Abstract

For assistive robots and virtual agents to achieve ubiquity, machines willneed to anticipate the needs of their human counterparts. The field of Learningfrom Demonstration (LfD) has sought to enable machines to infer predictivemodels of human behavior for autonomous robot control. However, humans exhibitheterogeneity in decision-making, which traditional LfD approaches fail tocapture. To overcome this challenge, we propose a Bayesian LfD framework toinfer an integrated representation of all human task demonstrators by inferringhuman-specific embeddings, thereby distilling their unique characteristics. Wevalidate our approach is able to outperform state-of-the-art techniques on bothsynthetic and real-world data sets.

 

Quick Read (beta)