NEURAL NETWORKS IN MACHINE LEARNING
We have been working on neural networks since 1990. Our neural systems have been used for computer vision, behaviour control, and also for learning system in robotics. We are currently developing novel applications of deep learning algorithms.
AUTONOMOS CARS FROM BERLIN
AutoNOMOS is an autonomous vehicle that has been designed and built by the Artificial Intelligence Group at Freie Universität Berlin. In this car the driver is the computer. The vehicles are equipped with sensors, computers, and actuators. The sensors collect information about the immediate environment.
The FUmanoids team is a student project supervised by Prof. Raúl Rojas. The FUmanoids are soccer playing robots. They participate every year in the RoboCup competition.
AUTONOMOUS AND BRAIN-CONTROLLED WHEELCHAIRS
At FU-Berlin we have developed a wheelchair that drives autonomously indoors and outdoors. The wheelchair can also be controlled using brain signals.
AUTONOMOUS MODEL CARS
We participate in Carolo-Cup, a yearly race of automonous model cars organized by the Technische Universität Braunschweig.
ANALYSIS OF THE HONEYBEE DANCE COMMUNICATION SYSTEM
Honeybees communicate valuable field resources to nestmates by performing the so called waggle dance. It encodes the polar coordinates of the food patch and serves as a positive feedback mechanism allocating and directing more foragers to newly found sites. However, it ist still unknown exactly how the information is decoded by the follower bees. We investigate this intriguing communication system with a honeybee robot that can mimic the dance and recruit bees to arbitrary field sites.
BIOMIMETIC FLYING PLATFORM FOR THE INVESTIGATION OF HONEYBEE NAVIGATION
NeuroCopter is a joint project of the Biorobotics Lab with the Computational Systems Neuroscience (Prof. Nawrot in Cologne) and Neurobiology (Prof. Menzel in Berlin). We investigate honeybee navigation through electrophysiology and behavioral experiments using quadcopter (with Prof. Menzel) and modeling neural circuits (with Prof. Nawrot) that might be involved in representing relevant environmental structures.
ANALYSIS OF SOCIAL NETWORKS IN HONEYBEE COLONIES
Honeybee colonies are amazing adaptive systems. However, the colony level behavior results from the individual responses of each of the (up to 30) thousand members. How does the colony adapt to vaying food supply, weather changes, honey robbers and so on? Local information has to be integrated and distributed. Just like in the brain, the interaction network defines the output of the collective information processing. The BeesBook project develops a system to track all bees in a colony and analyze the dynamic and multi-modal interaction networks.
AN OPEN-SOURCE VIDEO TRACKING FRAMEWORK FOR THE ANALYSIS OF ANIMAL BEHAVIOR
Most vision applications for the analysis of animal behavior use overlapping sets of functionality (video I/O, mouse and keyboard interaction, state representations of tracked objects, etc). The Biorobotics Lab has developed a modular, multiplatform, open source C++ framework saving time and money for biologists and computer vision developers.
ROBOTIC FISH SWARMS FOR THE ANALYSIS OF COLLECTIVE BEHAVIOR
Fish often organize in schools which improves each individual's probability of survival. However there are fascinating details on the underlying mechanisms that we investigate with Prof. Krause and his team (HU Berlin, Leibniz IGB). We have built fish robots that can autonomously interact with the fish, recruit them to locations in the tank or just swim along. We are currently investigating models of leadership in Guppies.
In the DFG funded project Leo-III a novel agent-based deduction system for classical higher-order logic (HOL) is developed which aims at exploiting massive parallelism at various levels in the reasoning process. The Leo-III system allows ad-hoc inclusion of independent specialist agents that add advanced functionality to the proof search such as consistency checks of the input axiomatization using model finders or augmented deduction rules for non-classical logics. The latter, very powerful, capability is enabled by semantical embedding of the desired goal logic in HOL. Several of such embeddings will be included in Leo-III, yielding an out-of-the-box automation tool for a great number of (quantified) non-classical logics relevant in mathematics (e.g. inclusive/free logic as relevant for category theory), philosophy (e.g. modal logics) and computer science (e.g many-valued logics, paraconsistent logics). In its current state, Leo-III is based on an ordered paramodulation calculus for typed lambda-terms, augmented with special means of extensionality treatment. The employment of agents allows parallelism on the search level by introducing and/or splits of the search space. The scheduling of the agents' actions is realized as optimization procedure using combinatorical auction games.
In this ongoing project we have achieved a landmark result that is bridging between mathematical logic/theoretical computer science, artificial intelligence and theoretical philosophy: we have for the first time automated Kurt Gödel’s ontological proof for the existence of God on the computer. Gödel’s ontological argument is in the tradition of the work of prominent philosophers, including St. Anselm, Descartes, and Leibniz. Most interestingly, the higher-order automated theorem provers we employed (in particular, our own prover LEO-II) found some new results on Gödel’s proof. This work has received media repercussion in a global scale, and yet it is only a glimpse of what can be achieved by combining these fields in the future.