System Architecture & Embodiment
Researching system concepts
The knowledge gained from this experience was condensated and refined in own research work and thesis and was put in the form of theories, methods and tools that allow us today to address new scientific and technological challenges in an innovative and efficient way. Our concepts and tools have been published and exported to other research and development institutions.
The new challenges are in the area of modeling interactive systems that have two or more strongly coupled actors with only partially observable internal representations and motivations.
Information processing and behavior generation in the automotive domain
The automotive domain has a strong demand for intelligent mobility solutions. Vehicles need to be safer, have better economics and provide more comfort and fun.
Those demands can only be met if the systems have a better understanding of the situation in the form of the environment and the driver intention and if they can act in a more foresighted fashion. This raises scientific questions about the Environment Perception (traffic structures and their conditions), about the Situation Understanding, Prediction and Behavior Choices (infer what will most likely happen, how it will influence my behavior options and how my choice will influence the overall situation in return), about the communication with and the inferred intention of the driver and finally about the successful execution of the selected behavior option. All those questions are tightly coupled and the success of giving one answer depends on the quality of the answers for the other questions. This is strongly reflected in the architecture research that aims at sorting the system's elements in such a way that a meaningful approach towards a well performing whole can be pursued.
Context based prediction
Advanced Driver Assistance Systems have a big share in the reduction of accidents and the increase of driver’s comfort. The value of these systems can be further increased by providing them with the human-like ability of foresighted driving, thus allowing them to predict the behavior of other traffic participants and (re-)act to their behavior before it becomes apparent.
To achieve this level of prediction, we are considering prediction as a pattern recognition problem, exploiting the maturity of pattern recognition algorithms. But the crucial research question is: What patterns should we look for?
In “Context Based Prediction”, we are considering a pattern as the specific relation between a traffic participant and its environment – the so called “context”. For example, from a vehicle quickly approaching its predecessor on a highway, we can predict that it will change lane if there is fitting gap.
Motion primitives & behavior generation in robotics
Future robotics applications will require complex movement skills that coordinate many degrees of freedom and incorporate sensory feedback of different kinds. Examples are multi-finger manipulation for assembling processes, bipedal locomotion in cluttered environments or human-robot cooperative tasks, to name only a few.
In order to tackle such problems, this research targets to devise movement representations and control concepts using the concept of Movement Primitives. It originates from biology and has been found to be a guiding mechanism to coordinate movement in animals and humans. We target to learn such building blocks of movement from human demonstrations, and acquire more complex skills that require the coordination of several of such primitives. While classical approaches design force interaction patterns, we think that it is important to learn a general representation that incorporates kinematic, tactile and interaction force measurements. Such learned representations will be capable to realize tasks that require the coordination of kinematics and force. They will have abilities to generalize to new situations, and to recover from large disturbances.
These abilities are an important prerequisite to achieve tasks in which a robot physically interacts with its surrounding.
Systems meet reality
The way from scientific research methods to convincing embedded proto-types can be long and cumbersome.
New results need to meet reality quickly in order to show their potential early or to guide the next research state towards the right challenges. Besides understanding the system architecture as a challenging research questions for its own we have researched and implemented several methods and tools for bringing new research results quickly into operation in embedded artifacts.
This comprises a software development and integration framework, means and tools for training and evaluation of learning systems as well as approaches for visualizing and analyzing integrated system properties.