Storing and Predicting User-Specific Action Chains in Assistive Robots

Dies ist die aktuellste Version dieses Eintrags.

Kurzfassung

Descriptive logging of robotic actions requires a structured knowledge base that is capable of describing a variety of information, ranging from task descriptions over object properties to decision-making processes. Researchers have developed ontologies to structure robotic knowledge and also providing a rich taxonomy of concepts. Integrating such an ontology into an existing robotic framework remains a challenge, unique for different robotic contexts. This thesis describes the work of integrating the SOMA ontology into the assistive robot EDAN and building an algorithm for semantic logging of episodic memories. This structured knowledge base not only allows ex-post evaluation of robotic experiments, but also enables online algorithms to make sense of the robotic actions. Leveraging the highly structured logging of task chains of the assistive robot EDAN, this thesis proposes a light-weight algorithm to predict the next action a user might want the robot to take based on the previously executed tasks, the available tasks, and the time of day. With privacy issues in mind, the algorithm renounces any large-scale training. With the N-Gram algorithm, which has been well established for next-word prediction, as a role model, this thesis proposes a next-task prediction based solely on a chain of previously executed tasks.

Verfügbare Versionen dieses Eintrags

• Storing and Predicting User-Specific Action Chains in Assistive Robots. (deposited 20 Okt 2025 08:25) [Gegenwärtig angezeigt]