Visual SLAM at Scale

A robot drives around a fixed loop, using a simulated camera to spot colour-coded landmark clusters and fold them into a growing feature map while it simultaneously estimates its own position from noisy step-by-step odometry alone. Because each tiny per-step error compounds into the next, the estimated trajectory and map slowly bow away from the true path — until the robot's odometer wraps back past the landmarks it mapped at the start of the loop. That revisit triggers a pose-graph loop closure: a single correction is distributed backward across the entire stored trajectory and map, snapping the estimate back into close alignment with the ground truth. Tune drive speed, odometry noise and sensor range, and watch the live drift readout plunge at every closure.