Modern automation challenges: towards seamless AGV localization

Modern warehouses constantly raise in complexity, widely spreading across many industrial fields. High flow efficiency, operation assurance, significant flexibility and pin-point repeatability are compelling in automated factories. Autonomous Guided Vehicles (AGVs) fleets are replacing manual lift trucks, guaranteeing solid warehousing performance and safe behaviours towards people, structures and other moving objects. They are usually equipped with forks to move products rapidly and safely. To correctly handle industrial materials, these vehicles require an estimation of position and orientation with accuracy respectively below 2 cm and 1 deg, especially during pickup and deployment actions. Furthermore, AGVs ought to incorporate versatile localization strategies to be able to navigate in many different industrial contexts. For this reason, each vehicle is equipped with a localization laser rangefinder to accurately perceive the surrounding environment. Inaccurate AGVs positioning, especially between vehicles of the same fleet, or uncertain localization performance may lead to stock damages or people injuries. The first contribution of this dissertation is the proposal of a novel automatic calibration method for four wheel AGVs. Parameter tuning is essential to guarantee that each vehicle of the fleet can reach the same point of the warehouse with adequate precision. Usually, the calibration of AGVs is an iterative manual activity: the operator runs a series of tests and tunes one parameter at a time. This operation requires a large amount of time, can be inaccurate, and strongly depends on the experience of the technician. The proposed calibration method estimates both the kinematic parameters and the laser rangefinder parameters by comparing the sensor egomotion and the motion commands. This calibration procedure is considerably faster, completely automatic, and independent for the ability of the operator, being able to estimate all the parameters at the same time with higher accuracy. This method has been tested on field on many different AGVs of different fleets to assess its correctness and to evaluate its actual performance. The second contribution of this thesis is the development of three different methods for localization in unstructured environments. Usually, the standard AGV localization is performed using the laser rangefinder to detect artificial landmarks, matching them with a static reflector map. The installation of those landmarks is expensive and time-consuming, as well as inflexible w.r.t. environmental changes. The reconstruction of industrial environments is built using a Terrestrial Laser Scanner (TLS) as part of an advanced plant setup workflow. Thereafter, a 2D map can be extracted from the TLS survey with advanced software processing. The localization methods proposed in this dissertation exploit such map either to extract environmental landmarks or to perform dense map registration. With these methods, no artificial landmark has to be deployed, so system installation becomes faster and less expensive. These localization approaches for unstructured environments return better or similar performances w.r.t. the canonical artificial landmark approach. Such methods have been tested on real AGVs in an industrial warehouse to assess their performance. The third contribution of this dissertation is the proposal of a novel signature for place recognition and loop closure with landmark maps, named GRD (Geometric Relation Distribution). Precise location signature is essential to reduce localization errors and inconsistencies during SLAM (Simultaneous Localization And Mapping) procedures and can also be used for Global Localization during AGV pose initialization. The proposed signature is suitable for artificial landmarks maps, like those currently used in industrial localization systems, as well as for feature-based maps. Performance have been evaluated through experiments on standard datasets. The combination of advanced automatic calibration methods, flexible localization strategies and place recognition approaches developed in this thesis can remarkably increase onpoint localization performance, greatly reduce plant installation times, and vastly enhance warehouse efficiency.