Visual SLAM (simultaneous localization and mapping) solves the problem of knowing where the robot is relative to some fixed pose. Robots can tell where they are relative to where they started through use of odometry such as kinematic odometry (for example, ticks from wheel encoders), visual odometry (tracking visual features as you move relative to them), or integrating IMU readings. Odometry inevitable begins to drift, however, because small errors when estimating motion between any two moments accumulate.
By building a graph of landmarks (distinct identifiable visual features on a map) from a stereo camera, Visual SLAM can recognize when it has visiting the same location again (closing the loop). This allows for an accurate localization to correct for any accumulating drift.