Learning Skills from Human Demonstrations

Many robots are designed for use in domestic environments where robots will be engaged in household chores. The robots need to learn ways to do the household chores that humans are now doing. We are taking a learning from demonstration (LfD) approach to this problem [1]. In terms of the household chores, a number of tasks are developed so far; for example, bringing a beer bottle from a refrigerator to a human, making pancakes [2], and folding towels [3]. However, a key issue for robots to do household chores is how to treat different versions of each task. Consider an opening task. There are a number of ways to open a container: rotating a cap on a plastic bottle, pulling a hinge cap of a ketchup bottle, pulling a poptab of a beer can, tearing a bag of potato chips, and so on. In addition, when opening a tight jar, we will use a different way to open it, like holding a cap with a wet towel. We call these methods skills. Learning these skills is essential for robots to fully handle tasks. In this research, we treat a pouring task to study skill learning; its purpose is to move material. Humans use many skills to pour, such as shaking a bottle to pour viscous liquid like ketchup, tapping a bottle to pour a little amount of coffee powder, squeezing a shampoo bottle, and pushing a soap pump. Thus, the pouring task is a good example for robots to learn skills. The goal of this research is making a general pouring behavior model from human demonstrations with which the robot can pour a wide variety of materials from a wide variety of containers. This problem is decomposed into three sub-problems: