DRL: Deep Reinforcement Learning for Intelligent Robot Control --\n Concept, Literature, and Future

Abstract

Combination of machine learning (for generating machine intelligence),\ncomputer vision (for better environment perception), and robotic systems (for\ncontrolled environment interaction) motivates this work toward proposing a\nvision-based learning framework for intelligent robot control as the ultimate\ngoal (vision-based learning robot). This work specifically introduces deep\nreinforcement learning as the the learning framework, a General-purpose\nframework for AI (AGI) meaning application-independent and\nplatform-independent. In terms of robot control, this framework is proposing\nspecifically a high-level control architecture independent of the low-level\ncontrol, meaning these two required level of control can be developed\nseparately from each other. In this aspect, the high-level control creates the\nrequired intelligence for the control of the platform using the recorded\nlow-level controlling data from that same platform generated by a trainer. The\nrecorded low-level controlling data is simply indicating the successful and\nfailed experiences or sequences of experiments conducted by a trainer using the\nsame robotic platform. The sequences of the recorded data are composed of\nobservation data (input sensor), generated reward (feedback value) and action\ndata (output controller). For experimental platform and experiments, vision\nsensors are used for perception of the environment, different kinematic\ncontrollers create the required motion commands based on the platform\napplication, deep learning approaches generate the required intelligence, and\nfinally reinforcement learning techniques incrementally improve the generated\nintelligence until the mission is accomplished by the robot.\n