Technical Constraints in Designing Recycling Systems
The client is a UK-based recycling technology startup focused on transforming the bottle recycling experience through intelligent automation. To address long-standing usability and operational challenges within the recycling industry, the company set out to develop stand-alone recycling machines capable of operating 24/7 and automatically collecting recyclable bottles in a faster, more user-friendly way.
As the organization expanded its vision for automated recycling machines, the company faced growing challenges in delivering accurate and real-time AI processing within highly constrained hardware environments.
- Running AI models on low-power Raspberry Pi hardware: The system needed to operate entirely on-device using Raspberry Pi infrastructure with limited processing power and memory resources.
- Accurate classification of recyclable materials and bottle conditions: The platform needed to reliably identify glass, plastic, and cans while evaluating item size and damage levels under varying real-world conditions.
- Preventing fraud and invalid recycling attempts: The solution required real-time mechanisms capable of detecting reverse conveyor movement and identifying non-recyclable organic waste.
A scalable edge AI platform was required to enable real-time recycling validation, improve operational reliability, and support fully automated recycling workflows.
Engineering an AI-Powered Recycling Intelligence Platform
The client required a solution capable of performing real-time object recognition, fraud detection, and semantic analysis directly on low-power edge devices was essential.
Our team engaged as a strategic engineering partner to design and implement a lightweight AI-powered recycling intelligence platform optimized for stand-alone recycling machines.
1. Building a Multi-Stage Object Classification System
Addepto implemented a two-stage object classification workflow capable of distinguishing between glass, plastic, and aluminum cans while also evaluating the degree of item damage for plastics and cans.
Classification outputs were used to determine recycling reward eligibility and improve operational accuracy across automated recycling workflows.
2. Leveraging Lightweight AI Models for Edge Processing
The solution used MobileNet for object classification due to its lightweight architecture and strong performance on low-power Raspberry Pi hardware. DeepLab v3 was implemented for semantic segmentation tasks to evaluate object size and shape with high precision.
Both models were fine-tuned using custom recycling data to improve performance within real operational recycling environments.
3. Enabling Real-Time On-Device AI Processing
The entire platform was engineered to run locally on Raspberry Pi infrastructure without relying on cloud-based processing.
Edge AI optimization reduced infrastructure complexity while enabling low-latency processing and fully autonomous recycling operations in real time.
4. Developing Real-Time Anti-Fraud Detection Mechanisms
Addepto implemented fraud detection systems capable of identifying reverse object movement on the conveyor belt using motion vector analysis from Raspberry Pi camera feeds.
Additional logic was introduced to detect organic waste and prevent invalid recycling attempts, helping ensure fair and transparent reward distribution.
“Although the initial concept seemed simple, the software imposed significant constraints. We needed a model that could run quickly on modest Raspberry Pi hardware, unlike cloud-based systems with extensive computing power. As a result, we focused on compact AI models rather than state-of-the-art computer vision solutions.”
Michał Pocztowski
Senior Data Scientist – Addepto
Transforming Bottle Recycling Through Intelligent Edge AI
The partnership between Addepto and the client transformed traditional recycling workflows into a fully automated AI-powered recycling experience capable of operating continuously in real-world environments. Recycling machines can now identify materials accurately, validate bottle conditions in real time, and prevent fraudulent activity without relying on cloud infrastructure.
By combining lightweight AI models with advanced computer vision techniques, the platform significantly improved recycling efficiency while maintaining fast and reliable processing on low-power Raspberry Pi devices.
Before
- Manual or semi-automated recycling workflows
- Limited material validation and fraud prevention
- Slow and inconsistent recycling verification processes
- Dependence on centralized infrastructure for advanced processing
After
- AI-powered automated bottle classification and validation
- Real-time fraud detection and anti-abuse mechanisms
- Fully local AI processing on Raspberry Pi hardware
- Scalable stand-alone recycling machines operating 24/7
Beyond operational improvements, the solution established a scalable foundation for future AI-driven recycling and sustainability initiatives.
As part of KMS Technology, Addepto continues to help organizations modernize operational workflows through intelligent edge AI, computer vision, and scalable automation platforms.
Ready to build your Edge AI solution? Contact us today!