ISimA: Intelligent Simulation for Testing and Training of Autonomous Single- and Multi-Vehicle Systems

For autonomous vehicles to be usable in time-critical missions, such as those that arise in the context of military applications, their teammates, be they human or robotic teammates, need to be able to count on the ability of the vehicles to accomplish the tasks they are given. Quite often, these tasks however must occur in highly unstructured environments, with lots of clutter, over wide range of terrains, and in hostile conditions. Currently, it is impossible to have confidence in autonomous vehicles completing their tasks under all conditions due to the limited amount of testing that can occur on real vehicles. For this reason, use of simulators provides an appealing alternative way to train and test autonomous vehicles. Unfortunately, current simulation tools don’t provide this functionality at an adequate level for several reasons. First, existing simulators are oriented much more towards structured tasks and don’t support the necessary level of fidelity of simulating outdoor clutter, various levels of terrain and adequate interaction of a vehicle with terrain, all within a single simulation software. Second, obtaining confidence in task execution requires much more than purely proper level of the simulation fidelity. We need to have a simulation software that automatically generates the conditions and controls the behavior of other dynamic entities in the virtual world in order to intelligently generate test data that builds a proper level of confidence in the ability of single- and multi-vehicle systems to accomplish desired tasks. To this end, in this project, we propose to develop a simulation software, called ISimA, that (a) provides a suite of simulation capabilities directed specifically towards the operation of single- and multi-vehicle systems performing missions in unstructured environments and on a wide range of terrains, (b)  has built-in intelligence in the generation and control of virtual world to dynamically provide test conditions that build up confidence in the ability of autonomous vehicles to accomplish their designated tasks, and (c) is open-source, and therefore easily extensible by other parties.