Programming Robots through Human Demonstration: Some Recent Challenges

Abstract:

In this talk, I will review some of our recent advances in the development of robust algorithms to enable robots to learn from imitating humans. Skills designate a range of competences to achieve tasks in multiple ways. Robots can exploit this multiplicity in the solutions to show more robustness in the face of sudden changes in the environment. Our algorithms retain the variability in human demonstrations and exploit this in two ways: to solve the correspondence problem and to explore for new solutions when humans fail at the task. Examples of application for flexible manipulation of object and for quick adaptation, such as catching an object that starts falling, will be presented.