Track: Artificial Intelligence

Abstract

In this paper, we present an autonomous vehicles (AV) research testbed with photo-realistic visualisation and
simulation capability. Compared to Oak Ridge National Labs CAVE system, Microsoft's AirSim, and MATLAB's Vehicle Dynamics
Blockset approach, the key difference is a decoupled and modular system architecture which allows developers
to focus on only one aspect of the whole system and be productive with minimal training. Basically, the proposed
system has a game engine, an external hardware accelerator for vehicle dynamical model, and a GPU equipped
computer to run the AI algorithms. Our vision is to have all vehicle dynamical models to be built in MATLAB/Simulink and run
on a hardware accelerator like Speedgoat, which is a leading software and hardware framework for scientific system
modelling and hardware in the loop simulation. The proposed photo realistic visualization environment is Unreal Engine based, and all AI models are keras based. This decoupled system architecture allows all vehicle dynamical models and AI decision making to be implemented outside the Unreal Engine, while the game engine doing all collision detections and reporting. Similarly, all traffic and pedestrian movement stochastic models can be implemented outside the game engine, allowing a framework agnostic approach to be used throughout the whole system.