Learning pouring skills from demonstration and practice

Abstract

This paper focuses on improving performance with practice for tasks that are difficult to model or plan, such as pouring (manipulating a liquid or granular material such as sugar). We are also interested in tasks that involve the possible use of many skills, such as pouring by tipping, shaking, and tapping. Although our ultimate goal is to learn and optimize skills automatically from demonstration and practice, in this paper, we explore manually obtaining skills from human demonstration, and automatically selecting skills and optimizing continuous parameters for these skills. Behaviors such as pouring, shaking, and tapping are modeled with finite state machines. We unify the pouring and the two shaking skills as a general pouring model. The constructed models are verified by implementing them on a PR2 robot. The robot experiments demonstrate that our approach is able to appropriately generalize knowledge about different pouring skills and optimize behavior parameters.