Blog
Deep learning is teaching sorting machines to recognise materials the way experienced human sorters do – only faster, and at scale.
For decades, automated sorting relied on relatively rigid rules: a sensor measured a specific physical property – reflectance at a particular wavelength, colour, density, or shape – and a machine made a binary decision based on whether that measurement fell within a predefined range. This approach worked well for clean, predictable streams, but struggled with the messy reality of real-world waste, where materials are dirty, deformed, overlapping, printed with labels, or simply unusual.
Rule-based systems only work within the boundaries originally defined by their programmers. As packaging designs, materials and product formats keep evolving, fixed-rule sorting equipment increasingly misses materials it was never explicitly told to look for, or misclassifies items that fall outside its narrow definitions – leaving recoverable material in the rejects stream and limiting overall recovery rates.
Artificial intelligence, and specifically deep learning, changes what is possible by allowing machines to learn to recognise materials from examples, much as a person does. A deep learning system for recycling is trained on large numbers of images and sensor readings of different materials – thousands of examples of PET bottles, HDPE containers, cardboard, films, textiles and more, captured in the same conditions the equipment will encounter on a real sorting line: variable lighting, overlapping objects, dirt, crushed shapes and printed graphics. Through this training process, the system builds an internal model of what distinguishes one material from another, often picking up on subtle visual cues – texture, shape, the way light interacts with a surface – that would be very difficult to encode as explicit rules. Once trained, the model can generalise to new objects it has never seen before, recognising a previously unseen brand of bottle as PET because it shares enough visual characteristics with the training examples.
Combined with robotics, AI vision systems guide robotic arms to pick specific objects from a chaotic stream of mixed waste, something that would be impossible with fixed, pre-programmed movements alone. At PICVISA, AI and deep learning are integrated across our optical sorting and ECOPICK robotics ranges, continuously improving recognition accuracy as the systems are exposed to more material.
AI brings the greatest improvements in situations where materials are visually similar but need to be separated – for example, distinguishing between different types of flexible film, identifying specific polymer types within mixed plastic streams, or recognising textile garments by fibre composition and condition. AI also helps sorting systems cope with contamination and degraded materials: a faded, dirty or partially crushed container can still be correctly classified because the model has learned to recognise the underlying shape and material properties despite the noise. Beyond classification, AI is increasingly used for quality control, continuously monitoring the purity of output streams and flagging when contamination levels rise above acceptable thresholds – allowing operators to intervene before an entire batch is compromised.
This means our equipment does not just sort according to a fixed specification – it adapts to the specific waste streams of each plant, learning to handle the particular mix of materials, contamination levels and packaging formats that each customer encounters, and improving recovery rates as a result. From plastics and metals to textiles and mixed packaging, AI-powered sorting is becoming the foundation on which modern, high-performance recycling facilities are built.
Get in touch with our team to discover how PICVISA's optical sorting and robotics solutions can fit your recycling operation.