Robot Formations as an Emergent Collective Task using Target-following Behavior

Abstract

Robot formations imply the establishment and the maintenance of a predetermined geometric shape by a group of robots. In this work, we achieve formations using robots with proximity sensors of short detection range and no inter-robot communication equipment. This fact turns the synthesis problem into a hard one compared to those cases where the robot has information about the absolute location of the other robots, or has sensors with a higher range limit. The robot formations emerge from an individual behavior called target-following which we synthesize using Reinforcement Learning (RL). We propose a task decomposition technique based on an action space transformation that allows us to reduce considerably the time needed for learning and overcomes some difficulties that arise with the use of RL for this particular problem. Preliminary results support the feasibility of this method for the synthesis of robot formation behaviors of different shapes.