ABSTRACT
In multirobot path planning, optimizing each robot’s position is important. Machine learning models, such as DARP, are used to optimize these paths and predict future paths given a set of starting positions and obstacles. This ensures complete coverage of the given maps. This paper explores the use of DARP to optimize the initialization and exploration of multirobot systems, and output the best possible initial positions for a set of robots, given a map and its set of obstacles. A dataset, REIOset, of 4000 elements containing, randomized maps, initial values, partition standard deviation, and co-visibility was created. A convolutional neural network (CNN) is then used to train the generated map and position data, to output optimized initial positions, trained on image and parameter data. Despite inconsistent predictions from our model, an improved version of this algorithm can be utilized for purposes such as robot swarm formation, other collaborative multi robot systems, or relative pose prediction.
Keywords: mCPP, DARP, Convolutional Neural Networks, Path Planning, Co-visibility