# The AI was the easy part: building an edge waste-sorter that actually moves the trash

> Lessons from SWCS — an offline Jetson + YOLO + MobileNetV3 recycling station where 61 ms of inference hid behind 3.3 seconds of mechanics.

- Author: Li Yanpei (Lucian)
- Type: Post
- Published: 2026-05-30
- Language: English (en)
- Reading time: 2 min
- Tags: Edge AI, Computer Vision, Jetson, YOLO, Robotics
- Canonical: <https://lucian-dev.com/blog/swcs>

---

For my final-year project I built **SWCS (Smart Waste Classification System)** — a small, offline recycling station that doesn't just _recognise_ trash on a screen, but physically sorts it. It runs entirely on the edge: no cloud, no network, made for places like office pantries, labs and school recycling points where nobody's watching the bin.

![image.png](https://lucian-dev.com/uploads/1780560283422-image.png)

The honest takeaway after months of work: **the machine learning was the easy part.**

[Demo](https://www.youtube.com/watch?v=jkAqK3vqILI)

## What it does

Drop an item in. An infrared sensor catches it and stops the conveyor. Cameras grab a frame, **YOLOv8** finds the object and crops it, **MobileNetV3** classifies the material (paper, plastic, metal, glass, cardboard), and a small **Decision Engine** fuses the probabilities and applies a confidence gate. If — and only if — the prediction is confident enough, a servo flap swings to route the item; otherwise it stays on the default path and nothing risky happens. A Jetson Orin Nano does the thinking; two Arduinos handle the belt and the servo, talking over USB-serial JSON.

## The numbers

On a 120-cycle labelled run (60 metal, 60 non-metal):

- Offline classifier accuracy: **93.59%**
- End-to-end routing accuracy: **90.8%**
- Recognition time: **\~61 ms**
- Full sort cycle: **\~3.29 s**

Look at those last two again.

## Where the time actually goes

Inference: 61 milliseconds. The whole cycle: 3,288 milliseconds. The model was \~2% of the latency. The other **98%** was mechanical — servo actuation (\~2.7 s) and conveyor recovery. Every instinct says "optimise the model"; the data said "optimise the _machine_."

## The bug that taught me the most

Early on, the sorting flap **jittered** — it wouldn't hold its commanded angle, so even a perfect classification could fail. The cause wasn't code: the high-torque servo drew more current than the Arduino's logic rail could supply, so the voltage and signal ground sagged. The fix was embarrassingly physical — an **independent 5 V supply with a shared ground** — and the jitter vanished. Closed-loop accuracy, it turns out, depends on power integrity and grounding as much as on network weights.

## What's next

Extend physical routing from the metal test path to all five materials, cut the mechanical latency, and run it for days to see what drifts — cameras, serial links, belt alignment, servo wear. The recognition is ready; the _system_ is where the next gains are.

If you're building anything that turns a model's output into a real-world action: budget your engineering time for the actuator, not the inference.