Learning motion primitives of object manipulation using Mimesis Model

In this paper, we present a system to learn manipulation motion primitives from human demonstration. This system, based on the statistical model “Mimesis Model”, provides an easy-to-use human-interface for learning manipulation motion primitives, as well as a natural language interface allowing human to modify and instruct robot motions. The human-demonstrated manipulation motion primitives are initially encoded by Hidden Markov Models (HMM). The models are then projected to a topological space where they are labeled, and their similarities are represented as their distances in the space. We then explore the unknown area in this space by interpolation between known models. New motion primitives are thus generated from the unknown area to meet the new manipulation scenarios. We demonstrate this system by learning bimanual grasping strategies. The implemented system successfully reproduces and generalizes the motion primitives in different grasping scenarios.