Taught robots to notice when their landmarks move, accepted at an IEEE conference

Robots often navigate using fixed visual markers, and when someone moves one, the robot's map quietly breaks. Viba built a system that continuously checks each landmark's health, using a Hidden Markov Model and a graph neural network to tell a moved marker apart from one that is merely blocked or blurry. When a move is confirmed, the map is repaired locally, and every repair must pass a safety gate before it is accepted. In simulation it kept trajectory error to 0.082 m while running in under 30% of the time of a full re-optimization.