Factor Graph Optimization

Factor graph optimization represents unknown poses, biases, landmarks, or object states as variables linked by measurements and priors, turning estimation into sparse nonlinear inference. Common approaches include Gauss–Newton and Levenberg–Marquardt solvers, incremental smoothing with fixed-lag windows, robust M-estimation, and belief propagation; newer work adds learned weighting, distributed optimization, and semidefinite relaxations that can certify global solutions. It underpins SLAM, visual and inertial navigation, mapping, and robot state estimation, and is commonly evaluated on pose-graph, bundle-adjustment, and sensor-fusion benchmarks as well as with field robots. A central challenge is balancing accuracy and robustness to outliers with runtime and memory, while global guarantees and calibrated uncertainty remain difficult to scale beyond local optimization.