Abstract
DeepRacer is a platform for end-to-end experimentation with RL and can beused to systematically investigate the key challenges in developing intelligentcontrol systems. Using the platform, we demonstrate how a 1/18th scale car canlearn to drive autonomously using RL with a monocular camera. It is trained insimulation with no additional tuning in physical world and demonstrates: 1)formulation and solution of a robust reinforcement learning algorithm, 2)narrowing the reality gap through joint perception and dynamics, 3) distributedon-demand compute architecture for training optimal policies, and 4) a robustevaluation method to identify when to stop training. It is the first successfullarge-scale deployment of deep reinforcement learning on a robotic controlagent that uses only raw camera images as observations and a model-freelearning method to perform robust path planning. We open source our code andvideo demo on GitHub: https://git.io/fjxoJ.