Object Rearrangement with Continual Imitation Learning

Abstract

In recent years, robots have been expected to replace human tasks, and approaches based on machine learning have been attracting attention as a method for realizing robots that can cope with diverse environments. In particular, imitation learning, which uses human manipulation data to learn, is a highly sample efficient method, and has been shown to achieve certain success rates for various tasks. However, it is still difficult to achieve a success rate close to 100\% for various conditions and to completely automate the task. Therefore, we propose a method to improve the performance of the system step by step by having the system operate autonomously under human supervision, intervening when the task fails, and using the data from the intervention for additional learning. In this study, we demonstrated the effectiveness of continual learning by intervention for the display operation of a manipulator.