AI-driven smart sorting: the evolution beyond traditional sensor-based sorting

For many years, optical sorting and sensor-based sorting (SBS) were treated as synonymous. The dominant paradigm was straightforward: select the appropriate sensor for a given waste stream, calibrate it properly, and interpret its signal through predefined thresholds, rule-based logic, or, to a limited extent, conventional machine learning methods.

This approach proved effective as long as material differences were sufficiently distinct and decision boundaries remained simple. However, as waste streams have become more heterogeneous and material compositions more complex, the limits of purely sensor-driven, threshold-based systems have become increasingly apparent.

Without innovation, a company is on life support!

Sylvain Montmory

The era of classical sensor-based sorting

Traditional sensor-based sorting systems relied on deterministic interpretation of sensor outputs. Once the correct sensing modality was selected, X-ray transmission for density contrast, visible imaging for color separation, inductive sensing for metals,  the classification logic was often based on thresholding or simple regression algorithms.

For example, in an X-ray transmission (XRT) system for metals, if the estimated effective atomic number or density proxy exceeds a predefined threshold, the object is classified as part of the heavy fraction. In a SWIR-based plastic sorting system, if reflectance within a specific wavelength band exceeds a calibrated threshold, the object is classified as the target material.

These systems were not simplistic from an engineering standpoint. On the contrary, they required careful calibration, environmental stabilization, white/dark balancing, and ongoing maintenance to ensure measurement consistency over time. Operating X-ray, hyperspectral, or spectroscopic systems in harsh recycling environments is technically demanding and often capital-intensive.

However, the classification logic itself remained largely binary and rule-based, partly because mechanical separation architectures were themselves binary (eject or pass). As a result, most systems operated under relatively simple decision frameworks.

This paradigm worked effectively when material differences were large and clearly separable under one sensing modality.

But as waste streams became more heterogeneous, these deterministic approaches reached structural limits.

The distruption introduced by artificial intelligence

Artificial intelligence has significantly reshaped industrial systems, and recycling is no exception. Traditional computer vision approaches required engineers to explicitly define the rules distinguishing one object from another, encoding shape descriptors, contour heuristics, texture filters, or spectral thresholds. Such rule-based systems are complex to maintain and often fail to generalize under real-world variability.

In contrast, supervised machine learning, particularly deep learning, enables models to learn discriminative features directly from labeled datasets. Instead of manually specifying decision rules, neural networks extract high-dimensional representations that can approximate, and in some cases exceed, human perceptual discrimination.

While AI applied to RGB imaging can match human sorting capabilities, its true impact emerges when extended to complex sensor data. The combination of advanced sensing modalities and data-driven models fundamentally changes how sorting decisions are derived.

The broader implications of AI in recycling, including dataset design, model drift, and long-term operational stability, warrant dedicated analysis and will be addressed separately. Here, we focus specifically on how AI transforms the interpretation of sensor data within sorting architectures.

Smart sorting: AI applied to advanced sensors

Smart sorting represents the convergence of advanced sensing and artificial intelligence. Rather than applying simple thresholds to X-ray, hyperspectral, or spectroscopic signals, AI models can now process high-dimensional sensor outputs and identify patterns that are not accessible through classical calibration logic. This unlocks capabilities that go beyond human perception.

For example, detection of trace elemental signatures indicating specific alloy compositions, identification of hazardous components such as embedded batteries in electronic waste reducing fire risks, recognition of subtle compositional differences within visually similar materials, and differentiation of alloy families that share density or color characteristics but differ in elemental profile.

AI models trained on multi-sensor data can exploit relationships between spectral features, geometric structure, and density proxies in ways that are not linearly separable using threshold-based logic. This marks a shift from sensor-based sorting, meaning sensor plus rule-based interpretation, to smart sorting, meaning sensor plus data-driven, adaptive interpretation.