Deep Reinforcement Learning for High-fidelity Vehicle Motion Simulation

NGA seeks to incorporate Artificial Intelligence (AI) and Machine Learning (ML) to Intelligence, Surveillance and Reconnaissance (ISR) missions in the aim to capture fleeting targets, thus a large amount of dynamic scenes with accurate target motions and behaviors will be needed for training and performance evaluation. Traditional microscopic model based approach for vehicle activity simulation is unable to produce enough fidelity for the training and the evaluation of ISR tracking, analytics, or collection strategies. Intelligent Automation Inc. (IAI) proposes to develop a Deep Reinforcement Learning (DRL) based approach to address the NGA’s critical needs. In our proposed approach, IAI innovatively incorporates a deep reinforcement learning process into a virtual traffic simulation environment, makes the vehicle activity simulation and data generation task becoming autonomously driving a virtual vehicle on any road conditions with the flexibilities to change various parameters and constraints on the fly. The proposed approach is built on top of IAI’s previous success and experience with deep learning predictive analytics and traffic simulation in various ONR, DARPA, AFRL and FHWA programs.