Curriculum learning for robotic peg-in-hole assembly

Abstract

Robotic assembly is very essential in the smart industry. With the help of Deep Learning technology, the robots gain the ability to perform more complex tasks with less human involvement, and become more adaptive to the environmental changes, compared to the conventional ways of fine tuning the robot strategies and parameters. One of the most interesting and trending tasks is the peg-in-hole insertion tasks where the robot needs to insert a peg or pin into a hole. A lot of research has been done to increase the performance for high precision insertion tasks and deep reinforcement learning is a convincing method to realize the purpose. In this project, a curriculum learning methodology is proposed and applied to train the robot from easier tasks, such as larger clearances, and transfer the learned knowledge to train with more difficult tasks. The training algorithm in each curriculum followed is a deep reinforcement learning method. The results have shown the potential that the curriculum learning is capable for training the robot to perform much more difficult tasks that are failed by the direct training.