GTSAM

Factor-graph optimization library powering many modern SLAM backends

License

BSD-2-Clause

Overview

Georgia Tech Smoothing and Mapping — a C++ library with Python bindings for factor-graph based state estimation and sensor fusion. It provides the iSAM2 incremental smoother, a rich set of ready-made factors for IMU preintegration, GPS, vision and range measurements, and the Lie-group machinery underneath. GTSAM is the optimization backend inside many modern SLAM systems, including LIO-SAM, Kimera and GLIM.

GTSAM models estimation problems as factor graphs: variables (poses, landmarks, velocities, biases) connected by factors encoding measurements. It ships batch optimizers (Levenberg-Marquardt, Gauss-Newton, dogleg) and the iSAM2 incremental smoother, IMU preintegration following Forster et al., smart factors for vision, and full Lie-group support for SO(3)/SE(3). Python bindings expose the complete API, and an active BSD-licensed community maintains it.

Why it matters

Factor graphs became the lingua franca of SLAM and multi-sensor fusion, and GTSAM is the reference implementation the field builds on — widely cited and embedded in several of the most-used open SLAM systems.

Problem it solves

Fusing IMU, lidar, camera and GPS measurements into a consistent state estimate is a large sparse optimization problem. GTSAM expresses it as a factor graph and exploits its sparsity, with iSAM2 updating the solution incrementally as new measurements arrive instead of re-solving from scratch.

Key takeaways

• Factor graphs turn multi-sensor fusion into a sparse optimization problem with reusable structure
• iSAM2 gives incremental, real-time smoothing instead of repeated batch solves
• Ready-made IMU, GPS, vision and range factors cover most robotics estimation problems
• The backend of choice for LIO-SAM, Kimera, GLIM and many research systems
• Python bindings make it accessible beyond C++ codebases

Who it is for

• SLAM system builders needing a proven optimization backend
• State-estimation researchers prototyping new factors
• Engineers fusing IMU/GPS/vision in production robots
• Graduate students learning estimation theory hands-on

Limitations

• Steep learning curve — factor-graph modeling and Lie-group concepts are prerequisites for productive use
• Building from source with Python wrappers and TBB can be involved on some platforms