A developmentally inspired transfer learning approach for predicting skill durations

As robots are increasingly integrated into daily life, one of the most important roles they will assume is that of collaboratively helping us perform physical tasks. Be it helping us put together furniture, transporting materials, or assisting with food preparation, a system´s ability to assess its (and others´) skill level regarding the performance of different tasks is essential to achieving efficient scheduling and collaboration. In this paper, we present preliminary work towards an observation-driven modeling approach allowing an agent to autonomously predict the amount of time required for different agents to complete actions. This approach utilizes insights and observations from the developmental psychology and operations research communities to accurately develop agent-personalized skill proficiency models. We demonstrate our model by evaluating its performance at estimating agent performance in a set of common assembly tasks. Our evaluation measures knowledge-transfer via novel task introduction, as well as extrapolation by predicting future performance given previous experience.