Waste sorting & recycling recognition

Material recovery facilities (MRFs) and sorting plants separate mixed waste by hand or with sensors that see only material type. Vision models add object-level understanding: they recognise packaging formats, brands, colours and contaminants, drive air-jet or robotic sorting, and measure what is really on the belt. The result is higher purity of recycled fractions, fewer sorting-line staff in hazardous work, and measurable data for regulatory reporting.

Problem & value

• Today: incoming mixed waste is sorted by a mix of mechanical separation (screens, magnets, eddy current), near-infrared (NIR) optical sorters that detect polymer type, and manual pickers who remove contaminants and select fractions. Manual picking is repetitive, dust-, injury- and odour-exposed work with high staff turnover; quality of bales (purity) drives price and rejection by recyclers. Households and businesses also mis-sort at source, which contaminates streams.
• AI value: vision models see shape, texture, label and context that NIR alone misses (e.g. food-grade vs non-food-grade PET trays, multilayer cartons, black plastics, film, batteries as fire risk). They enable robotic picking, auditing of belt composition without manual sampling, early detection of hazardous items and per-customer contamination feedback.
• Regulatory pull: the EU Packaging and Packaging Waste Regulation (EU) 2025/40 sets recyclability, recycled-content and collection requirements, raising demand for verifiable, item-level composition data; the Waste Framework Directive and national extended producer responsibility (EPR) schemes need reliable fraction data [1][2].
• Benefit figures vary strongly by plant and vendor; use vendor-reported numbers only as hypotheses and measure purity/recovery on your own stream (see KPIs).

Variants

• MRF belt sorting: optical + AI sorters and sorting robots for lightweight packaging, paper, C&D and e-waste lines.
• Belt/bale analytics only: cameras above belts or at weighbridges quantify composition without actuating sorting.
• Smart bins and collection vehicles: cameras in public or commercial bins and refuse-truck hoppers detect contamination and wrong fractions to feed back to the customer or municipality.
• Reverse vending and deposit-return: recognition of returned containers (barcode + shape) — see VIS-018 for codes.
• E-waste, battery and metal scrap: X-ray/LIBS plus vision to identify battery chemistries, alloys and fire-risk items.
• Construction and demolition waste: skip and truck-load classification for pricing and diversion reporting.
• Private use: phone apps that suggest the correct bin for a photographed item (consumer-grade, local rules vary).

How it works

1. bulk waste on belt (singulated/spread)
2. sensor head: RGB + optional NIR/hyperspectral/3D/metal detection
3. real-time inference (detection or instance segmentation + classification)
4. tracking over belt speed
5. ejection point prediction
6. actuation: air-jet valves or delta/gantry sorting robot with learned grasp point
7. bunker / bale; belt analytics stored with timestamps
8. dashboards: purity, throughput, contaminants, per-load reports
9. human review: audit samples, relabel unknowns, handle exceptions

• Models: YOLO-class detectors or Mask R-CNN/Mask2Former-style segmentation trained on plant-specific classes (e.g. PET clear/blue/coloured, HDPE, PP, film, cartons, aluminium cans, steel, paper grades, batteries). Cluttered, deformed, overlapping objects are the core difficulty, which public datasets such as ZeroWaste and TACO were built to expose [3][4].
• Sensor fusion: NIR gives polymer chemistry, RGB gives shape and printing; fusing both is standard in 2026 products.
• Vision-language models: used offline to pre-label images, discover new classes and answer audit queries on stored frames, not in the millisecond sorting loop.
• Edge runtime: inference must meet belt speed (tens of milliseconds per frame); models run on industrial GPUs or edge modules next to the sorter with TensorRT/ONNX optimisation.
• Human-in-the-loop: operators set target purity and recovery, review low-confidence frames, relabel new items and verify bale quality; safety systems remain independent of AI.

Data preparation

This is the most data-hungry use case in the VIS family: models must learn what your waste stream actually looks like. Data means images of material on belts, in bins or at collection points, with labels for material type and contamination. In an SMB or municipal operator these images mostly do not exist; what exists are weighbridge records, delivery notes and quality complaints in ERP or Excel. Typical gaps are no labelled images, huge variation between seasons and sources, dirty, crushed and overlapping items, and unclear category schemes (what exactly counts as contamination).

Preparation steps: first agree the category scheme with the sorting staff (PREP-019), then record camera footage under real conditions over weeks and seasons, select representative frames, label them with a tool such as CVAT or Label Studio with experienced sorters, and keep a separate test set (PREP-018). Reuse existing records as weak labels where possible (PREP-020). Plan 50 percent or more of total effort for data collection and labelling; vendors with pre-trained models reduce this but still need site-specific tuning.

Implementation paths

Which path for whom

• Enterprise: Path B, because plant operators need the sorting data in their own systems and waste streams differ by site, which favours an own pipeline on cloud APIs.
• Public: Path C, because municipal and public waste operators face procurement rules and sovereignty needs, and open models on EU or German infrastructure meet them.
• Developer: Path C, because open detection and segmentation models can be trained on your own waste images and embedded in sorting products.
• SMB: Path A, because small recyclers and waste-handling firms are best served by ready-made sorting-camera and robot products rather than building their own.
• Private: Path A, because consumer apps that identify waste types are good enough for household sorting and need no setup.

Path A — Buy (recommended)

• AI sorting robots and sorters: AMP Robotics (US), ZenRobotics (Finland, EU; part of Terex), Bulk Handling Systems / MAX-AI, Machinex (Canada), Waste Robotics (Canada), TOMRA Recycling (Norway, EU/EEA), Steinert (Germany), Pellenc ST (France), Sesotec (Germany), Redwave/BT-Wolfgang Binder (Austria), Wittmann/Stadler (Germany) — EU-based vendors favoured for service and data-location reasons.
• Waste analytics: Greyparrot (UK) belt-monitoring analytics, Recycleye (UK), plus camera analytics from sorting vendors (e.g. TOMRA Insight).
• Smart bins/vehicles: vendors for bin-fill and contamination cameras (e.g. Enevo, Bin-e, CleanRobotics — check EU availability and GDPR posture).
• Choose A when: you operate a plant or municipality fleet and need guaranteed throughput and purity; vendors bring pre-trained models, hardware, safety certification and service.

Path B — Build on cloud APIs

• Use cloud vision (Azure AI Vision / Custom Vision, Google Vertex AI Vision, AWS Rekognition Custom Labels, EU regions) or multimodal models (Claude, Gemini, GPT, Mistral) for non-real-time tasks: photo-based classification at drop-off points, load inspection at a weighbridge, audit of stored frames, labelling assistants.
• Not suitable for the real-time sorting loop because of latency and connectivity; keep that on-site.
• Orchestration: MQTT/OPC UA from the line, object storage in EU region, dashboards in Grafana/BI.

Path C — Self-host (open source)

• Models: Ultralytics YOLO (AGPL, check licence for commercial use), RT-DETR, Mask2Former, SAM 2 / Grounding DINO for label bootstrapping, DINOv2 features for classification; export to ONNX/TensorRT.
• Data: ZeroWaste, TACO, TrashNet, plus plant images [3][4].
• Stack: CVAT or Label Studio for annotation, MLflow/DVC for versioning, NVIDIA DeepStream or GStreamer for streams, Jetson Orin or one HW1/HW2 GPU per line.
• Effort: needs ML, camera/optics and PLC integration skills; ejection mechanics and lighting are as important as the model.

Path D — Custom training

Usually required in some form: each plant has its own stream mix, lighting, belt colour and target classes. Plan for thousands to tens of thousands of labelled images, continuous relabelling of new packaging, and a retraining cycle (active learning from operator corrections). Vendors often do this for you in Path A; custom training pays off for niche streams (e-waste, textiles, C&D) or research.

State of the art (2026)

• Instance segmentation + sensor fusion on conveyors is mature in leading products; sorting robots are widely deployed for pick tasks, with grasp-point learning for deformable items [3].
• Brand- and format-level recognition (e.g. separating food-grade from non-food-grade packaging) is an emerging differentiator, supported by packaging-labelling and Digital Product Passport initiatives (see DATA-021).
• Foundation models (SAM-class, open-vocabulary detectors, VLMs) cut labelling effort and help to find unknown items.
• Limits: dirt, crushed and overlapping objects, black plastics, wet material, new packaging designs, domain shift between plants and seasons; accuracy reported on benchmarks does not transfer to your line without testing [3]. Vision cannot see through opaque layers or identify additives. Fire-risk items (lithium batteries) must be handled by dedicated detection and safety measures, not only by AI.

Shared resources & combinations

• Same stack (object-detection + segmentation + edge-runtime): VIS-001 (inspection), VIS-006 (food sorting), VIS-026 (counting & volume), VIS-018 (labels/serial OCR) — shared annotation tooling, edge GPUs and MLOps.
• Downstream: DATA-012 (CSRD/ESG reporting), DATA-021 (product carbon footprint & DPP), VIS-033 (food waste tracking), ROBO-017 (robotic disassembly and battery recycling).

Effort, cost & KPIs

• Path A: quarters from tender to commissioning; capital cost for sorter/robot plus service contract; some vendors offer robots-as-a-service.
• Path B: weeks for a pilot at drop-off points or weighbridges.
• Path C/D: 6–12 months to a robust line deployment incl. optics, ejection and PLC integration.
• Cost drivers: hardware (cameras, lighting, valves/robots), labelling, retraining, downtime during integration.
• KPIs: purity and recovery per fraction, throughput (t/h), picks per minute and pick success rate, false ejection rate, contamination rate of incoming loads, manual picker FTE hours, bale rejection rate, share of correct bin disposal.

People & change

Involve management for the investment case and the plant or department lead as process owner; sorting staff are the domain experts for labels and must be there from the start, because their knowledge defines the categories and the test set. IT covers cameras, network and integration with weighbridge or ERP; suppliers (waste deliverers, robot vendors) matter for input quality and contamination feedback. Works council: it depends. Cameras above sorting lines capture workers, and logs of sorting performance can be suited to monitor behaviour, so §87(1) no. 6 BetrVG likely applies where such data is stored per person or shift; objective suitability suffices per the BAG. Inform the works council under §90(1) no. 3 in any case; in public bodies, involve the Personalrat. Personal data is usually minimal; involve the DPO if workers or citizens are identifiable.

Support AI literacy per Art. 4 with training on reading confidence scores and handling errors, especially for safety around robot arms. Pilot on one line or one material stream, keep manual sorting as fallback. Typical concerns are job loss and safety; address them with retraining towards quality control. See GUIDE-004, GUIDE-008 and GUIDE-009.

Risks, regulation & compliance

• EU AI Act: typically minimal risk; AI literacy (Art. 4) applies. The system could become a safety component of machinery (robots) under the Machinery Regulation (EU) 2023/1230 and relevant standards; assess case by case [5]. If cameras in bins or on trucks capture people or link contamination to identifiable households, assess fairness and transparency; no Art. 5 practice applies as long as no biometric or social-scoring use is made.
• GDPR: belt cameras rarely capture personal data, but public bins, vehicle cameras and photo apps can capture people, faces and plates; use minimisation, blur/crop, DPIA, clear signage, and avoid household-level penalties based on solely automated decisions (Art. 22).
• Waste law: Waste Framework Directive, PPWR and national EPR/packaging acts define recycling targets and reporting [1][2]. Hazardous waste and batteries require separate handling duties.
• Occupational safety: robots and ejection systems need machine safety assessment; fire detection for batteries.
• Mitigations: physical audits of bales, independent sampling, drift monitoring, retraining after packaging changes.

Getting started

1. Characterise your stream: composition audit, target fractions and current purity, with photos.
2. Fix the physical basics (singulation, belt speed, lighting) before buying AI.
3. Pilot a belt-analytics camera first (read-only) to quantify composition and find the largest loss.
4. Run a vendor trial on your own material; compare purity/recovery against the baseline.
5. Build a labelling and retraining loop with operator feedback; plan EPR/CSRD reporting data from the analytics.

Sources

1. Regulation (EU) 2025/40 on packaging and packaging waste (PPWR) — https://eur-lex.europa.eu/eli/reg/2025/40/oj
2. European Commission — Waste Framework Directive — https://environment.ec.europa.eu/topics/waste-and-recycling/waste-framework-directive_en
3. Bashkirova et al. — ZeroWaste Dataset: Towards Deformable Object Segmentation in Cluttered Scenes (CVPR 2022) — https://ai.bu.edu/zerowaste/
4. TACO: Trash Annotations in Context — http://tacodataset.org/
5. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. TOMRA Recycling — AI-based sorting — https://www.tomra.com/waste-metal-recycling
7. ZenRobotics — https://www.terex.com/zenrobotics