Research

Heuristic Problem-Solving

Real-world problems are often difficult to model in a small set of variables, which are needed to define a classical optimization problem. Inspired by human problem-solving abilities, wie explore heuristic strategies to problem-solving. [more information]

 


 

Representation of Everyday Objects

In AI objects are usually represented in the form of ontologies. Observations from psychology show that humans represent objects in a completely different way. Especially, the boundary between object classes is neither fixed nor can it easily be defined. In this project, which is funded by the ministry of science of Baden-Württemberg, we explore the use of prototype theory from psychology to represent everyday objects. Our test scenario is a kitchen, where a robot has to store objects according to their functional and geometric properties, so  that humans can easily find them. [more information]

 

 

 

Personal Mobility and Habits

 

In times of fast growing cities and urbanisation people's needs changes continuously in terms of personal mobility. In large cities the user can choose among a lot of mobility options, e.g. subway, car, cab, Uber and bike, to get from location A to destination B. It is not trivial to find the best mobility options for a user regarding all available mobility options in respect to his daily agenda. With more options usually more possible intermodal combinations are possible to reach a destination. Mostly, the consideration of all available information is extremely time-consuming, for instance traffic, transportation schedules, disturbances or local (football) events, to find the best commuting option for each specific day. According to former research a person is strongly driven by mobility habits. To ease the daily commute planning, an intelligent mobility assistant could learn a user's individual mobility habits and inform the user about the best mobility option based on the user's habits / daily agenda. [more information]

 

 

Opportunistic Action Selection

Robots are often considered as inflexible and "only do what they have been programmed to do" (although current systems often do not show the behaviour that was intended by their creators). Their behaviour is often incomprehensible to humans, who are well-trained to adapt their actions to their changing environment. The MORPH (Model-based Opportunistic Robot Planning for Human-robot Collaboration) project explores how a robot can recognize and use opportunities in a highly dynamic household environment to show more flexible and understandable behavior. [more information]

 

 

Human-aware Navigation

Robot path planning has traditionally concentrated on collision-free paths. For robots that collaborate closely with humans, however, the situation is different in two respects: 1) the humans in the robot's environment are not randomly moving objects, but cognitive beings who can deliberately make way for a robot to pass and 2) the quality of a navigation plan depends less on quantitative efficiency criteria, but rather on the acceptance of humans. We work on a robot navigation approach that takes into account human-centered requirements and the collaborative nature of the interaction between the human and the robot. [more information]

 

 

Simulation for Human-Robot Interaction

Realistic simulators are a widespread tool for developing intelligent robots. For human-robot interaction, a human has to be inserted realistically into the simulation. We contribute to open source simulation development to and evaluate the use of avatars in such simulations for human-robot interaction research. [more information]

 

 

 

Robot Learning

The interaction with humans requires knowledge about the abilities of the human and the robot as well as models about preferences of individual humans. We use machine learning techniques to learn prediction models and use these models in efficient search algorithms to determine parameters for the program execution. [more information]

 

 

 

Expectation-based failure recognition

For intelligent interaction and for the recognition of failures, a robot needs knowledge about the execution of actions (both human and robot actions). Similar to operators in planning problems, the robot needs to know the expected outcome of actions. In addition, a robot needs knowledge about events that should not happen during the action or how an action typically proceeds. We develop complex action models based on hybrid system modeling, combined with learned predictions. [more information]

 

 

Measuring Legibility in Human-Robot Interaction

A fundamental question for human-robot interaction is how to measure the acceptance of a robot in everyday working and living environments. We have particularly looked at the notion of legibility, which is the degree to which a person expects and can predict the robot's action. As a specific task, we have examined robot navigation, but the basic approach can also be applied to other tasks. [more information]