Currently my research interests concentrate in autonomous mobile robots (AMRs) and computational intelligence (fuzzy logic, genetic algorithms). Also my previous research interests in microelectronics is included in this framework. Since I studied computer science and electrical engineering I combine the theoretical aspects of cognition and intelligence phenomena with practical applications for robots. To electrical engineers my work seems highly theoretical and for computer scientists too much practical. Though these two aspects may seem contradicting, there synthesis leads to desirable results. The following list gives an overview of the main topics. Please look at my publication site for detailed papers to the topics.
Reading robots
is more or less a synonym for acquiring semantic information from sensor streams. This could be textual information out of indoor environment e.g. reading door plates but also textual information from objects. A service robot should distinguish between e.g. normal beer and alcohol free beer. This is difficult since the bottles look very similar but could be easily distinguish while reading the labels.
Fast moving or high speed autonomous mobile robotics.
2003, i build a robot (KURT2)
that drives with 4.0 m/s in our corridor (reliable controlled !!) which was during that time the fastest indoor robot of the world.
The robot is controlled by some special real-time
algorithm. The main sensor is the 3D laserscanner and wheel encoders. More informations can
be found at the high speed KURT2 homepage.
2D and 3D World modeling
A world model is a (some times fuzzy) plan of the environment of the robot. I developed a 3D world model
format for administrative buildings. The format is used to reposition the robot when it
loses his position and to plan missions and actions. The job scheduler uses also the world
model to build its optimization function.
In buildings repositioning is usually done at the end or beginning of
a corridor. Doors especially door frames could also be found by the sensor evaluation
routines and are used as additional natural landmarks. My world model is also the
fundament of several other modules. 2D and 3D descriptions (e.g. VRML) are generated from
this building description for the human machine interface HMI. Also 2D grids for high
speed robot simulation and robot behavior prediction (robot behavior anticipation) are
generated. The functionally orientated world-, robot- and sensor-modeling is also
important for designing different behaviors of the robot.
With our IAIS 3D laser scanner it is now possible to accquire 3D information completely
automatic and to build 3D maps from the environement (SLAM). Extracting and labeling objects from the 3D data e.g. for symbol grounding or symbol anchoring is an challenge task for robots also with the kinect.
Robot control architectures and robot hardware
In 1995 i specify a control architecture and hardware of the MORIA robot and the ARIADNE service robots Marvin, Odysseus and Thor .
This includes path and action planning and navigation behaviors. The navigation behavior
of the MORIA robot which is based on recurrent fuzzy systems was prized with the
intelligence award at the FUZZ/IEEE 1995 in Yokohama, Japan. Today we use ROS to control our robots.
Computational Intelligence / Machine learning
This mainly includes fuzzy logic / control and genetic algorithm but also standard machine learning, neural networks and deep learning. I combine
fuzzy logic controller and genetic algorithm for the automatic design of
robot control behaviors (Evolutionary Robotics) or the combination of different
behaviors. I am also interested in theoretical questions e.g. the difference between
fuzzy logic and probabilistic approaches, Simulated Annealing and genetic algorithms etc..
Real-time control
This general topic influences the robot architecture and especially the design and
implementation of algorithms (e.g. for navigation, planing, World modeling, sensor
evaluation and fusion ...). Real-time control includes also real-time communication. To
achieve this, the ARIADNE robots had one main PC and two Microcontrollers which share
several informations. I like the modified linux kernel
RT-linux and ROS to achieve
real-time capabilities. The servos from the 3D scanner are controlled by RT-linux
Sensors and sensor fusion
Information sensing is the basis for autonomous navigation. The robots may use ultra
sonic, laser range finder or cameras. Currently I prefer laser range finder and
wheel encoders as main sensors for robot navigation assisted by internal or external
cameras and ultra sonic sensors. Further sensors e.g. smell sensors are added on demand. I
developed an algorithm for information extraction (e.g. different directions) and obstacle
avoidance (humans, other robots) which evaluates a complete laser scan of 361 (Sick) or
500 (Schmersal) values in
O(n) (linear with the number of sensor points).
On the base of a 2D laser range finder i also designed and realized a cheap, precise,
reliable and real-time capable
3D laser range finder.
Navigation and planning
Navigation and planning are important for the robot to do useful things like post
delivery. Especially the integration of high-level reasoning and low-level execution and
the designing of robust behavior modules is important. I developed different behaviors
based on a recurrent fuzzy system that are robust behavior modules. The compositional rule
of inference is my favorite method for the integration of different behavior blocks as
well as high level planning and low-level execution. New behaviors could be added
modularly as new fuzzy rules or as a block of fuzzy rules through the system. I used
a fuzzy case tool that I developed on my own for developing the robot behavior.
Standard graph search algorithms are used to find the robot paths and the actions (e.g.
opening a door) in the world model.
Human Machine Interfaces (HMI)
Since my robots work really autonomously, they do not have to be teleoperated. But the
robots need an adequate user interface because they have to be controlled, ordered,
monitored or maintained by humans. Internet technologies, e.g. client/server models, offer
the possibility to realize such an user interface highly comfortable for the user, because
the clients (e.g. Netscape Navigator or MS Internet Explorer) work on different platforms
with the user's operating system and known environment. So, currently I implement a server with the following features:
Spatial Cognition in indoor environments
This includes spatial representations and actions, representation and activation of route
knowledge etc. In principal this means the analysis of intelligence phenomena. In real
world applications only a few aspects of the whole theoretical framework are realized.
Currently work concentrates in the automatic generation of (recurrent) fuzzy rule based
systems for service robots. This is done with a distributed Genetic Algorithm on a
standard computer network. The fuzzy rules should be more or less equal to rule
description of human experts.
High disposability, reliability and robot security
Today in the context of internet/intranet and client/server technologies security aspects
gain increasing interest. Therefore I added the following security features to our system: