Research

Research Issues

Since 1994, I am working on methods that allow mobile robots to autonomously act in unknown, dynamic environments. On this page you will find a brief overview of the topics I have addressed over the last years. Most of the work has been done jointly with Wolfram Burgard and Sebastian Thrun. I higly recommend to look at the page introducing my most recent work.

Mobile robot navigation and planning

The key design principle of our software architecture for mobile robots is the application of probabilistic methods for dealing with the inherent uncertainty in the robot's sensors and actuators. To test the reliability of our system, we installed two of our robots in densely crowded museums, where they successfully acted as robotic tour-guides over several weeks.

Look at related papers

Position estimation

This was the main focus of my work over the last years. Most of the issues are addressed in detail in my doctoral thesis. Our grid-based approach to Markov localization meets the following requirements:

Global Localization:: It is able to estimate the position of a mobile robot without knowledge of its initial location. Furthermore, it detects situations in which the position of the robot is lost and can recover from such situations.
Localization in dynamic environments:: In order to reliably localize a mobile robot even in dynamic environments such as a crowded museum, our approach uses a technique which filters sensor data. These filters are designed to eliminate the damaging effect of sensor data corrupted by unmodeled dynamics.
Active Localization:: The efficiency especially of global localization can be improved by actively disambiguating between different possible locations. Key open issues in active localization are ``where to move'' and ``where to look'' so as to best localize the robot. In order to derive means for determining the best action with respect to localization, we introduced a decision-theoretic extension of Markov localization. By choosing actions to minimize the expected future uncertainty, our approach is capable of actively localizing a mobile robot from scratch.

Monte Carlo Localization: Recently, we replaced the grid representation of the density over the robot's state space by an efficient sample based representation. See my most recent work.

Look at related papers

Map building

The problem of map building is even harder than that of map-based position estimation. Here, in addition to estimating the position of a robot, the map has to be estimated simultanuously. Based on previous experiences in map building and global position estimation we came up with an approach to concurrent map building and localization. The approach uses the EM-algorithm to estimate the most likely map given the robot's observations.

Look at related papers

Collision avoidance

Most existing approaches to safe navigation rely on a purely sensor-based, reactive collision avoidance. In order to overcome the limitations of this paradigm, we combined our method for position estimation with our dynamic window approach to reactive collision avoidance. The resulting hybrid approach to collision avoidance differs from previous approaches in that it considers the dynamics of the robot and avoids collisions with invisible obstacles even if the robot is uncertain about its position.

Look at related papers

Selected papers on Mobile robot navigation and planning