Offline reinforcement learning application in robotic manipulation with a COG method case

Artificial intelligence now has different applications in various industrial fields. Reinforcement learning (RL) is one of the hot topics in the artificial intelligence, also in robotics. It is an important learning method in the field of robotic manipulation. The training policies of reinforcement learning can be divided into online learning policy and offline learning policy. Besides, the reinforcement learning algorithm of offline policy has great potential in transforming large data sets into powerful decision engine. To solve the problem that most of robot applications involve collecting data from scratch for each new task, offline learning combined with online learning is to make the training more efficient and convenient. The aim of this paper is to clearly introduce the application of offline reinforcement learning in the field of robotic manipulation. The basic formulation of reinforcement learning includes two points: First, it introduces Markov Decision Process and one of method of solution – policy gradients. Then through analyzing an application of offline learning in the field of robotic manipulation - COG algorithm, this paper analyzes the process of offline learning combining the prior data to learn new robotic skills and uses this method to solve specific tasks of robotic, such as the problems of sample efficiency. The results show that the offline learning policy has important research value in the field of robotic manipulation by reducing training time and make process efficient, and it fully embodies its advantages in solving the problems of robotic sample efficiency.