pose-graph optimization (graph SLAM)

Pose-graph optimization is a way of finding a robot's most likely path by treating the whole journey as a network of connected points and gently nudging every point until all the connections are as happy as they can be. Each 'pose' is one snapshot of where the robot was and which way it faced at some moment, drawn as a dot. Between two dots the robot draws an edge that records a constraint: a measured belief like 'I moved roughly two meters forward and turned a bit to the left to get from here to there.' String many poses together and you get a graph — a chain of dots joined by these relative-motion edges.

Most edges link a pose to the one just after it, the ordinary step-by-step record of driving. But the special, powerful edges are the ones added by loop closure: 'this much-later pose is actually back at the same place as this much-earlier pose.' Now the graph is over-determined — there are more constraints than there is slack — and they will conflict, because each measurement carried a little error. Optimization is the act of sliding all the poses around at once to the single arrangement that violates the fewest constraints by the least amount, like a tangle of springs settling into the shape where the total tension is lowest. Every edge is a spring, and the answer is the resting pose of the whole web.

This 'graph SLAM' view became popular because it separates a messy problem into two clean jobs: building the graph (deciding which poses connect and how strongly) and then solving it (finding the best layout). The solving step can lean on fast, well-understood math because the graph is mostly local — each pose touches only a few neighbors — so even thousands of poses can be balanced in a fraction of a second.