Dr-BA: First direct radar bundle adjustment for all-weather autonomous navigation

Dr-BA, developed by Daniil Lisus, Cedric Le Gentil, and Timothy D. Barfoot at the University of Toronto, is a groundbreaking direct radar bundle adjustment framework. Unlike traditional methods that first extract sparse point clouds from radar scans and then apply point cloud alignment, Dr-BA works directly on full 2D spinning radar intensity images. This allows the system to use all radar returns from multiple scans simultaneously for dense mapping and pose estimation. The key innovation is a separable optimization that elegantly decouples pose estimation from mapping, making the problem computationally efficient and generalizable. The same formulation naturally extends to direct radar-only localization within an existing map, eliminating the need for other sensors.

The method was validated on over 200 kilometers of on-road data from five distinct routes, demonstrating state-of-the-art performance in both radar-based bundle adjustment and cross-session localization. This is a major step toward truly all-weather autonomous driving, as radar is largely unaffected by precipitation that cripples cameras and lidar. The implementation is publicly available on GitHub, allowing the robotics community to replicate and build upon the work. Accepted at RSS 2026, Dr-BA solves a long-standing challenge of precise, direct radar SLAM without intermediate point cloud extraction.