Multispectral AI transforms plastic waste recycling and sorting

Plastic is one of Singapore's top four waste streams, yet only 5% was recycled in 2023 — the lowest rate of any category. The bottleneck is sorting: recyclables go into blue bins, are trucked to materials recovery facilities, then manually sorted and baled for export. Manual sorting is slow, error-prone and a health risk for workers.

The team replaces that step with multispectral AI. By reading reflectance patterns across wavelengths, the system identifies plastic types in real time as items move on the conveyor. Self-supervised learning cuts labelled-data requirements by 90% while outperforming existing state-of-the-art classifiers. The model also separates clean bottles from contaminated ones, so only recyclable material continues down the line. The full stack runs inside a robotic platform built for real environments, reducing manpower cost and making large-scale plastic recycling viable. The team is now working with industry partners and government agencies to expand deployment.

Caption language: en · Fetched: 2026-05-02

Plastic waste is one of the top four waste streams generated in Singapore. Yet, it had the lowest recycling rate, only 5% in 2023. A critical step in the recycling chain is the sorting of plastics. In the current system, recyclables are first disposed of in blue bins and transported to materials recovery facilities. There, materials are manually sorted and bailed for shipment overseas. However, manual plastic sorting remains inefficient. It is slow, labor intensive, errorprone, and poses health risks to workers. Introducing a novel approach that uses multisspectral AI technology to transform the plastic sorting process. By capturing reflectance patterns across different wavelengths, the system enables accurate identification of plastic types, even as materials move along a conveyor belt.

Our self-supervised learning technology achieves high classification performance as compared to existing state-of-the-art while reducing the need for labeled data by 90%. To further improve sorting quality, the AI system is also capable of distinguishing between clean and contaminated bottles, ensuring that only recyclable materials continue down the processing line. These capabilities are integrated into a robotic platform designed for seamless operation in real world environments. In practice, the system automatically identifies and sorts plastic waste in real time, minimizing the need for human intervention while maintaining high precision. Our proposed multisspectral AI technologies can automate passive weight sorting, reducing reliance on manual sorting.

This technologies decreases manpower cost and enables large scale plastic recycling. We'll continue to scale up the system and cooperate with industrial partners and government agencies to transform plastic recycling.