Enhancing sorting efficiency in cluttered construction and demolition waste streams via boundary-guided grasp...

Area of Science:

• Robotics and Artificial Intelligence
• Waste Management and Recycling
• Computer Vision

Background:

• Robotic automation is crucial for construction and demolition waste (CDW) valorization, but grasp detection for recycling is underexplored.
• Current AI-driven computer vision (CV) excels at CDW recognition but struggles with action-focused tasks like grasp detection in cluttered scenes.
• A significant limitation is the scarcity of grasp-annotated data for training robotic systems.

Purpose of the Study:

• To develop a robust method for robotic grasp detection in CDW.
• To address the challenge of identifying optimal, collision-free gripper poses for CDW recyclables.
• To introduce a novel dataset for training and evaluating CDW instance segmentation and grasp detection models.

Main Methods:

• A boundary-guided grasp detection model using attentional feature fusion is proposed.
• The method leverages advances in shape-aware CDW instance segmentation and LLM-enabled automated labeling.
• The ReCoDeWaste dataset, featuring over 100,000 annotated instances in cluttered scenes, is introduced for RGB-D CDW instance segmentation and grasp detection.

Main Results:

• The boundary-guided model successfully predicts collision-free grasps in cluttered CDW streams.
• The approach outperforms state-of-the-art methods in standard evaluation metrics.
• Up to 94.36% grasp detection accuracy was achieved, demonstrating the model's effectiveness.

Conclusions:

• This research enhances CDW valorization by advancing action-focused CV tasks beyond mere recognition.
• The developed method and dataset facilitate more efficient and accurate robotic sorting of construction and demolition waste.
• Future work can build upon these findings to further automate waste recycling processes.