Enhancing Unsupervised Language Grounding through Online Learning


To enable natural human-robot collaboration robots need to understand natural language, which in turn requires mechanisms to ground language by connecting words to corresponding percepts. This paper, presents an unsupervised online grounding framework for grounding of synonymous words through corresponding percepts. The framework is evaluated through a human-robot interaction experiment and compared to a Bayesian grounding framework, which requires an offline training phase. The results show that the proposed framework outperforms the baseline, thereby illustrating the benefit of online learning for natural language grounding.

ICRA 2020 Workshop ``Shared Autonomy: Learning and Control''