Unique Presentation Identifier:
P23
Program Type
Undergraduate
Faculty Advisor
Dr. Afsana Ahamed, Mr. Stanton Apple
Document Type
Poster
Location
Face-to-face
Start Date
29-4-2025 9:30 AM
Abstract
Title: Semi-Autonomous Trash Robot
Authors: Benjamin Leon, Ethan Mashburn, Connor Olivera, Dillon Wood
Abstract: Environmental pollution is becoming more worrying every day as megatons of human waste persist in vast amounts in our ecosystems. The existing waste collection methodologies, with the labor, time, and cost components combined, make the process cumbersome. A potential solution could be an automated waste collection robot that incorporates AI into a cheap system that autonomously detects trash. This report proposes a machine that may be manufactured to adequate standards in terms of size and cost and deployed in various environments, with the ultimate goal of operating in a fully autonomous mode. Our project iteration outlines an AI model that recognizes 18 classes of litter in the environment, and its integration onto Raspberry Pi 4 to reduce the cost of the robot. The Ultralytics library is used to train a baseline YOLOv11 model on the TACO dataset, which contains more than 4,000 images covering various trash categories with diverse backgrounds such as rocky patches, grass, asphalt roads, etc. Subsequently, transfer learning is used to reduce the image set to a custom 150 images over several training sessions in order to further enhance the accuracy of the model with respect to the campus wastage of Arkansas Tech University. With an average inference time of 10 milliseconds, the trained model demonstrates near real-time performance at 80% classification accuracy over 18 classes of litter. All necessary components, including motors, arms, chassis, and a processor, have been acquired to create a fully functional semi-autonomous robot. The next iterations of the project should incorporate navigation algorithms, charging stations, wireless transmission of summarized data results, a mechanism to empty a full load into a receptacle, and an enhanced battery management system to eliminate human-induced pollution successfully and autonomously.
Key Words: Artificial Intelligence (AI), Machine Learning, YOLOv11, Transfer Learning, Raspberry Pi 4, Classification Accuracy
Recommended Citation
Leon, Ben; Mashburn, Ethan; Olivera, Connor; and Wood, Dillon Michael, "Semi-Autonomous Trash Bot" (2025). ATU Student Research Symposium. 27.
https://orc.library.atu.edu/atu_rs/2025/2025/27
Included in
Applied Mechanics Commons, Computer-Aided Engineering and Design Commons, Electrical and Electronics Commons, Other Electrical and Computer Engineering Commons
Semi-Autonomous Trash Bot
Face-to-face
Title: Semi-Autonomous Trash Robot
Authors: Benjamin Leon, Ethan Mashburn, Connor Olivera, Dillon Wood
Abstract: Environmental pollution is becoming more worrying every day as megatons of human waste persist in vast amounts in our ecosystems. The existing waste collection methodologies, with the labor, time, and cost components combined, make the process cumbersome. A potential solution could be an automated waste collection robot that incorporates AI into a cheap system that autonomously detects trash. This report proposes a machine that may be manufactured to adequate standards in terms of size and cost and deployed in various environments, with the ultimate goal of operating in a fully autonomous mode. Our project iteration outlines an AI model that recognizes 18 classes of litter in the environment, and its integration onto Raspberry Pi 4 to reduce the cost of the robot. The Ultralytics library is used to train a baseline YOLOv11 model on the TACO dataset, which contains more than 4,000 images covering various trash categories with diverse backgrounds such as rocky patches, grass, asphalt roads, etc. Subsequently, transfer learning is used to reduce the image set to a custom 150 images over several training sessions in order to further enhance the accuracy of the model with respect to the campus wastage of Arkansas Tech University. With an average inference time of 10 milliseconds, the trained model demonstrates near real-time performance at 80% classification accuracy over 18 classes of litter. All necessary components, including motors, arms, chassis, and a processor, have been acquired to create a fully functional semi-autonomous robot. The next iterations of the project should incorporate navigation algorithms, charging stations, wireless transmission of summarized data results, a mechanism to empty a full load into a receptacle, and an enhanced battery management system to eliminate human-induced pollution successfully and autonomously.
Key Words: Artificial Intelligence (AI), Machine Learning, YOLOv11, Transfer Learning, Raspberry Pi 4, Classification Accuracy