Towards a Simulator for Imitation Learning with Kinesthetic Bootstrapping

Abstract

This paper presents a physics based simulator that allows kinesthetic interactions between a human and a robot to be recorded, and later used for imitation learning. We argue that kinesthetic interac- tion can be a very important tool for programming robots, if properly supported by a simulation engine. For this, we propose a new scheme for robot motion learning based on kinesthetic bootstrapping. Interactions with the robot are used to create a low-dimensional posture space. The posture space together with the presented simulation engine allow for fast imitation learning of behaviors. Early results of this approach, using Genetic Algorithms as learning technique, will be presented.