Recipient Organization
L5 AUTOMATION INC.
4802 LA CANADA BLVD
LA CANADA FLINTRIDGE,CA 91011
Performing Department
(N/A)
Non Technical Summary
US strawberry growers facea worsening labor shortage that is threating to become a crisis. Harvesting accounts for39% of grower operating expensesand while automation is an obvious solution, it has proven technically challenging.L5 Automation is working to change that. L5's proposed automated in-field strawberry characteristics assessor addresses USDA SBIR Topic Area 8.13 Plant Production and Protection, specifically improved harvesting of specialty crops and development of cyber-physical systems to support sustainable agriculture. This research and development lays the foundation for a key component of L5's automated strawberry harvester, a system that can alleviate the labor shortages currently hindering the industry. Strawberries are a major U.S. specialty crop and mechanized harvesting will be requiredto keep strawberry production viable in the United States.A drop-in automated solution forstrawberry harvesting must perform the three key functions currently performed by human harvest laborers: berries must be picked from the plants, inspected, and finally packed into retail clamshells or placed intrays for jucing or discard. This project focuses on the inspection function and aims to developthe ability for a mechanized system to rapidly and accurately assess four characteristics of strawberries: ripeness, damage, shape and size. Based on this assessment,a pick / no-pick decision can be made for berries in the bed and pickedberries can be categorized for retail sale, juicing or discard. This research aims to implement algorithms to estimate the four characteristics strictly from images using a mix of classical, deep learning and geometric vision techniques. A custom dataset of at least several hundred strawberries will be gathered from an operational farm, with characteristics labeled via human input and 3-D scans. Iterations of a field-ready assessor will be prototyped to determine best practices and estimate operational throughput. The anticipated result is a field operable cyber-physical system which assesses characteristics and meshes with the workflow of L5's automated harvester. If successful, this technology will ensure that harvest quality from a mechanized platform matches or exceedsthe status quo, unlocking the $1.8B market for harvest labor.
Animal Health Component
80%
Research Effort Categories
Basic
5%
Applied
80%
Developmental
15%
Goals / Objectives
The major goal of this SBIRproject is to research,design and prototype a field operable strawberry assessment capability that canestimate strawberry ripeness, damage, shape and size from several camera views of the berries. This goal is in support ofL5 Automation's largeraimto develop an autonomous strawberry harvesting solution that is a drop-in substitue for human hand harvestlaborers. Such a solution must perform the same tasks that humans do today which includes picking the berries, inspecting and sorting the picked berries according to their quality, and finally packing the retail grade berries into clamshells and depositing other grades of berries into an appropriate bin. The technology and knowledge gained in the develompent of the proposedassessment system will aid all three operations required of the harvester: Pick, Inspect and Pack.To reach this goal L5 has identifiedtechnical objectives for Phase I of the SBIR:1. Create a labeled dataset of harvested strawberries which will support the implementation ofalgorithms to assess required strawberry characteristics. This will include recording images ofall conditions of berries found in the fieldand precisely measuring and labeling their characteristics. The exact size of the dataset may vary, but a sufficient dataset will enable machine learning algorithms to be trained to an accuracy that meets or exceeds that of a human inspector.2. Design and build a prototype portable automated strawberry characteristics assessment system that can categorize strawberries according to grower business requirements and USDA inspection guidelines. This prototype must:be sufficiently portable to be operated in the fieldcapture enough views of the strawberries to enable the trained algorithms to estimate the strawberry's characteristics as or more accurately than a human inspector.With success in Phase I, L5 will be well positioned to extend the work in future Phases to aid in making a pick / no pick decision as well as categorizing the berries for retail sale, juice or discard and finallyprecisely packing the clamshells to the correct weight. Furthermore, the system can be optimized for performance andbe integrated into the larger autonomous harvesting solution.
Project Methods
The methods employed are expected to follow a fairly typical machine learning algorithm development, training, and prototyping cycle.Effortscollect and label strawberry datasetsplit the dataset into train and test segments, being careful to avoid cross contamination between themtrain machine learning algorithmsbased on collected dataset (with expert ground truth labels)iterate the configuration of the algorithms to improve performanceexpand the dataset especially to compensate for discovered gaps or weaknesses in the trained algorithmsiteratively adjust the hardware and softwareconfiguration of the prototype assessment system to improve performanceEvaluationcompare accuracy of assessments on the training dataset segment vs. the test dataset segment to ensure training is successfulmeasure expert human performance on test set and compareto algorithm performanceonce a trained algorithm performs comparably to an expert human on the test set, measure the system performance in the field using new strawberries and the prototype assessment system compared with expert assessment