AI-Powered Failure Detection for 3D Printing: Obico, Spaghetti Detective, and Self-Hosted Alternatives

Our OctoPrint and print farm fleet management guides cover getting a printer online and keeping a queue of jobs organized, but neither answers a more basic question: how do you know a print has failed before it's wasted eight hours of filament and left a rat's nest of plastic wrapped around the hotend? AI-based failure detection watches the print through a webcam and flags spaghetti, warping, and detachment in real time - and it's cheap enough now that there's little reason to run a printer unattended without it.

How It Works

These systems run a lightweight computer vision model - usually a fine-tuned YOLO variant trained on thousands of labeled images of failed prints - against a webcam feed pointed at the bed. The model looks for the visual signature of a failure: strands of filament floating above the print (spaghetti), a print that's detached from the bed and is being dragged around, or a nozzle clogged and extruding into a blob instead of a layer. When confidence crosses a threshold, the system can pause the print, send a notification, or in more integrated setups, cut power to the printer entirely.

This is a meaningfully different problem from general object detection - failures look different across materials, colors, and lighting, which is why the better tools use models trained on large, continuously updated failure datasets rather than a one-off script someone wrote in a weekend.

Options Compared

ToolHostingCostNotes Obico (formerly The Spaghetti Detective)Cloud (free tier) or self-hostedFree tier limited detections/month; paid tiers for unlimited + higher accuracyMost mature option, OctoPrint plugin and Klipper/Moonraker support, mobile app notifications Obico self-hosted serverFully local (Docker)Free, your hardwareNo cloud dependency or per-detection limits, but you run the ML inference yourself PrintWatcher / custom YOLOLocalFree, DIYRoll-your-own with a Roboflow-trained model or public failure-detection weights; more setup, full control Bambu Lab AI cameraOn-printer (X1 series, some P/A series with camera)Included with compatible printersBuilt into firmware, no separate host needed, but locked to Bambu hardware Klipper + Moonraker webcam pluginsLocalFreeGenerally simpler heuristics (e.g. comparing frame deltas) rather than trained models - lower accuracy, no extra dependencies

Setting Up Obico Self-Hosted with OctoPrint

The self-hosted route avoids the free-tier detection cap and keeps your print farm's video feed off someone else's servers - a reasonable default if you're already running the headless OctoPrint setup covered elsewhere on this site.

1. Stand up the Obico server component with Docker Compose on a machine with enough CPU to run inference - a Raspberry Pi 4/5 can do it for one or two printers, but a small x86 box or a GPU-equipped machine scales much better across a farm.
2. Install the Obico plugin from the OctoPrint Plugin Manager on each printer's Pi, and point it at your self-hosted server's address during the linking step instead of the public app.obico.io URL.
3. Position the webcam with a clear, well-lit view of the full bed - the model needs the print outline visible, not just a close-up of the nozzle.
4. Tune the detection sensitivity. Too aggressive and you'll get false positives on normal supports and complex overhangs; too conservative and you'll miss real spaghetti until it's substantial.
5. Wire up the pause/notify action - most setups pause the print and push a phone notification rather than fully cancelling, since a human can often resume a print that stalled for a benign reason (filament runout, a bumped bed).

What It Won't Catch

• Dimensional drift - a print that's slowly warping out of tolerance but hasn't detached looks fine to a camera watching for gross failure.
• Layer shifting from a skipped step on one axis - if it's subtle, current models often miss it until it compounds into visible chaos.
• Silent under-extrusion - weak layers that still look geometrically correct from a top-down camera view.
• The first few layers - most models are tuned to detect established spaghetti, not to judge whether a first layer is adhering correctly. Pair this with a proper first-layer calibration routine, not as a replacement for one.

Treat AI failure detection as a safety net for catastrophic, obvious failures during unattended printing - not a substitute for good calibration, a leveled bed, and dry filament.

Scaling Across a Print Farm

Running inference per-printer on a fleet of Pis gets expensive in CPU quickly. For farms beyond a handful of printers, centralize inference on one beefier host (an old gaming PC, a Pi 5 with an AI accelerator HAT, or a small server) and have each printer's plugin stream its webcam feed there rather than running the model locally on each print's own single-board computer. This is the same principle covered in our print farm fleet management piece - centralize what doesn't need to be distributed, and keep the per-printer footprint light.

Safety Note

AI failure detection is a monitoring convenience, not a fire safety system. It watches for filament spaghetti, not thermal runaway or an electrical fault, and a pause command sent over a flaky WiFi connection is not a substitute for a smoke detector in the room and a printer that shuts off power (not just heaters) when genuinely unattended for long stretches. Keep these systems as a second layer, not the only layer, especially for prints left running overnight.

For a home shop or small farm, self-hosted Obico paired with a decent webcam is the highest return-on-effort addition you can make to an unattended printing setup - it turns "I came home to a nest of melted plastic" into a phone notification sent hours earlier, while the print was still salvageable.