RTDRNet-lite: A lightweight real-time detection framework for robotic waste sorting.

Area of Science:

• Environmental Science and Engineering
• Artificial Intelligence
• Robotics

Background:

• Urbanization presents significant waste management challenges globally.
• Current waste recycling systems often lack efficiency and automation.
• Intelligent sorting is crucial for effective waste management and resource recovery.

Purpose of the Study:

• To develop a comprehensive, automated waste management framework for intelligent, real-time waste sorting.
• To integrate AI-based detection with robotic hardware for enhanced waste processing.
• To address the limitations of existing waste recycling technologies.

Main Methods:

• Development of the RTDRNet-lite model, a lightweight variant of RT-DETR, achieving 97% mAP@50.
• Hybrid training approach using real-world and Stable Diffusion-generated synthetic waste images.
• Integration with a custom 4-degree-of-freedom robotic arm for live sorting validation.

Main Results:

• RTDRNet-lite model demonstrates high accuracy (97% mAP@50) and reduced computational complexity.
• Hybrid training enhanced model generalizability and handling of complex object boundaries.
• Successful validation of the integrated system in live waste sorting tasks.

Conclusions:

• The proposed AI and robotic system offers a robust and accurate solution for automated waste sorting.
• The framework shows significant potential for deployment in industrial-scale waste management facilities.
• This research advances intelligent waste management towards greater sustainability.