End-to-End Learning Driver Policy using Moments Deep Neural Network

A common practice for autonomous driving is to train a model to mimic expert actions. However, the actions are randomly drawn based on a underlying policy. The policy specifies the probability distribution of the expert taking different actions under each circumstance. An ideal driving model should mimic the policy, rather than the action. In this paper, we propose a new method to mimic the policy. Our method employs deep neural network to map each circumstance to the probability distribution of actions. In practice, the distribution is approximately represented by its low order moments. Hence, our model is called Moments Deep Neural Network (MDNN). To improve the efficiency, MDNN employs a new type of dilated convolution we term the Dilated Tensor Convolution (DTC). DTC applies the idea of dilated convolution on every dimension of the input, reducing the number of parameters in MDNN. On multiple experiments, MDNN is demonstrated to outperform the classical deep neural network PilotNet using much less parameters.