UC San Diego

The simultaneous Localization and Mapping (SLAM) problem for single robot systems has been a topic of intense research for many years. However, in scenarios where a single robot is not sufficient to explore and reconstruct a large or complex environment, multi-robot SLAM can prove to be a promising solution. By working and communicating together, multiple robots can cover larger areas and improve the accuracy of localization and mapping.

A traditional approach for multi-robot SLAM includes using a central server, which communicates with each robot to exchange information and updates. However, this approach is not ideal for real-time scenarios, as it can cause delays and render the computation power of each agent futile. Moreover, if the central server fails, the whole system will break down.

The main aim of this thesis is to propose a distributed filtering algorithm to build an accurate sparse global map by leveraging the computation power of multiple robots and minimizing communication across them. The approach involves each robot estimating its state locally and sharing its observations with its neighbors. The robots must agree upon an estimate of the common observations, which is achieved through consensus optimization. The algorithm is inspired by the distributed stochastic mirror descent approach to solve a constrained variational inference problem that can be decomposed. The optimization algorithm ensures consistency among agents and their measurements and is systematically evaluated in both simulation and real-time datasets.