Enhancing Robotic Manipulation: AR-Powered Data Collection for Learning from Demonstration

Thilina Tharanga Malagalage Don, Hamid Khayyam, Ehsan Asadi

Keywords: AR, LFD, HRI

TL;DR: Paper presents a study on the performance of machine learning models used for robotic manipulation, trained using task demonstrations powered by AR based methods.

Abstract:

Integrating robotic manipulators into everyday
households faces the significant challenge of allowing them to
be taught skills in a natural and humanly understandable way.
Although learning-from-demonstration (LFD) shows promise, its
reliance on quality data and cumbersome demonstration methods
limits its broader application. This paper presents a comparison
study on the performance of machine learning models, trained
using task demonstration carried out via two traditional methods,
two traditional methods augmented with augmented reality (AR), and one
augmented reality based method. We compare the performance
of these input methods against four ML models and two input
data modalities. The results demonstrate the advantage of using
AR augmented methods in data collection for LFD and the
pure AR method nearly matches the performance of the highest
performing AR augmented traditional method while having no
drawbacks of the traditional methods.

Submission Number: 3