Multi-UAV target search using decentralized gradient-based negotiation with expected observation

This paper presents a novel approach for the coordination of a team of autonomous sensor platforms searching for lost targets under uncertainty. A real-time receding horizon controller in continuous action space is developed based on a decentralized gradient-based optimization algorithm and by using the expected observation as an estimate of future rewards. The expected observation is a cost-to-go heuristic that estimates the goodness of the states that the platforms could reach. It permits the decision making algorithm to take into account the information on the whole environment, reducing the time needed to detect the target. The heuristic, modeled as a sensor, allows us to develop a new team utility function with low computational cost and high performance. It can be applied to challenging scenarios such as multi-target search with complex and non-uniform target probability distributions. Through simulation and statistical analysis, we show the advantages of using the expected observation heuristic in multi-vehicle coordination for search applications.