Data Sharing and Assimilation in Multi-Robot Systems for Environment Mapping

Abstract

We consider scenarios where a mobile multi-robot system is used for mapping a spatial field. Gaussian processes are a widely employed regression model for this type of tasks. For the sake of generality, scalability, and robustness, we assume that planning and control are fully distributed and that robots can only communicate via range-limited channels. In such scenarios, one core challenge is how to let the robots efficiently coordinate in order to maintain a shared view of the mapping process, and, accordingly, make plans minimizing overlaps and optimizing joint information gain from obtained measurements. A simple approach of sharing and utilizing all the sampled data would not scale to large teams, neither for computation nor for communication (assuming a general ad hoc robot network). Building on previous work where robots adaptively plan where to sample data by selecting convex containment regions, we propose a data sharing and assimilation strategy which aims to minimize the impact on communication and computation while minimizing the loss on accuracy in map estimation. The strategy exploits convexity of the regions to create compact meta-data that are locally shared. Submodularity of information processes and properties of GPs are used by the robots to create highly informative summaries of the sampled regions, that are shared on-demand based on the meta-data. In turn, a received summary is assimilated by a robot into its local GP only if/when needed. We perform a number of studies in simulation using real data from bathymetric maps to show the efficacy of the strategy for supporting scalability of computations and communications while guaranteeing learning accurate maps.