Full State History Cooperative Localisation with Complete Information Sharing

Abstract

This thesis presents a decentralised localisation method for multiple robots. We enable reduced bandwidth requirements whilst using local solutions that fuse information from other robots. This method does not specify a communication topology or require complex tracking of information. The methods for including shared data match standard elements of nonlinear optimisation algorithms. There are four contributions in this thesis. The first is a method to split the multiple vehicle problem into sections that can be iteratively transmitted in packets with bandwidth bounds. This is done through delayed elimination of external states, which are states involved in intervehicle observations. Observations are placed in subgraphs that accumulate between external states. Internal states, which are all states not involved in intervehicle observations, can then be eliminated from each subgraph and the joint probability of the start and end states is shared between vehicles and combined to yield the solution to the entire graph. The second contribution is usage of variable reordering within these packets to enable handling of delayed observations that target an existing state such as with visual loop closures. We identify the calculations required to give the conditional probability of the delayed historical state on the existing external states before and after. This reduces the recalculation to updating the factorisation of a single subgraph and is independent of the time since the observation was made. The third contribution is a method and conditions for insertion of states into existing packets that does not invalidate previously transmitted data. We derive the conditions that enable this method and our fourth contribution is two motion models that conform to the conditions. Together this permits handling of the general out of sequence case.