Navigazione autonoma e ispezione visiva intelligente in ambienti industriali

This dissertation examines two independent research directions: autonomous mobile navigation and industrial visual inspection. Each direction is developed as a complete and self-contained system, from problem formulation to implementation and controlled experimental evaluation. The two domains respond to different industrial needs, employ different technical philosophies, and are assessed against different operational criteria. The navigation study pursues a minimalist sensing and control strategy for dependable operation in structured indoor environments, whereas, in a separate research domain, the visual inspection study adopts a high-fidelity deeplearning pipeline for complex component-verification tasks in automotive assembly. The first direction focuses on a minimalist navigation framework intended for robots performing sterilisation-related tasks in corridor-like indoor spaces representative of the pharmaceutical industry. The system uses a single two-dimensional LiDAR together with wheel odometry, avoiding complex sensing configurations that are difficult to maintain in environments with strict operational constraints. A ROSbased navigation pipeline is configured around AMCL localisation, static-map global planning, a rolling local costmap, and a tuned Dynamic Window Approach planner. Parameter refinement covers footprint geometry, obstacle-layer behaviour, inflation radii, update frequencies, and velocity limits. Navigation behaviour is assessed through repeated clockwise and counter-clockwise waypoint cycles in a controlled indoor layout shaped to reflect corridor constraints. The results demonstrate reliable localisation in most scenarios, though systematic lateral drift (up to 290 mm) was observed during clockwise right-hand turns. This indicates that route completion and localisation recovery were achieved, but that waypoint repeatability remained direction-dependent and requires cautious interpretation in the absence of loadedplatform sterilisation trials. Complementary to the navigation task, the second research direction develops a multi-class inspection system for identifying missing components on fuel-tank assemblies. A dedicated dataset of real components is constructed and annotated to capture typical visual variability in such assemblies. YOLOv8 and Faster RCNN models are trained under matched conditions to allow direct comparison. YOLOv8 achieves stronger overall performance in precision, recall, and mAP, and cross-dataset evaluation with the NPU-BOLT benchmark demonstrates that the detector generalises beyond the training setup. Real-time testing in a controlled iinspection environment, including staged removal of components, confirms that the detector can identify missing parts at frame rates compatible with in-line quality control. Taken together, the two research directions illustrate how carefully matched engineering choices, parameter strategies, and experimental validation can yield practical solutions in different industrial contexts. The work provides empirical evidence that deployable performance depends not on enforcing a single design doctrine, but on aligning system architecture with the operational objective of each task.