Dr-LiSA Aligns Radar and Lidar Directly for 3D Robot Localization

A team of robotics researchers has published Dr-LiSA, a new technique for localizing robots and vehicles by combining two very different sensors: spinning radar and lidar. According to the paper posted on arXiv, Dr-LiSA is the first direct method capable of estimating a full 3D pose — known as SE(3), covering both position and orientation in three dimensions — for 2D spinning radar readings measured against a 3D lidar map.

Radar and lidar each have strengths that make combining them appealing for autonomous navigation. Radar continues to work reliably in rain, fog, and other harsh weather, while lidar produces much more detailed 3D maps under good conditions. The catch, the researchers note, is that the two sensors observe the world so differently that aligning their data has been difficult. Until now, radar-lidar localization systems were limited to flat, planar SE(2) alignment and typically trailed the accuracy of systems that pair matching sensor types, such as lidar-to-lidar or radar-to-radar setups.

Dr-LiSA tackles this modality gap with a learned forward model that predicts what a radar scan should look like from a given lidar submap and candidate pose. This predicted radar image is then directly compared, pixel by pixel, against the actual radar measurement — a technique the authors call photometric alignment — allowing the system to refine a full SE(3) pose estimate rather than being confined to a flat plane.

In testing across over 90 km of on-road driving, Dr-LiSA outperformed earlier radar-lidar localization methods on planar accuracy and reached results competitive with state-of-the-art radar-radar systems. The paper is currently under review and spans 8 pages with 6 figures.