Reliable Multi-robot Coordination Using Minimal Communication and Neural Prediction

In many multi-robot applications, such as robot soccer, robot rescue, and exploration, a reliable coordination of robots is required. Robot teams in these applications should therefore be equipped with coordination mechanisms that work robustly despite communication capabilities being corrupted. In this paper we propose a coordination mechanism in which each robot first computes a global task assignment for the team that minimizes the cost of achieving all tasks, and then executes the task assigned to itself. In this coordination mechanism a robot can infer the intentions of its team mates given their belief states. Lack of information caused by communication failures causes an increase of uncertainty with respect to the belief states of team mates. The cost of task achievement is estimated by a sophisticated temporal projection module that exploits learned dynamical models of the robots. We will show in experiments, both on real and simulated robots, that our coordination mechanism produces well coordinated behavior and that the coherence of task assignments gracefully degrades with communication failures.