Towards vision-based deep reinforcement learning for robotic motion control

Abstract

Manipulation in highly dynamic and complex environments is challenging for robots. This paper introduces a machine learning based sys-tem for controlling a robotic manipulator with visual perception only. The capability to au-tonomously learn robot controllers solely from camera images and without any prior knowl-edge is shown for the first time. We build upon the success of recent deep reinforcement learn-ing and develop a system for learning target reaching with a three-joint robot manipulator using external visual observation of the ma-nipulator. In simulation, a Deep Q Network (DQN) was demonstrated to perform target reaching after training. Transferring the net-work to real hardware and real observation in a naive approach failed, but experiments show that the network works when replacing cam-era images with synthetic images generated by a simulator according to real-time robot joint angles. 1