Skip to content

Home

Introduction

GLIM

GLIM is a versatile and extensible range-based 3D mapping framework.

  • Accuracy: GLIM is based on direct multi-scan registration error minimization on factor graphs that enables to accurately retain the consistency of mapping results. GPU acceleration is supported to maximize the mapping speed and quality.
  • Easy-to-use: GLIM offers an interactive map correction interface that enables the user to manually correct mapping failures and easily refine mapping results.
  • Versatility: As we eliminated sensor-specific processes, GLIM can be applied to any kind of range sensors including:
    • Spinning-type LiDAR (e.g., Velodyne HDL32e)
    • Non-repetitive scan LiDAR (e.g., Livox Avia)
    • Solid-state LiDAR (e.g., Intel Realsense L515)
    • RGB-D camera (e.g., Microsoft Azure Kinect)
  • Extensibility: GLIM provides the global callback slot mechanism that allows to access the internal states of the mapping process and insert additional constraints to the factor graph. We also release glim_ext that offers example implementations of several extension functions (e.g., explicit loop detection, LiDAR-Visual-Inertial odometry estimation).

Tested on Ubuntu 22.04 / 24.04 with CUDA 12.2, and NVIDIA Jetson Orin.

GLIL CPU Reproduction Fork

This fork adds CPU-focused reproduction configs and validation notes for local LiDAR odometry experiments. The 2026-04 manifest-verified bundle records 3/3 MegaParticles APE PASS plus a clean official Ouster sample smoke check.

Reproduction scorecard

dataset kind status RMSE playback note
indoor_easy_01 APE PASS 1.019250 1.000x Track B+C PASS
outdoor_hard_01a APE PASS 0.906313 1.000x 5/5 byte-identical hard recipe
outdoor_kidnap_a APE PASS 20.349845 1.000x Track B+C PASS
os1_128_01_downsampled smoke WARN NA 0.201x clean Ouster smoke, no GT APE

Track B is upstream GLIM RMSE + 20%. Track C is playback mean >= 0.95x. The Ouster sample is a completion/stability smoke check because this workspace does not include a matching ground-truth APE file. See the Reproduction scoreboard for the shareable summary and recommended configs.

Build GitHub stars License: MIT

Video

Robustness test

Mapping with various range sensors

Outdoor driving test with Livox MID360

See more in Extension modules and Demo pages.

Contact

Kenji Koide
National Institute of Advanced Industrial Science and Technology (AIST), Japan