This project will take new strides in robot learning by working towards unified understanding of the large variety of representations used currently in Learning from Demonstration (LfD), Reinforcement Learning (RL), and classical motion planning. A practical manifestation will be correcting mistakes by LfD, regardless of how the robot has initially been taught or programmed. This will require research on both theoretical base of learning, as well as Human-Robot Interaction (HRI): the user must understand why the robot failed, in order to understand how to correct the issue. Also, to properly display this to the user, we must be able to transfer the skill from any representation to one what can be easily visualised and modified by LfD. As the variety of different representations of robot skills is wide, we will focus on good ways to find key points in robot skills, which would make transfer between representations easier.

2025

2029