Deep reinforcement learning based multi agent pathfinding

AirSim simulation results from the MAPF controllers developped in the ME5001 (master's) project "Deep reinforcement learning based multi-agent pathfinding" by Zhiyao Luo. In this video, we setup various 2D and 3D scenarios to illustrate how PRIMALc and 3D-M* plan collision-free paths for multiple UAVs. These drones are controlled directly by position commands, and their altitude is considered constant in the 2D cases. Note how the drones can move smoothly along the discrete path point controlled by built-in controller without any collisions. 2D cases are solved with the AI-based, decentralized path planner PRIMALc, described in the associated report, while the 3D cases show results from a 3D extension of M* (since training of 3D-PRIMAL could not be completed due to time).