Soft continuum arms for agricultural manipulation: model based control, grasp state classification, and visual servoing

Agricultural robotics is a growing field due to challenges such as a growing labor shortage and changing agricultural practices. This dissertation seeks to explore and expand the use of Soft Continuum Arms (SCA) in agricultural manipulation tasks. The SCA is a type of soft robotic actuator which is both highly dexterous and compliant. These features make it a good fit for working in the complex, unstructured environment of the agricultural field without damaging the crops. However, SCA are also challenging to model and control which makes their implementation difficult. This work explores fundamental problems related to making SCA effective at agricultural manipulation. It offers insight into the entire grasping process. It begins with methods to identify and reach targets in unstructured environments, moves to controlling the SCA, and finally to grasping the target fruit and separating it from the plant. SCA have infinite degrees of freedom and to effectively use them requires a simpler, reduced order model. This work develops the Constant Curvature and Torsion model which uses a constant strain assumption to describe the deformation of the SCA. This model is used to develop task space feedback control of an SCA. The importance of feedforward in achieving good results is developed. The control method is extended to work on hybrid systems that combine an SCA with rigid degrees of freedom to expand their workspace. Additionally, a dynamic model of an SCA is developed. Additionally, this dissertation explores how to identify and move the SCA toward a fruit target using visual servoing. Using the Constant Curvature and Torsion model, a simulation was developed to train a Reinforcement Learning policy to servo an SCA to view a target fruit with a tip mounted camera. Because the model works in the strain-based configuration space, it was possible to achieve a zero-shot sim-to-real transfer of the policy and strong results on real hardware. To efficiently harvest fruit in a cluttered and occluded environment requires an understanding of the current state of the grasp - including not only if slip is occurring, but also if the grasp has been lost or the fruit is successfully picked. SCA, with their small payloads, require special consideration on what sensors are effective at providing this feedback. This work explores combinations of lightweight, low-cost sensors and their ability to classify the current grasp state. It additionally explores different machine learning methods and their ability to perform this classification.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Walt, Benjamin Thomas

Contributors dc:contributor

• Krishnan, Girish
• Mehta, Prashant G
• Chowdhary, Girish
• Yuan, Wenzhen
• Xu, Siyi

• Soft Continuum Arms
• Model Based Control
• Agricultural Grasping
• Agricultural Robotics
• Visual Servoing With Soft Continuum Arms
• Soft Continuum Arm Dynamics

Statement dc:rights

• Copyright 2025 Benjamin Thomas Walt

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/130091

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F130091