DARPA Urban Challenge, a C++ based platform for testing Path Planning Algorithms: An application of Game Theory and Neural Networks

摘要

The DARPA Grand Challenge in which the Cornell Racing Team participates requires the completion of aSimulator, which purports all errors in the artificial intelligence path planning down below and back up.The simulator comes as the last layer in the top down approach followed by the Cornell Racing Team.The Strategic layer is charged of global route planning, the tactical layer of collision avoidance and maneuverplanning, while the operational layer controls lane tracking and safe following.The simulator is the last layer. Through a COBRA interface the C++ or C# version of the simulator will bereceiving commands from the Artificial Intelligence Strategic Layer concerning maneuvers such as Turn Left,Turn Right, Change Lane, Increase Speed, and Stop.The simulator induces from its current situation, using controls such as bounding boxes and the World class,pointing to every object in the World, a set of more detailed commands.Apart from writing a simplified version of the simulator in C++, we also concentrated my efforts onto findinga solution aside from dynamic programming for Path Planning and the Behavioral Modeling of Visible andNeighboring Vehicles on the road network.We have built an efficient and self-correcting C++ GUI Interface including some random moving vehicles aswell as a smart vehicle named Autosmart.The Path Planning algorithm is written and implemented although may be missing a more significant round oftesting.To do so, we are using the approach of game theory and artificial intelligence’s neural networks. We representthe world as nature, resulting in decisions independent of the drivers (types: turn left or right at the nextintersection); nature being in this case the DARPA Challenge organizers. Moreover the drivers chose theirbehaviors (aggressive, altruist) on the road and keep updating their anticipations about the other playersbehavior and types, as mentioned above.The end result is to train these neural networks to react to previously categorized behaviors and situations bystoring necessary information about the ‘game’.Every player runs its own network, although in our case we limited the simulation to one smart vehicle,Autosmart and 2 random vehicles; therefore by nature the algorithm the algorithm would lead to biasedresults.It is meant for simplicity since if not for programming the set of commands which lead to adequate behavior atintersections and on segments, such as being done for the smart vehicle; sometimes the random vehicles getinto trouble, being too much off the road network.In most cases, the simulator will self-correct their path however.