We are working towards a fully decenatralize lifelong path planner for multi-agent path finding. To that end we have built an Open.ai gym environment which can be found here: https://github.com/pehuenmoure/warehouse_rrl. Our environment environment will be used for training our agents and analyzing the improvements and limitations of implementaitons when compared to closed form and bounded path planners. We're having trouble in converging to a reasonable solution so I've been working on implementing imitation learning for a closed form solution ODRM* to speed up training. An example of the flawed agent can be seen below.

We have implemented two closed form solutions one where each agent independtly calculating A* and one with a coordinated planner. Please find the gredy planner below. Note the potentional for causing a blocking situaiton:

Below you can find the coordinated closed form planner of ODrM*, following implementation of tbe paper: https://ieeexplore.ieee.org/document/6631119. Note the yellow agent backtracking in order to clear a hallway for the red agent to be able to pass at the end of the simulation.

Next Post Previous Post