GLIM

GPU-accelerated 3D lidar SLAM framework for any range sensor, from AIST

License

MIT

Overview

GLIM is a 3D point-cloud-based localization and mapping framework by Kenji Koide (AIST, Japan). It combines a tightly-coupled lidar-IMU odometry front-end with a global-optimization back-end based on global registration error minimization, runs its heaviest kernels on GPU, and deliberately supports 'any range sensor' — mechanical spinning lidars, non-repetitive solid-state units like Livox, and even RGB-D cameras — under one extensible, ROS 2-native architecture.

GLIM splits SLAM into an odometry estimation front-end and a global mapping back-end, both formulated as factor-graph optimization on GTSAM. The front-end tightly couples lidar and IMU with keyframe-based fixed-lag smoothing, staying robust through aggressive motion; the back-end performs global trajectory optimization with GPU-accelerated voxelized GICP matching cost factors, closing loops via a global registration error minimization approach instead of fragile explicit loop detection. Everything beyond the core is a plugin: odometry variants (CPU, GPU, IMU-centric), loop-closure modules, and extension modules from the companion glim_ext repository. The interactive viewer is built on the author's Iridescence library, ROS 2 and ROS 1 wrappers are provided, and the same pipeline handles Velodyne/Ouster spinning lidars, Livox solid-state units and depth cameras — a breadth few open SLAM systems attempt.

Why it matters

GLIM is the most complete actively-developed successor to the classic open lidar-SLAM stacks: research-grade global optimization and GPU acceleration in an engineered, documented, plugin-extensible package from the author of hdl_graph_slam and small_gicp, whose libraries already underpin much of the open lidar-mapping ecosystem.

Problem it solves

Building accurate 3D maps with modern sensors — solid-state lidars, aggressive platform motion, long trajectories — breaks the assumptions of older lidar SLAM systems. GLIM's tightly-coupled IMU fusion, GPU-accelerated global optimization and sensor-agnostic design handle these cases in one framework.

Key takeaways

• Tightly-coupled lidar-IMU odometry robust to aggressive motion and sparse geometry
• Global optimization back-end minimizes registration error over the whole map (no fragile loop detector)
• CUDA-accelerated voxelized GICP factors make dense global optimization tractable in real time
• Sensor-agnostic: spinning lidars, Livox solid-state units and RGB-D cameras
• Plugin architecture (glim_ext) and ROS 2/ROS 1 wrappers for clean extension and integration

Who it is for

• Robotics engineers building 3D mapping pipelines
• SLAM researchers extending factor-graph systems
• AMR and outdoor-robot developers
• Surveying and digital-twin practitioners using mobile lidar

Limitations

• Best performance requires an NVIDIA GPU (CPU fallbacks are slower)
• IMU required for the standard tightly-coupled pipeline
• Global optimization memory grows with map size on very large-scale runs