Development of a multicell capacitive tactile sensing system for stable robotic fruit grasping

Autonomous fruit harvesting remains challenging because vision-only robotic systems cannot evaluate contact quality during grasping. This thesis develops a vision–tactile framework for robotic apple harvesting using a YOLOv8 RGB-D perception module and an origami-inspired multicell capacitive tactile sensor integrated into three robotic fingertips. A center of pressure (CoP)-aware offline labeling method was used to transform tactile responses into three grasp-quality classes: Class 0, stable; Class 1, moderate; and Class 2, failed. During deployment, a Random Forest classifier predicts grasp quality from a 12-dimensional normalized capacitance vector, avoiding online CoP computation and enabling low-latency inference. The classifier achieved 98.38% accuracy on the held-out tactile dataset. An independent binary fail/non-fail validation achieved 86.7% accuracy across 15 physical trials. In controlled mock-orchard harvesting trials, the tactile-enabled system reduced fruit damage from 34.4% to 3.1%, corresponding to a 31.3 percentage-point absolute reduction. These results show that CoP-informed tactile sensing can improve damage-aware robotic fruit harvesting, although field validation under real orchard conditions remains necessary.