Advances in automated plant phenotyping and target-driven visual navigation for breeding candidates

Advances in agricultural robotics are transforming field operations by enhancing perception and planning capabilities, particularly in automated plant phenotyping and visual navigation. Leaf angle, a critical factor for solar light absorption and photosynthetic efficiency, directly impacts the growth and yield of row crops such as maize. As such, plant breeders are interested in selecting specific leaf angle traits, which require measurement of leaf angle. However, traditional manual measurement of leaf angle in field settings is nearly impossible due to the significant labor involved. Furthermore, measurement tactics, such as with protractors, do not scale to field settings, and the results have been found to vary widely between different individuals. In this thesis, we present a field robotic system equipped with low-cost sensors to automatically measure the angle of the leaf. The proposed method is capable of producing quantitative and qualitative estimates of leaf angle for numerous maize leaves consistently using an RGB-D camera. Our method provides a high-throughput low-cost tool for the measurement of leaf angle and is expected to enable new agricultural research. Additionally, autonomous navigation in unstructured and dynamic outdoor environments is critical for field robotics applications. For example, the breeding candidate after phenotyping needs to be located and their seeds collected for further breeding. Traditional Deep Reinforcement Learning-based visual navigation techniques face challenges in outdoor settings, particularly in the absence of high-resolution maps and GPS signals. This thesis presents a deep reinforcement learning-based approach for target-driven visual navigation in outdoor settings, using the successor feature (SF) framework to enhance the model's generalization and transfer learning capabilities. Our method constructs a grid-world environment for navigation tasks and employs a goal-conditioned reinforcement learning (GCRL) strategy that leverages SFs to capture environmental dynamics. This approach enables the model to transfer knowledge across tasks, making it adaptable to new environments with zero-shot or few-shot tuning. The experimental results demonstrate the adaptability of our method in new outdoor environments within the same domain. Furthermore, although the model is trained in a discrete grid-world environment, it is successfully deployed in real-time across different seasons within the same area, highlighting its robust transferability across domains and continuous state spaces.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Agricultural & Biological Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Wu, Junzhe

Contributors dc:contributor

• Tran, Huy
• Chowdhary, Girish
• Eveland, Andrea
• Lipka, Alexander

• Leaf Angle Prediction
• Automated Plant Phenotyping
• Target-driven Navigation

Statement dc:rights

• Copyright 2025 Junzhe Wu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130037

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130037