Robust Deep Reinforcement Learning for Autonomous Driving

Self-driving cars are set to transform the transportation industry in a few years. While a lot of end-to-end supervised deep learning approaches have achieved success in making cars autonomous, the compounding errors that exist in the environment compel us to employ a Reinforcement Learning approach to solving this problem. We present here our results in using deep reinforcement learning methods to tackle the challenge of autonomous driving. Considering the complexity of the task we describe how we are using the TORCS environment for simulation, and a Deep Deterministic Policy Gradient approach to train our autonomous race car. We introduce ways and methods for training the algorithm so that it is robust to facing noisy sensor inputs, a common issue when dealing with complex hardware systems, without significant loss of performance as compared to an ideal noiseless system.