TOYON RESEARCH CORPORATION — Department of Defense SBIR Phase I: A16-120

Description

Toyon Research Corp. proposes to develop neural-network-based algorithms for the autonomous control of a vehicle, to follow a leader vehicle driven by a human. Traditional methods are rule-based, requiring responses to be explicitly programmed for each state encountered by the vehicle. However, such methods perform poorly when the vehicle encounters a new state, a disadvantage our approach will address. Taking a deep learning approach, by training the neural networks with a large corpus of data, the follower vehicle will be able to learn to respond a wide range of scenarios, even those it has never encountered before. A combination of Convolutional and Recurrent neural network architectures is proposed, to combine the spatial feature extraction capability of the former, and the ability to learn long-term temporal relationships of the latter. To allow further training of the neural network, state-of-the-art Reinforcement Learning algorithms will be investigated to allow unsupervised training of the networks. By allowing a limited degree of freedom to the follower vehicle, it can explore the environment and build further knowledge of the environment and the task, improving performance. Through the deployment of these algorithms on driving simulators, the neural networks can be trained for a large number of driving