Pose-graph optimization is how modern SLAM builds a consistent map: treat every robot pose as a node and every measurement as a constraint, then solve one big least-squares problem that best satisfies them all — including loop closures.
Pose-graph optimization treats each place the robot was as a dot and each movement or recognition as a rubber band between dots. It then relaxes all the rubber bands at once so every constraint is satisfied as well as possible — snapping the map into shape.
Instead of tracking one running estimate like EKF-SLAM, modern SLAM keeps the whole trajectory and solves for it all at once. The tool is pose-graph optimization.
The graph
Nodes are the robot's poses over time (and sometimes landmarks).
Edges are constraints: "odometry says pose 5 was 1.2 m ahead of pose 4," or a loop closure saying "pose 40 is actually the same place as pose 2."
Every edge carries an uncertainty. The optimizer finds the set of poses that best satisfies all constraints simultaneously — minimizing the total weighted disagreement:
Odometry links consecutive poses; a loop closure links a pose to a much earlier one. Solving the graph distributes the accumulated drift across the whole loop.
Why it scales
The trick is sparsity: each pose only connects to a few others, so the matrix the solver handles is mostly zeros. Gauss-Newton or Levenberg-Marquardt exploit that, letting graphs with hundreds of thousands of poses be optimized in real time — impossible for EKF-SLAM's dense O(N²) covariance. Libraries like g2o, GTSAM, and Ceres do the heavy lifting.
The loop-closure payoff
The dramatic moment in any SLAM run is a loop closure: recognizing a previously-visited spot adds one edge linking now to long ago. Re-solving the graph then spreads the accumulated drift backward across every pose in the loop — the map visibly snaps into alignment. This global correction is exactly what a filter can't do cleanly and what makes graph SLAM the standard.
Why it matters
Pose-graph optimization is the backbone of essentially every production SLAM system today — Cartographer, ORB-SLAM, slam_toolbox. It reframes mapping as a clean optimization problem, which is why it scales and why it corrects globally.
