Modelling wine grapevines for autonomous robotic cane pruning
Henry Williams, David A. Smith, Jalil Shahabi, Trevor Gee, Mahla Nejati, Benjamin McGuinness, Kale Black, Jonathan H. Tobias, Rahul Jangali, Shen Hin Lim, Mike Duke, Oliver Bachelor, Josh McCulloch, Richard Green, Mira O’Connor, Sandhiya Gounder, Angella Ndaka, Karly Burch, Jaco Fourie, Jeffrey Hsiao, Armin Werner, R.H. Agnew, R. G. Oliver, Bruce A. MacDonald
Abstract
Aotearoa (New Zealand) has a strong and growing winegrape industry struggling to access workers to complete skilled, seasonal tasks such as pruning. Maintaining high-producing vines requires training agricultural workers that can make quality cane pruning decisions, which can be difficult when workers are not readily available. A novel vision system for an autonomous cane pruning robot is presented that can assess a vine to make quality pruning decisions like an expert. The vision system is designed to generate an accurate digital 3D model of a vine with skeletonised cane structures to estimate key pruning metrics for each cane. The presented approach has been extensively evaluated in a real-world vineyard as a commercial platform would be expected to operate. The system is demonstrated to perform consistently at extracting dimensionally accurate digital models of the vines. Detailed evaluation of the digital models shows that 51.45% of the canes were modelled entirely, with a further 35.51% only missing a single internode connection. The quantified results demonstrate that the robotic platform can generate dimensionally accurate metrics of the canes for future decision-making and automation of pruning.