BIEVR-LIO: Robust LiDAR-Inertial Odometry through Bump-Image-Enhanced Voxel Maps

A team from ETH Zurich and the University of Zurich has published a new paper on arXiv detailing BIEVR-LIO, a novel approach to LiDAR-Inertial Odometry (LIO) designed to solve a critical problem in robotics: navigation in geometrically uninformative environments. Traditional LIO systems, which fuse LiDAR scans with inertial data to track a robot's movement, often fail in places like long corridors, empty warehouses, or tunnels where there are few distinct features for point cloud registration. BIEVR-LIO tackles this by introducing a high-resolution map representation that stores surfaces as compact, voxel-wise oriented height images, which the authors term 'bump-images.' This allows the system to exploit even subtle variations in flat surfaces for registration without calculating intermediate geometric primitives.

Beyond the novel map representation, BIEVR-LIO employs a smart, map-informed point sampling strategy. Instead of processing every point in a LiDAR scan globally, it focuses computational resources on the most geometrically informative regions. This dual innovation leads to two major benefits: substantially improved robustness in challenging, feature-sparse scenarios where other methods diverge, and a reduction in computational cost compared to brute-force high-resolution sampling. The paper demonstrates state-of-the-art performance across multiple sensors and platforms. Furthermore, the fine-grained geometry captured by the system is shown to be directly useful for downstream tasks, such as creating detailed elevation maps to aid in legged robot locomotion over uneven terrain.