Online, Scalable and Decentralized Smoothing and Mapping

Many applications for field robots can benefit from large numbers of robots, especially applications where robots are trying to cover or explore a region. A key enabling technology for robust autonomy in these teams of small and cheap robots is the development of collaborative perception to account for the shortcomings of the small and cheap sensors on the robots.

In this seminar, I present DDF-SAM to address the decentralized data fusion (DDF) problem with a decentralized inference based on the smoothing and mapping (SAM) approach to single-robot mapping that is online, scalable and consistent while supporting a variety of sensing modalities. The DDF-SAM approach performs fully decentralized simultaneous localization and mapping through a process in which robots choose a relevant subset of variables from their local map to share with neighbors. Each robot summarizes their local map to yield a density on exactly this chosen set of variables, and then distributes this summarized map to neighboring robots, which then further distribute the summarized map throughout the network. Each robot fuses summarized maps it receives to extend its local sensor horizon.

I will present two primary variations on DDF-SAM, one that uses a batch nonlinear constrained optimization procedure to share maps, DDF-SAM 1.0, and one that uses an incremental solving approach for substantially faster performance, DDF-SAM 2.0. I will demonstrate results from these systems using a combination of real-world and simulated experiments. In addition, I evaluate design tradeoffs for operations within DDF-SAM, with a focus on efficient approximate map summarization.