only for machines but also
The goal of this project is to develop a modelling approach to enable a human to learn a complex motor skill quickly. Learning of motor skills such as piano playing involves mastering of multiple subskills, such as reading music score, simultaneously controlling fingers and hands.
Learning of a new skill is a process that is commonly accompanied by an expert. Inspired by the MAML [1] the goal of this thesis is to implement a meta-learner to accelerate human learning. We will develop a tool that teaches a novice to play piano, guided by the dexmo exoskeleton.