ABC: Adaptive, Biomimetic, Configurable Robots for Smart Farms - From Cereal Phenotyping to Soft Fruit Harvesting

Currently, numerous factors, such as demographics, migration patterns, and economics, are leading to the critical labour shortage in low-skilled and physically demanding parts of agriculture. Thus, robotics can be developed for the agricultural sector to address these shortages. This study aims to develop an adaptive, biomimetic, and configurable modular robotics architecture that can be applied to multiple tasks (e.g., phenotyping, cutting, and picking), various crop varieties (e.g., wheat, strawberry, and tomato) and growing conditions. These robotic solutions cover the entire perception–action–decision-making loop targeting the phenotyping of cereals and harvesting fruits in a natural environment. The primary contributions of this thesis are as follows. a) A high-throughput method for imaging field-grown wheat in three dimensions, along with an accompanying unsupervised measuring method for obtaining individual wheat spike data are presented. The unsupervised method analyses the 3D point cloud of each trial plot, containing hundreds of wheat spikes, and calculates the average size of the wheat spike and total spike volume per plot. Experimental results reveal that the proposed algorithm can effectively identify spikes from wheat crops and individual spikes. b) Unlike cereal, soft fruit is typically harvested by manual selection and picking. To enable robotic harvesting, the initial perception system uses conditional generative adversarial networks to identify ripe fruits using synthetic data. To determine whether the strawberry is surrounded by obstacles, a cluster complexity-based perception system is further developed to classify the harvesting complexity of ripe strawberries. c) Once the harvest-ready fruit is localised using point cloud data generated by a stereo camera, the platform’s action system can coordinate the arm to reach/cut the stem using the passive motion paradigm framework, as inspired by studies on neural control of movement in the brain. Results from field trials for strawberry detection, reaching/cutting the stem of the fruit with a mean error of less than 3 mm, and extension to analysing complex canopy structures/bimanual coordination (searching/picking) are presented. Although this thesis focuses on strawberry harvesting, ongoing research is heading toward adapting the architecture to other crops. The agricultural food industry remains a labour-intensive sector with a low margin, and cost- and time-efficiency business model. The concepts presented herein can serve as a reference for future agricultural robots that are adaptive, biomimetic, and configurable.