[Paper] A.R.I.S.: Automated Recycling Identification System for E-Waste Classification Using Deep Learning

Source: arXiv - 2602.17642v1

Overview

The paper introduces A.R.I.S. (Automated Recycling Identification System), a low‑cost, portable device that uses deep‑learning vision (YOLOx) to sort shredded electronic waste (e‑waste) into metals, plastics, and circuit boards in real time. By boosting detection accuracy while keeping inference latency low, A.R.I.S. promises to close the material‑recovery gap that plagues current e‑waste recycling streams.

Key Contributions

• End‑to‑end hardware‑software prototype: a compact sorter that can be deployed on the shop floor or in collection centers.
• Real‑time classification with YOLOx: adapts a state‑of‑the‑art object detector to the noisy, fragmented nature of shredded e‑waste.
• High performance metrics: 90 % overall precision, 82.2 % mean average precision (mAP), and 84 % sortation purity on a diverse test set.
• Cost‑effective design: uses off‑the‑shelf cameras and inexpensive compute (e.g., edge GPU/TPU) to keep the system affordable for small‑to‑medium recyclers.
• Open‑source dataset & training pipeline: the authors release annotated images of shredded e‑waste, enabling further research and industry adoption.

Methodology

1. Data collection – The team shredded a variety of consumer electronics (smartphones, laptops, TVs) and captured high‑resolution images of the resulting fragments under controlled lighting.
2. Labeling – Each fragment was manually annotated as metal, plastic, or circuit board.
3. Model selection – YOLOx was chosen for its balance of speed and accuracy. The network was fine‑tuned on the custom dataset using transfer learning from a COCO‑pretrained backbone.
4. System integration – A compact conveyor feeds shredded pieces under the camera. Detected objects trigger pneumatic or mechanical actuators that divert the piece into the appropriate bin.
5. Evaluation – Precision, recall, mAP, and sortation purity were measured by comparing the system’s output bins against ground‑truth material composition.

Results & Findings

• Precision: 90 % of items the system classified were correctly identified.
• Mean Average Precision (mAP): 82.2 % across the three classes, indicating robust detection even with overlapping or partially occluded fragments.
• Sortation purity: 84 % of the material in each output bin belonged to the intended class, a substantial improvement over baseline manual sorting (≈60 %).
• Latency: Inference time per frame stayed under 30 ms on an edge GPU, enabling a throughput of ~200 items/min, suitable for typical recycling line speeds.

These numbers demonstrate that a deep‑learning model can reliably operate in the noisy visual environment of shredded e‑waste without sacrificing speed.

Practical Implications

• Scalable recycling operations – Small recyclers can adopt A.R.I.S. without large capital expenditures, increasing overall material recovery rates across the industry.
• Reduced manual labor & safety risks – Automated visual sorting lessens the need for workers to handle sharp or hazardous fragments.
• Higher-quality feedstock for downstream processes – Cleaner separation improves the economics of downstream metal smelting and plastic re‑processing, making circular‑economy business models more viable.
• Integration with existing infrastructure – The system can be retrofitted onto existing conveyor belts and sorting lines, allowing incremental upgrades rather than full plant overhauls.
• Data‑driven process optimization – Real‑time detection logs can be fed into analytics dashboards to monitor material composition trends, informing procurement and product‑design decisions aimed at easier end‑of‑life handling.

Limitations & Future Work

• Dataset diversity – The current training set covers a limited range of device types and brands; performance may drop on exotic or heavily corroded components.
• Fragment size constraints – Very fine particles (<2 mm) are beyond the camera’s resolution and thus remain unsorted.
• Environmental robustness – Lighting variations and dust accumulation can affect detection accuracy; future versions will explore adaptive illumination and self‑cleaning optics.
• Extended material taxonomy – Adding subclasses (e.g., copper vs. aluminum, PET vs. PVC) could further improve downstream recycling value but requires more granular labeling and larger models.

The authors suggest exploring multimodal sensing (e.g., hyperspectral imaging) and reinforcement‑learning‑based actuator control as next steps to push both accuracy and throughput higher.

Authors

• Dhruv Talwar
• Harsh Desai
• Wendong Yin
• Goutam Mohanty
• Rafael Reveles

Paper Information

• arXiv ID: 2602.17642v1
• Categories: cs.LG
• Published: February 19, 2026
• PDF: Download PDF