Robust AMR Localization in Dynamic Industrial Environments

Autonomous Mobile Robots (AMRs) in warehouses commonly localize with Adaptive Monte Carlo Localization (AMCL) against a static Occupancy Grid Map recorded when a site is commissioned. In busy areas, pallets and goods placed on the floor hide the mapped walls, and the robot can lose track of its position and stop until an operator intervenes. This thesis, carried out in collaboration with Robotize, starts from such lost robot events and attributes them not to the localization filter but to the map it is given. The goal is to keep the Occupancy Grid Map up to date while the robot operates, using only the onboard 2D LIDAR and no external infrastructure. Detected changes are admitted into the map only when the pose estimate can be trusted and the change has been confirmed over time, so that pallets enter the map while people and vehicles do not. Since a 2D LIDAR only observes the face of an object turned towards it, the method also completes the footprints of box shaped objects, which prevents a partially mapped object from misleading the localization later on. Inferred cells are marked as estimated rather than measured, and are removed again if later observations contradict them. The method is evaluated in a simulated warehouse over 100 scenarios with an exact ground truth. It captures nearly all real changes, including object faces the sensor never observed, runs well within the sensor period, and produces maps that are safe to localize against. The same code was then run unchanged on a public real-world warehouse dataset and on a Robotize AMR. In a corridor lined with unmapped pallets, AMCL lost its position on the original site map and followed the corridor on the map built by the method, and the robot completed a drive that had previously left it lost.