Mixed Rubble, No Way to Sort It

Construction and demolition (C&D) waste exceeds 3.5 billion tonnes per year globally. When source-separated on site, recycling rates reach 75–90%. But mixed C&D debris — the dominant real-world condition — achieves only 50–70% recovery, with contamination from dust, moisture, adhesives, and material composites rendering many streams unrecyclable. Current sorting technology (optical sensors, NIR spectroscopy, robotic arms with CV) degrades severely in the harsh conditions of actual demolition sites and material recovery facilities: high dust, variable lighting, wet conditions, and sensor fouling.

C&D waste represents the largest waste stream by mass in most countries (>35% of total waste in the EU). The circular economy for construction cannot function without high-quality material recovery from mixed debris. Landfilling C&D waste wastes embodied energy and virgin materials. EU regulations now mandate 70% C&D recovery, but mixed-debris processing cannot reliably meet this target.

Optical sorting (TOMRA, BHS) separates bulk material types (wood, metal, concrete) at high throughput but fails on composite materials, contaminated fractions, and fine debris. Robotic sorting with computer vision (ZenRobotics) achieves high accuracy on individual material types but operates at pick rates too slow for C&D waste volumes. Manual sorting remains the norm for complex fractions — labor-intensive, costly, and hazardous. Deep learning models for waste classification (trained on benchmark datasets) degrade sharply on real debris because the scarcity of large-scale labeled C&D datasets prevents generalization across diverse building types, regions, and demolition methods. A 2025 benchmark dataset covers only limited materials and conditions.

Open, large-scale labeled datasets of C&D waste captured in real operating conditions (variable lighting, dust, moisture) across multiple facility types and regions. Sensor fusion approaches combining NIR spectroscopy, RGB imaging, and 3D depth sensing that are robust to environmental degradation. On-site pre-sorting systems at the point of demolition using real-time classification — moving intelligence upstream before materials are mixed.

A team could build and benchmark a CV classifier for C&D materials using real mixed-debris imagery (collected from a local demolition site or MRF), comparing performance against models trained only on clean laboratory images. Alternatively, a team could prototype a sensor-fusion sorting station combining low-cost NIR and RGB sensors with ML classification, testing on actual mixed debris samples. Computer vision, robotics, and waste management skills apply.