Deep Regression Models for Local Interaction in Multi-agent Robot Tasks

A direct data-driven path planner for small autonomous robots is a desirable feature of robot swarms that would allow each agent of the system to directly produce control actions from sensor readings. This feature allows to bring the artificial system closer to its biological model, and facilitates the programming of tasks at the swarm system level. To develop this feature it is necessary to generate behavior models for different possible events during navigation. In this paper we propose to develop these models using deep regression. In accordance with the dependence of distance on obstacles in the environment along the sensor array, we propose the use of a recurrent neural network. The models are developed for different types of obstacles, free spaces and other robots. The scheme was successfully tested by simulation and on real robots for simple grouping tasks in unknown environments.