Robotics Motion Control by Using Deep Reinforcement Learning

Abstract

Recently, with the development of Artificial Intelligence and Deep Learning in the field of robotics, especially to control robots for specific tasks. The traditional approach uses a large number of models to calculate the most suitable and stable operating points, however, this makes it necessary to spend a lot of time to collect models and therefore this approach is not generalizable. Other methods use neural networks to predict possible operating points from the model's 3D data, which is faster but also depends on the training of the neural network. However, Deep Reinforcement Learning combines the benefits of Deep Learning in the vision domain with the generalization of Reinforcement Learning, which allows robots to learn to perform specific tasks by interacting with their environment. The purpose of this paper is to implement and improve the traditional deep reinforcement learning approach by applying popular Deep Reinforcement Learning methods to the Franka Emika Panda robot. Implementing different Deep Reinforcement Learning algorithms to control the robot to perform specific tasks such as reach, touch and grasp. Additional, improving feature extraction through secondary processing of observations. Finally, the impact of different feature exactors on the algorithm is compared.