From Place Recognition to Multi-Robot SLAM Based on Imaging Radar

From Place Recognition to Multi-Robot SLAM Based on Imaging Radar

I started with a narrow question — can a robot recognize where it is from a single radar scan? — and ended up building a multi-robot SLAM system around the answer. This post walks through that path, what changed at each step, and what the sensor kept teaching me along the way.

Hogyun Kim Oct 20, 2026

TL;DR

Place recognition on radar works best when you stop trusting the bright returns. ReFeree describes a scan by its free space as much as by its features.
Global localization needs a pose, not just a match. RadLoc turns retrieval into a full 3-DoF localization module that is fast enough to run inside SLAM.
Multi-robot SLAM changes what “correct” means: every inter-robot loop is expensive to send and hard to undo. The bugs that hurt most were sign and direction conventions, not algorithms.
Radar measures Doppler velocity for free. Putting it into the back-end turned out to keep the problem in the same certifiable form as SE(2) pose-graph optimization.