Homogeneous Transformation Matrix Based Neural Network for Model Based Reinforcement Learning on Robot Manipulator

This paper introduces a novel architecture of neural networks based on a homogeneous transformation matrix for model-based reinforcement learning in robotics. It shows how the homogeneous transformation neural network outperforms the standard feed-forward neural network in the robot manipulator area. The homogeneous transformation matrix can be an alternative for modeling robotic manipulators as a neural network. Since, there are a lot of differences in the robotic manipulator model, especially in geometry, a geometry transfer is introduced in this paper. It shows how to transfer the feature to the other homogeneous transformation matrix for a different type of geometry of the robot manipulator.