01Recycling does not fail at the plant, it fails at the bin
Municipal solid waste keeps climbing with urbanisation, and the default handling, manual collection, late monitoring, mixed disposal, produces overflowing bins, wasteful collection routes, and contaminated streams. There is no shortage of smart bins that report when they are full. What there is a shortage of is anything that actually separates waste at the moment it is thrown away.
That distinction matters more than it sounds. Once wet organic matter touches dry recyclables, the dry stream is contaminated, and contamination is the single most common reason a load of recyclables gets rejected downstream. Sorting later, at a facility, is expensive and imperfect. The cheapest place to keep the two streams clean is the household, at the bin, before they ever mix.
A fill-level sensor tells you the bin is full. It does not tell you the bin is sortable. That gap is the whole point of this paper.
So the contribution here is narrow and concrete: a physical sorting device, not another monitor. It runs on a sub-cost microcontroller, it makes its decision locally, and it moves the waste into the right bin on its own.
02Why two sensors and not a camera
The obvious modern answer to "classify this waste" is a convolutional network on a camera feed. Those models reach high accuracy, and the paper does not dispute that. The problem is everything around the model: a camera, more compute than a hobbyist board can spare, and almost always a cloud connection that adds latency, consumes bandwidth, and assumes an always-on network that rural and resource-limited settings simply do not have.
For the binary question, is this wet or dry, none of that is necessary. Wet organic waste and dry waste differ along one axis that a two-dollar sensor can read directly: water content. So the design fuses a moisture sensor with an infrared proximity sensor and lets an Arduino Uno do the deciding. The whole thing runs at the edge, responds in real time, and keeps working when the network does not.
This is the same instinct that runs through the rest of the portfolio: do the work where the data is, instead of shipping the data somewhere to do the work.
03The hardware, deliberately boring
Every part was chosen to be cheap and easy to source. An Arduino Uno is the controller. A moisture sensor reads the water content of whatever lands in the slot, and an infrared sensor detects that something has landed there at all. A servo motor drives the lid, and a stepper motor drives the sorting flap. A 5 V buzzer gives audible feedback, jumper wires tie it together, and a regulated 9 V supply powers the lot.
The layout is modular so any single component can be swapped or upgraded, with the Arduino sitting in the middle of a dual-bin frame and the sensors mounted right at the input slot for a clean reading.
04One threshold does the classifying
The decision logic, written in embedded C through the Arduino IDE, is about as simple as a classifier gets. When the infrared sensor sees an object, the moisture sensor takes a reading. If that reading is above 65%, the item is wet and is routed to the wet bin. Anything at or below the threshold is treated as dry.
The actuation follows a fixed choreography. The servo opens the lid, the stepper rotates the flap by ±120° to align the item with the correct bin, the lid stays open for five seconds to let the item drop, and then the lid closes and the flap returns home. The system goes back to standby and waits for the next object. No model to load, no parameters to tune in the field, nothing to retrain.
05What the numbers say, and what they do not
On the test samples, a tomato for wet, paper and dry leaves for dry, the system classified correctly every time, for a reported 99.1% accuracy. Average response time, from infrared detection to the item landing in the right bin, stayed under two seconds across every trial.
Stability was checked over 20 consecutive operating cycles with no misclassification and no hardware failure, which is the property that actually matters for an appliance meant to run unattended. Power analysis showed the actuators, the servo and stepper, drew the most current while the sensors and Arduino drew almost nothing, leaving total per-cycle energy low enough for long-duration home use at minimal cost.
How the proposed device sits against the alternatives:
| System type | Method | Cost | Cloud | Sorts? | Accuracy |
|---|---|---|---|---|---|
| CNN smart bin | Vision | High | Yes | No | 92-98% |
| IoT monitoring bin | Level sensing | Medium | Yes | No | n/a |
| Edge-intelligent | Hybrid ML | Med-high | Partial | Limited | ~95% |
| Proposed | Moisture + IR | Low | No | Yes | 99.1% |
06The honest limitations
It is worth being exact about what the 99.1% does and does not mean. It was measured in controlled laboratory conditions, on a small set of representative samples, for a binary wet-versus-dry decision. It is a clean prototype result, not a claim about messy real-world streams or municipal-scale throughput, and the paper says so plainly.
Three limits follow directly. First, the system only does binary separation, it does not distinguish plastic from metal from glass from paper. Second, validation was on a narrow sample set in the lab, not in the field. Third, moisture alone can misjudge edge cases: a sealed wet item reads dry, a damp-but-dry item reads wet. Naming these honestly is what makes the future-work section credible rather than aspirational.
07Where this sits in the portfolio
This is the most tangible end of a longer thread on sustainable, resource-constrained IoT. The lightweight oneM2M interoperability work is the same idea at the protocol layer, getting real capability onto constrained gateways without the usual heavyweight stack. The AI-optimized VLSI work is the same idea at the silicon layer, buying energy headroom so capable IoT endpoints can run on a tiny power budget. This paper is that instinct turned into an object you can build for very little that still does something useful on its own.
Different layer, same through-line: keep the intelligence at the edge, keep the cost on the floor, and design for the deployments that do not get a reliable network.
08What comes next
- Multi-class sorting. Move beyond wet and dry to plastics, metals, glass, paper, and biodegradables, using lightweight edge-intelligence models and added sensors rather than a cloud classifier.
- Field hardening. Toughen the sensors for real disposal conditions and validate on large, messy, non-curated waste streams instead of representative samples.
- Municipal integration. Connect bins to municipal IoT networks for fill-level reporting and dynamic collection-route optimisation, cutting fuel, cost, and carbon.
- Self-powered operation. Solar and low-power hardware so the unit runs sustainably and self-reliantly, with predictive analytics on the gathered data to help planners forecast waste-generation trends.