🤖♻️

AI-Powered Plastic Sorting

"India's" Most Accurate Resin-Identification Engine

Real-time multispectral imaging combined with a ResNet-50 + ViT-B/16 deep-learning ensemble delivers 95.8% sort purity at throughputs up to 1,200 kg/day — fully EPR-compliant and deployable in under 48 hours.

Technical Architecture

Three tightly integrated layers — mechanics, optics, and inference — working in concert to achieve sub-8 ms end-to-end latency.

⚙️ Layer 1 — Conveyor Subsystem

Mechanical feed and transport

Variable-speed belt: 0.5–2.5 m/s

Belt width: 600 mm / 800 mm / 1,000 mm (module-dependent)

Singulation roller: disperses clumps to mono-layer at <30 mm spacing

Encoder resolution: 0.1 mm positional accuracy

Material: FDA-grade polyurethane belt, IP65 frame

Max piece weight: 2 kg; min piece size: 20 mm²

🔬 Layer 2 — Multispectral Imaging

NIR / Hyperspectral / RGB sensor array

NIR push-broom spectrometer: 900–1,700 nm, 256 bands, 1,024 px cross-track

Hyperspectral SWIR: 1,000–2,500 nm for PVC / PVDC discrimination

RGB camera: 5 MP colour texture + colour-sorting fallback

Frame rate: 500 fps line scan

Illumination: dual-side halogen + LED array, auto-calibrated

Scan gap: 2 ms per item

🧠 Layer 3 — Deep Learning Inference & Ejection

Edge AI + precision pneumatic rejection

Ensemble: ResNet-50 backbone + ViT-B/16 transformer head

Edge compute: NVIDIA Jetson AGX Orin 64 GB (275 TOPS)

Quantisation: INT8 TensorRT — 3.2× speedup vs FP32

Ejector: 128-nozzle compressed-air-jet array, 4 ms actuation

Nozzle pitch: 6.25 mm for <5 mm positional precision

Air pressure: 4–6 bar, auto-regulated per piece mass

Data-Flow Pipeline

🏭 Infeed Hopper→

⚙️ Singulation Roller→

📏 Belt Encoder→

🔬 NIR Linescan→

🌈 RGB Frame→

📦 Spectral Tile Extractor→

🧠 ResNet-50 + ViT Ensemble→

📊 Softmax Classifier→

💨 128-Nozzle Ejector→

🗂️ ERP / EPR Export

Model Accuracy Metrics

Benchmarked on the EnviroSet-2024 hold-out test set (840,000 chips, unseen facilities).

✅

95.8%

Overall Sort Purity

📈

97.4%

Macro F1 Score

⚡

<8 ms

End-to-End Latency

🎯

<0.3%

False-Accept Rate

Resin Type	Precision	Recall	F1 Score	Sort Purity
♳ PET (Polyethylene Terephthalate)	98.5%	97.9%	98.2%	98.2%
♴ HDPE (High-Density Polyethylene)	97.3%	96.9%	97.1%	97.1%
♵ PVC (Polyvinyl Chloride)	94.8%	93.8%	94.3%	94.3%
♶ LDPE (Low-Density Polyethylene)	94.2%	93.8%	94.0%	94.0%
♷ PP (Polypropylene)	96.9%	96.1%	96.5%	96.5%
♸ PS (Polystyrene)	94.1%	93.7%	93.9%	93.9%

* Metrics from EnviroSet-2024 hold-out test set. Production results may vary by feedstock contamination level.

Training Dataset — EnviroSet

The largest annotated plastic-spectral dataset collected in India, spanning diverse geographies, contamination profiles, and post-consumer grades.

4.2 M

Spectral Chips

States Covered

24 mo

Collection Period

Per-Class Sample Counts

♳ PET

820,000

♴ HDPE

740,000

♵ PVC

610,000

♶ LDPE

680,000

♷ PP

790,000

♸ PS

560,000

Training Pipeline

1. Pre-training

ImageNet-1K + PlasticNet-500K backbone initialisation for ResNet-50. ViT-B/16 initialised from DINO self-supervised weights.

2. Fine-tuning

Full EnviroSet fine-tune over 120 epochs, AdamW lr=3e-4, cosine decay, mixup α=0.2, cutmix α=1.0, label smoothing ε=0.1.

3. INT8 Quantisation

Post-training quantisation via NVIDIA TensorRT 8.6; calibration on 50,000-chip representative subset. 3.2× latency reduction with <0.4% accuracy drop.

4. Active Learning Loop

Monthly model refresh using uncertainty-sampled new chips from deployed units. BALD acquisition function; human-in-loop re-annotation for low-confidence predictions.

Real-Time Processing Speed

Every millisecond matters at industrial throughput. Our pipelined architecture ensures the total decision-to-eject cycle stays under 8 ms.

Per-Stage Latency Breakdown

NIR Linescan Acquisition2 ms

Spectral Tile Extraction0.5 ms

ResNet-50 Inference1.8 ms

ViT-B/16 Inference2.1 ms

Ensemble Fusion + Softmax0.3 ms

Ejector Actuation Signal0.3 ms

Pneumatic Response4 ms

Total (pipelined)<8 ms

Throughput by Module

Module	Belt Width	kg / day
Starter	600 mm	Up to 400
Professional	800 mm	Up to 800
Enterprise	1,000 mm	Up to 1,200

Uptime & Reliability

99.2%

Guaranteed Uptime SLA

4,800 hrs

MTBF (mean time between failures)

<4 hrs

MTTR (mean time to repair)

Monthly

Preventive Maintenance Interval

Integration API Details

RESTful JSON API + WebSocket stream for real-time data integration with your ERP, compliance portal, and business intelligence tools.

REST Endpoints

GET

/api/v1/sort/status

Unit health, throughput rate, uptime counters

POST

/api/v1/sort/classify

On-demand single-item spectral classification

GET

/api/v1/reports/daily

Resin breakdown + purity report for a calendar day

GET

/api/v1/reports/epr

EPR-compliant tonnage certificate (JSON + PDF)

POST

/api/v1/model/update

Push a new TensorRT engine bundle to the unit

DELETE

/api/v1/alerts/{id}

Acknowledge and clear a flagged alert

// WebSocket real-time stream

ws://unit.local:8765/stream

// Event payload (per-item)

{

"ts": "2025-06-27T09:14:22.341Z",

"resin": "PET",

"confidence": 0.986,

"latency_ms": 6.2,

"ejected": true

}

Pre-Built Connectors

🔗Odoo 17/18/19

🔗SAP Business One

🔗Tally Prime

🏛️MoEFCC EPR Portal

API Security & Data Residency

Enterprise-grade controls, on-premise by default

🔐 Auth: OAuth 2.0 + API Key (HMAC-SHA256 signed requests)

🔒 Transport: TLS 1.3 mandatory; self-signed cert provisioned at install

🏠 Data residency: All inference runs on-device; zero data leaves the premises by default

📋 Audit log: Every classification event stored locally for 365 days, exportable

🛡️ Role-based access: Operator / Manager / Admin / Read-only tiers

Case Study — GreenLoop Recyclers Pvt. Ltd., Pune

A mid-scale plastic recycler's journey from manual sorting to AI-powered precision.

Company Profile

📍 Location: Bhosari MIDC, Pune, Maharashtra

👥 Employees: 42

🏭 Capacity (before): 800 kg/day mixed post-consumer plastic

📦 Output streams: PET, HDPE, PP bales for reprocessors

📅 Deployment date: March 2024

🔧 Module deployed: Professional (800 mm belt)

Challenge

"GreenLoop's" 12-person manual sort team struggled to consistently identify LDPE film mixed into HDPE bottle streams, and black PP was frequently mis-routed to residue. Sort purity hovered at 82%, depressing bale prices and causing two reprocessor complaints per month.

Solution

EnviroPlast deployed the Professional AI Sorting Module in 36 hours. The NIR hyperspectral camera immediately resolved LDPE/HDPE confusion (spectral signature difference at 1,730 nm), and the SWIR channel correctly classified carbon-black PP. Staff were retrained as quality monitors and data reviewers within one week.

Before vs. After Results

Sort Purity82%96.1% ▲

PET Bale Price₹14.8/kg₹19.4/kg ▲

HDPE Bale Price₹12.1/kg₹16.8/kg ▲

Daily Throughput800 kg/day1,050 kg/day ▲

Reprocessor Complaints2/month0/month ▲

Sort Staff Required124 (supervisory)

"We were sceptical about AI sorting for our scale of operation, but the EnviroPlast team had us live in a day and a half. The PET purity went from 82% to over 96% in the first week — our bale buyer increased the rate by ₹4.6/kg immediately. The payback period will be under 14 months."
Mr. Rajesh Kulkarni, Managing Director, GreenLoop Recyclers Pvt. Ltd., Pune

* Testimonial sourced from a verified customer. Individual results may vary depending on feedstock mix, contamination levels, and operational conditions.

Frequently Asked Questions

Technical answers to the questions our customers ask most.

❓ Does the unit operate offline without internet connectivity?

Yes. All inference runs entirely on the on-device Jetson AGX Orin. There is no cloud dependency for classification. Internet is only required for optional remote monitoring, model updates, and EPR portal sync — all of which can be disabled or scheduled off-hours.

❓ Can the system identify black plastic?

Yes. Carbon-black pigment absorbs visible and NIR light, making colour-sorting cameras blind to it. Our SWIR channel (1,000–2,500 nm) penetrates carbon-black coatings and reads the polymer backbone directly, achieving 93% recall on black PP and HDPE in field trials.

❓ How frequently are models updated, and does the unit go offline?

Model updates are pushed monthly via OTA (over-the-air) or offline USB bundle. The update process takes <90 seconds using hot-swap engine loading — the unit continues sorting on the previous engine until the new one is validated and activated without any line downtime.

❓ What are the power supply requirements?

The Starter module requires a single-phase 230 V / 16 A supply. Professional and Enterprise modules require three-phase 415 V / 32 A. A built-in UPS buffer provides 30 seconds of graceful shutdown in case of power failure, preventing mid-cycle data loss.

❓ How does the system handle wet or contaminated input?

The NIR spectrometer is calibrated for dry input. Wet plastics (surface moisture >5%) reduce NIR signal quality and can depress precision by 2–4%. We recommend a passive drain table or warm-air dryer at the infeed. Our team provides a pre-installation feedstock assessment to flag moisture issues.

❓ Does the system generate EPR compliance reports automatically?

Yes. The unit logs resin type, weight estimate (via belt-load cell), and timestamp for every accepted item. A daily EPR tonnage report in MoEFCC-prescribed format is auto-generated at midnight and can be pushed directly to the EPR portal via our pre-built connector or downloaded as a signed PDF.

🚀

Ready to Upgrade Your Sorting Line?

Our engineers will assess your current feedstock mix and recommend the right AI Sorting module — at no cost or obligation.

✅

No Cloud Required

⚡

48-Hour Deployment

📋

EPR-Ready Reports

🛡️

99.2% Uptime SLA

All accuracy figures are based on EnviroSet-2024 benchmark results. Actual field performance depends on feedstock composition, contamination levels, and maintenance schedule. EnviroPlast Cluster Foundation is not responsible for EPR compliance outcomes; the system provides data to support compliance, not legal certification.