Active Incremental Learning of a Contextual Skill Model
Sähkötekniikan korkeakoulu | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Control, Robotics and Autonomous Systems
AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)
AbstractContextual skill models enable robot to generalize parameterized skills for a range of task parameters by using regression on several optimal policies. However, the task difficulty and task sequence of learning a contextual skill model is usually neglected. Thus, the learning process is usually time consuming since some tasks might be easier to learn or the knowledge of these tasks might be easier to share with other tasks. In this thesis, we introduce active incremental learning framework for actively learning a contextual skill model based on dynamical movement primitives which are widely used to learn parameterized policies on trajectory level as a dynamical system for robot. The proposed framework will first select a task which maximizes the expected improvement in skill performance over entire task parameters and then optimize the corresponding policy with a fixed number of iterations in policy search. We model the learning rate of policy search for predicting reward improvement over a single iteration. We evaluated the improvement of the skill performance in two tasks, ball-in-a-cup and basketball, with a simulated KUKA robot arm. In both, the results show that active task selection can improve the skill performance continuously over a baseline.
Thesis advisorHazara, Murtaza
contextual skill model, active incremental learning, dynamical movement primitives, policy search