A Deep Reinforcement Learning Solution for the Low Level Motion Control of a Robot Manipulator System

Motion planning and control is a necessary aspect of incorporating robots into the real world. There are a variety of different types of control tasks that involve collision avoidance and fine control, that are difficult to program without the use of artificial intelligence (AI), especially in an non-stationary environment. In this paper, one method for applying deep reinforcement learning (RL) to the motion planning of a manipulator robot is described. Using a soft actor-critic (SAC) network, a model is trained to direct the manipulator to various locations so as to avoid colliding either its hand or the object it carries with a game tower. This demonstrates a simple and effective method for training an agent to achieve its goal that generalizes to similar but different environments.