Unsupervised Bayesian Online Learning for Multi-Agent Exploration in an Unknown Environment

Abstract

Exploring an unknown environment with multiple robots is an enabling technology for many useful applications. This paper investigates decentralized motion planning for multi-agent exploration in a field with unknown distributions such as received signal strength (RSS) and terrain elevation. We present both supervised with RSS distribution and unsupervised methods with terrain data. The environment is modelled with a Gaussian process using Bayesian online learning by sharing the information obtained from the measurement history of each robot. Then we use the mean function of the Gaussian process to infer the multiple source locations or peaks of the distribution. The inferred locations of sources or peaks are modelled as the probability distribution using Gaussian mixture-probability hypothesis density (GM-PHD) filter. This modelling enables nonparametric approximation of mutual information between peak locations and future robot positions. We combine the variance function of the Gaussian process and the mutual information to design an informative and noise-robust planning algorithm for multiple robots. At the end, the proposed algorithm is extended by applying an unsupervised method with Dirichlet process mixture of Gaussian processes. The experimental performance of supervised method and unsupervised method are analysed by comparing with the variance-based planning algorithm. The experimental results show that the proposed algorithm learns the unknown environmental distribution more accurately and faster.