Training-Based Robotic Arm for Automated Sorting
University
Shawnee State University
Major
Computer Engineering
Student Type
Undergraduate Student
Presentation Types
Poster Presentation (Live)
Keywords:
Reinforcement Learning, Genesis, Proximal Policy Optimization, Robotic Control
Abstract
Training-Based Robotic Arm for Automated Sorting began in Fall 2025 and is now in Phase 2 of a reinforcement-learning-based system. Phase 1 established the framework for integrating the Pi0 pretrained model with the Proximal Policy Optimization (PPO) algorithm within the Genesis physics simulation environment. A structured reward and penalty system and domain randomization were created to improve generalization and reduce the gap between simulation and a physical robotic arm.
Phase 2 transitions the project from design to implementation. A 3-DOF robotic arm model has been constructed and validated within Genesis to establish a foundation for stable reinforcement learning. Testing has confirmed motor-controlled articulation and target reaching in simulation, providing measurable results prior to reinforcement learning integration. The long-term goal is to validate a trained PPO policy in simulation and deploy it onto a Raspberry Pi-controlled Yahboom robotic arm.
Human and Animal Subjects
no
IRB or IACUC Approval
no
Faculty Mentor Name
Jeong Tae Ok
Faculty Mentor Title
Assistant Professor
Faculty Mentor Department
Engineering Technologies
Recommended Citation
Rowe, Elijah J., "Training-Based Robotic Arm for Automated Sorting" (2026). Celebration of Scholarship. 5.
https://digitalcommons.shawnee.edu/cos/2026/PosterSession/5
Training-Based Robotic Arm for Automated Sorting
Training-Based Robotic Arm for Automated Sorting began in Fall 2025 and is now in Phase 2 of a reinforcement-learning-based system. Phase 1 established the framework for integrating the Pi0 pretrained model with the Proximal Policy Optimization (PPO) algorithm within the Genesis physics simulation environment. A structured reward and penalty system and domain randomization were created to improve generalization and reduce the gap between simulation and a physical robotic arm.
Phase 2 transitions the project from design to implementation. A 3-DOF robotic arm model has been constructed and validated within Genesis to establish a foundation for stable reinforcement learning. Testing has confirmed motor-controlled articulation and target reaching in simulation, providing measurable results prior to reinforcement learning integration. The long-term goal is to validate a trained PPO policy in simulation and deploy it onto a Raspberry Pi-controlled Yahboom robotic arm.