Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revisionPrevious revision
Next revision		Previous revision
Next revisionBoth sides next revision
jobs [2014/05/13 09:29] – tenorthjobs [2015/09/08 12:24] – [Theses and Jobs] nyga
Line 5: Line 5:
  
  
-== GPU-based Parallelization of Numerical Optimization Techniques (BA/MA/HiWi)==+== Kitchen Activity Games in a Realistic Robotic Simulator (BA/MA/HiWi)== 
 + {{ :research:gz_env1.png?200|}} 
  
-In the field of Machine Learning, numerical optimization techniques play a focal role. However, as models grow larger, traditional implementations on single-core CPUs suffer from sequential execution causing a severe slow-down. In this thesis, state-of-the-art GPU frameworks (e.g. CUDA) are to be investigated in order implement numerical optimizers that substantially profit from parallel execution.+Developing new activities and improving the current simulation framework done under the [[http://gazebosim.org/|Gazebo]] robotic simulator. Creating a custom GUI for the game, in order to launch new scenarios, save logs etc.
  
 Requirements: Requirements:
-  * Skills in numerical optimization algorithms +  * Good programming skills in C/C++ 
-  * Good programming skills in Python and C/C+++  * Basic physics/rendering engine knowledge 
 +  * Gazebo simulator basic tutorials
  
-Contact: [[team:daniel_nyga|Daniel Nyga]]+Contact: [[team:andrei_haidu|Andrei Haidu]]
  
-== Online Learning of Markov Logic Networks for Natural-Language Understanding (MA)==+== Integrating Eye Tracking in the Kitchen Activity Games (BA/MA)== 
 + {{ :research:eye_tracker.png?200|}} 
  
-Markov Logic Networks (MLNs) combine the expressive power of first-order logic and probabilistic graphical models. In the past, they have been successfully applied to the problem of semantically interpreting and completing natural-language instructions from the web. State-of-the-art learning techniques mostly operate in batch mode, i.e. all training instances need to be known in the beginning of the learning process. In context of this thesis, online learning methods for MLNs are to be investigated, which allow incremental learning, when new examples come in one-by-one.+Integrating the eye tracker in the [[http://gazebosim.org/|Gazebo]] based Kitchen Activity Games framework and logging the gaze of the user during the gameplay. From the information typical activities should be inferred.
  
 Requirements: Requirements:
-  * Experience in Machine Learning. +  * Good programming skills in C/C++ 
-  * Experience with statistical relational learning (e.g. MLNs) is helpful. +  * Gazebo simulator basic tutorials
-  * Good programming skills in Python.+
  
-Contact: [[team:daniel_nyga|Daniel Nyga]]+Contact: [[team:andrei_haidu|Andrei Haidu]]
  
 +== Hand Skeleton Tracking Using Two Leap Motion Devices (BA/MA)==
 + {{ :research:leap_motion.jpg?200|}} 
  
-==HiWi-Position: Knowledge Representation & Language Understanding for Intelligent Robots==+Improving the skeletal tracking offered by the [[https://developer.leapmotion.com/|Leap Motion SDK]], by using two devices (one tracking vertically the other horizontally) and switching between them to the one that has the best current view of the hand.
  
-In the context of the European research project RoboHow.Cog [1,2] we +The tracked hand can then be used as input for the Kitchen Activity Games framework.
-are investigating methods for combining multimodal sources of knowledge (e.g. video, natural-language recipes or computer games), in order to enable mobile robots to autonomously acquire new high level skills like cooking meals or straightening up rooms.  +
  
-The Institute for Artificial Intelligence is hiring a student researcher for the +Requirements: 
-development and the integration of probabilistic methods in AI, which enable intelligent robots to understand, interpret and execute natural-language instructions from recipes from the World Wide Web.+  * Good programming skills in C/C++
  
-This HiWi-Position can serve as a starting point for future Bachelor's, Master's or Diploma Theses.+Contact: [[team:andrei_haidu|Andrei Haidu]]
  
-Tasks: +== Fluid Simulation in Gazebo (BA/MA)== 
-  * Implementation of an interface to the Robot Operating System (ROS). + {{ :research:fluid.png?200|}}  
-  * Linkage of the knowledge base to the executive of the robot. + 
-  * Support for the scientific staff in extending and integrating components onto the robot platform PR2.+[[http://gazebosim.org/|Gazebo]] currently only supports rigid body physics engines (ODE, Bullet etc.), however in some cases fluids are preferred in order to simulate as realistically as possible the given environment. 
 + 
 +Currently there is an [[http://gazebosim.org/tutorials?tut=fluids&cat=physics|experimental version]] of fluids  in Gazebo, using the [[http://onezero.ca/fluidix/|Fluidix]] library to run the fluids computation on the GPU. 
 + 
 +The computational method for the fluid simulation is SPH (Smoothed-particle Dynamics), however newer and better methods based on SPH are currently present 
 +and should be implemented (PCISPH/IISPH). 
 + 
 +The interaction between the fluid and the rigid objects is a naive one, the forces and torques are applied only from the particle collisions (not taking into account pressure and other forces). 
 + 
 +Another topic would be the visualization of the fluid, currently is done by rendering every particle. For the rendering engine [[http://www.ogre3d.org/|OGRE]] is used. 
 + 
 +Here is a [[https://vimeo.com/104629835|video]] example of the current state of the fluid in Gazebo. 
  
 Requirements: Requirements:
-  * Studies in Computer Science (Bachelor's, Master's or Diploma) +  * Good programming skills in C/C++ 
-  * Basic skills in Artificial Intelligence +  * Interest in Fluid simulation 
-  * Optional: basic skills in Probability Theory +  * Basic physics/rendering engine knowledge 
-  * Optional: basic skills in Machine Learning +  * Gazebo simulator and Fluidix basic tutorials
-  * Good programming skills in Python and Java+
  
-Hours: 10-20 h/week+Contact: [[team:andrei_haidu|Andrei Haidu]]
  
-Contact: [[team:daniel_nyga|Daniel Nyga]] 
  
-[1] www.robohow.eu\\ +== Automated sensor calibration toolkit (BA/MA)==
-[2] http://www.youtube.com/watch?v=0eIryyzlRwA+
  
 +Computer vision is an important part of autonomous robots. For robots the image sensors are the main source of information of the surrounding world. Each camera is different, even if they are from the same production line. For computer vision, especially for robots manipulating their environment, it is important that the parameters for the cameras in use are well known. The calibration of a camera is a time consuming task, and the result depends highly on the chosen setup and the accuracy of the operator.
  
-== Depth-Adaptive Superpixels (BA/HiWi)==+The topic for this thesis is to develop an automated system for calibrating cameras, especially RGB-D cameras like the Kinect v2.
  
-We are currently investigating a new set of sensors (RGB-D-T), which is a combination of a kinect with a thermal image camera. Within this project we want to enhance the Depth-Adaptive Superpixels (DASP) to make use of the thermal sensor data. Depth-Adaptive Superpixels oversegment an image taking into account the depth value of each pixel. {{ :research:dt_dasp.png?200|}}+ {{ :kinect2_calibration_setup_small.jpg?200|}} 
 +The system should: 
 +  * be independent of the camera type 
 +  * estimate intrinsic and extrinsic parameters 
 +  * calibrate depth images (case of RGB-D) 
 +  * integrate capabilities from Halcon [1] 
 +  * operate autonomously
  
-Since the current implementation of DASP is not very performant for high resolution images, there are several possibilities options for doing a project in this field like reimplementing DASP using the CUDA, investigating how thermal data can be integrated, ...+Requirements:  
 +  * Good programming skills in Python and C/C++ 
 +  * ROS, OpenCV
  
-Requirements: +[1] http://www.halcon.de/
-  * Basic knowledge of image processing +
-  * Good programming skills in C/C++. +
-  * Experience with CUDA is helpful+
  
-Contact: [[team:jan-hendrik_worch|Jan-Hendrik Worch]]+Contact: [[team:alexis_maldonado|Alexis Maldonado]] and [[team:thiemo_wiedemeyer|Thiemo Wiedemeyer]]
  
 +== On-the-fly 3D CAD model creation (MA)==
  
-== Tools for knowledge acquisition from the Web (BA/MA/HiWi) ==+Create models during runtime for unknown textured objets based on depth and color information. Track the object and update the model with more detailed information, completing it's 3D model from multiple views improving redetection. Using the robots manipulator pick up the object and complete the model by viewing it from multiple viewpoints.
  
-There are several options for doing a project related to the acquisition of +Requirements:  
-knowledge from Web sources like online shops, repositories of object models, +  * Good programming skills in C/C++ 
-recipe databases, etc.+  * strong background in computer vision  
 +  * ROS, OpenCV, PCL
  
-Requirements: +Contact: [[team:thiemo_wiedemeyer|Thiemo Wiedemeyer]] 
-  * Programing skills (Java) + 
-  * Experience with Web languages and datamining techniques is helpful +== Simulation of a robots belief state to support perception(MA) == 
-  * Depending on the focus of the project, experience with database technology, natural-language processing or computer vision may be helpful+ 
 +Create a simulation environment that represents the robots current belief state and can be updated frequently. Use off-screen rendering to investigate the affordances these objects possess, in order to support segmentation, detection and tracking of these in the real world.  
 + 
 +Requirements:  
 +  * Good programming skills in C/C++ 
 +  * strong background in computer vision  
 +  * Gazebo, OpenCV, PCL 
 + 
 +Contact: [[team:ferenc_balint-benczedi|Ferenc Balint-Benczedi]] 
 + 
 +== Multi-expert segmentation of cluttered and occluded scenes == 
 + 
 +Objects in a human environment are usually found in challenging scenes. They can be stacked upon eachother, touching or occluding, can be found in drawers, cupboards, refrigerators and so on. A personal robot assistant in order to execute a task, needs to detect these objects and recognize them. In this thesis a multi-modal approach to interpreting cluttered scenes is going to be investigated, combining the results of multiple segmentation algorithms in order to come up with more reliable object hypotheses.
  
-Contact: [[team:moritz_tenorth|Moritz Tenorth]]+Requirements:  
 +  * Good programming skills in C/C++ 
 +  * strong background in 3D vision  
 +  * basic knowledge of ROS, OpenCV, PCL
  
 +Contact: [[team:ferenc_balint-benczedi|Ferenc Balint-Benczedi]]
- Impressum - Datenschutz



Prof. Dr. hc. Michael Beetz PhD
Head of Institute

Contact via
Andrea Cowley
assistant to Prof. Beetz
ai-office@cs.uni-bremen.de

Discover our VRB for innovative and interactive research


Memberships and associations:


Social Media: