Promotionsvortrag: “Experience-Based Behavior Adaptation of Kinematically Complex Robots”

In the last decades, robots have increasingly evolved from blindly acting machines towards more and more intelligent agents that feature many actuators and sensors in combination with advanced processing units. Besides precisely repeating actions, these kinematically complex robots provide the flexibility and mobility to react to changing conditions with the help of sophisticated control approaches. Thereby, reactive and/or planning-based motion controls are usually used that need to be configured to generate solutions for a limited context range, i.e. for specific commands, environmental properties, or conditions of the robot. In addition, different configurations of the motion control may fulfill the same action but with different scopes of performance, e.g. precision, stability, or efficiency. Consequently, to obtain a generic robot that can perform actions in various situations while maintaining an application-specific optimal performance, the robot must be able to adapt its motion control during runtime according to the context.
The goal of this thesis is to increase the contextual range in which a robot can act by autonomously adapting its motion control, and thus its behavior, on the base of context information and gained experiences. In this thesis, a concept is presented how experiences - i.e. values for application-specific evaluation criteria of an executed behavior in a detected context - can be generated, managed, and used for behavior adaptation. The presented approach allows the continuous integration of new experiences, whereby the foundation for autonomous behavior adaptation grows with the lifetime of the robot and thus considers possible condition changes such as wearout. To derive the supposedly best behavior for the current context based on experience, two different methods are implemented and compared to exploit their pros and cons. The first approach is based on case-based reasoning, while the second implements a model-based approach that uses an incremental Gaussian process regression as an internal simulation model to search for an optimal behavior.
The experience-based behavior adaptation is analyzed for the use case of adapting walking behaviors during runtime using two kinematically complex robots as examples, the hexapod SpaceClimber and the four-legged walking robot Charlie. Therefore, a generic locomotion control is developed that generates walking behaviors for various types of robots, i.e. it generates different poses and walking patterns with their advantages and disadvantages depending on its configuration. In addition, evaluation criteria and metrics to determine the context are defined and applied. Knowledge bases are generated through manual and automated testing of different behavior parameter configurations. The experimental evaluation shows that the gathered experiences in combination with the presented approach allow a successful behavior adaptation for the target systems. Both investigated adaptation strategies show advantages and disadvantages that justify their respective utilization depending on the application. The case-based reasoning approach has its strengths, if behavior parameters need to be adapted on the foundation of few experiences. The model-based approach is best suited for large knowledge bases to derive fine-grained adaptations.