Online Gaussian Process Learning and Posterior Optimization for Path Planning in a Multi-Agent Observation System

Abstract

In heterogeneous multi-agent robotic systems, each agent must adapt its decisions to new information from other agents with complementary sensing capabilities. In particular, the task of path planning in environmental exploration and monitoring scenarios depends highly on the current belief of the world. We present a Gaussian Process based path planning method that adapts to new data and incorporates path planning constraints. This method can run in real time and integrate data from multiple sources. We optimize a cost function dependent on the posterior of the Gaussian Process at the end of the optimization horizon, parametrized by the future waypoints. The optimal choice of waypoints are found by solving an optimal control problem. We demonstrate this method on a multi-agent sensing scenario in which an autonomous surface vessel monitors algal blooms in an ocean region while periodically receiving noisy exogenous data from a satellite.