Future computer programs will contain a growing part of 'intelligent' software modules that are not conventionally programmed, but that are learned either from data provided by the user or from data that the program autonomously collects during its use.
In this spirit, the Machine Learning Lab deals with research on Machine Learning techniques and the integration of learning modules into larger software systems, aiming at their effective application in complex real-world problems. Application areas are robotics, control, forecasting and disposition systems, scheduling and related fields.
Research Areas: Efficient Reinforcement Learning Algorithms, Intelligent Robot Control Architectures, Learning in Multiagent Systems, (Un-)Supervised Learning, Autonomous Robots, Industrial Applications
Combining EEG and Reinforcement Learning for Robot Control 18.4.2013
As a demonstration of the idea behind the NeuroBots project we developed a prototype system to control an autonomous robotic arm's actions using an EEG device. The user merely selects an object, which the arm then grasps by itself using behaviours acquired through reinforcement learning. A video of the system is available in the media section.
BrainLinks-BrainTools Cluster of Excellence Launched 18.4.2013
Last week, the BrainLinks-BrainTools Cluster of Excellence here at the University of Freiburg was officially launched. It aims to push research on brain-machine interfaces for disease prevention and rehabilitation forward. The Machine Learning Lab will be participating in the context of three different project dealing with the development of intelligent prostheses, detection of epileptic seizures, and reinforcement learning in real neural networks. Our project listing has been updated with more detailed descriptions and links.
System Design Project Competition 11.2.2013
The annual final competition of the System Design Project will take place this Friday, February 15. First year students will compete with LEGO robots developed over the past semester to finish an unknown course as quickly as possible. This year, a livestream will be available as well, which you can watch here.
Software Releases and Updates 15.1.2013
We are happy to announce our brand new download section. Aside from the long-anticipated fundamental upgrade to the CLS2 machine learning framework, several new open-source projects have been added: Tapir, a complementary modular object detection tool; libgp, a stand-alone Gaussian Process library; and motld, a C++ implementation of the TLD object detection algorithm with multi-object support. Find these projects in the Downloads section to the left.
Visual Deep Learning 9.8.2012
For the challenging pole balancing task we developed a system which uses raw visual input data to learn a control strategy. Using a neural network - a deep autoencoder - we were able to compress the relevant information in the camera images to a low dimensional feature space which was used as system states for reinforcement learning. (see Video)
Open Source RoboCup Mid-Size Robot 20.4.2012
We are proud to introduce the new Website for openTribot, our Open Source RoboCup Mid-Size Robotic Platform. The DFG Transfer Projects Goal was the creation of a robust, easy to use, performant Robot that can be used in RoboCup Mid-Size Soccer. Our Prototype Robots are complete and can be programmed with the ROS Robotic Operating System. Project Website
We congratulate Sascha and Patrick to their recently awarded “EXIST-Gründerstipendium”. This grant will allow them to work full-time on their idea of founding a new, highly-innovative games company here in Freiburg (see http://www.5dlab.com). Their next product will be a location-aware strategy-MMO.
Learning to drive a virtual racing car - this idea was realized in the bachelor thesis written by Thorsten Engesser. A neural network based driver learns to drive fast but save, controlling the car by steering, breaking and accelerating. The learning car is realized in the Torcs-framework, showing highly competitive driving behaviour, finally learned from the experience of success and failure (see video). Our data-efficient batch-RL scheme NFQ was used to realize the learning. More learning videos can be found in this list.