How I spent ~400 hours and ~$830 building my own rat terminator with machine learning (instead of hiring an exterminator).

Justin Thomas

Why I Can’t
Work From Home

Rats chew wires for a mix of biological needs, environmental factors, and opportunistic behavior. The behavior is common in urban, agricultural, and domestic settings and has predictable causes and consequences.

— Rando on Quora

⚠️ Disclaimer · pulse check

!! This talk contains: rats, ballistics, and at least one dead rat !!

Chewed Cat-6 cable held in a gloved hand, attic in background

Dramatization

News article: Mice destroy NSW family home as plague continues — rodent-caused fire

Act I · framing

The Trap Didn't Work

Rat ignoring a cage trap, captured on Reolink IR camera

          08/15/2025 · 02:32 AM · rat walks past baited cage
        

Wile E. Coyote falling with a tangled parachute

Me (unbeknownst)

If the rat won't go to the trap, the trap can go to the rat.

The plan, roughly

Train ML model → 3D print → ... → Profit

Act I · the reasonable option

Normal Rat Extermination Project

Monday

Tuesday

Wednesday

Thursday

Friday

Saturday

Call Exterminator

Exterminator Sets Traps

Rats Removed

Go to the beach with family

More Beach 🏖️

~$300/month. Done in a week. Zero emotional scarring.

Act I · the unreasonable option

Rat Extermination With ML

August

September

October

November

December

January

Research

Train ML Model to Detect Rats

Experiment with 3D Printing Designs

Program & Wire up Servos

Ballistic Testing

Maybe Rat Eliminated?

Hours

400

Out of pocket

~$800

Rats confirmed dead (maybe)

Act II · training loop

YOLO != YOLO

Reolink IR capture — rat near trap, detected by YOLO

rat · 0.92

YOLOv8n · MacBook Pro · 640×640

01 Label draw boxes, auto split 70/30

Rat Image Labeler — custom bounding-box tool

Python/Tk labeler

02 Train 49 images · YOLOv8n · 640×640

03 Test + repeat ↺ mAP@0.5 0.94precision 0.91recall 0.88

Act II · the v1 rig

A servo that pulls a trigger.

Daisy pellet rifle Single-shot, then reset by hand Fusion 360 · PLA

Trigger mount CAD overlaid on the actual Daisy pellet rifle trigger

Isolated Fusion 360 CAD of the servo-to-trigger mount

Raspberry Pi

Feetech FT9035M 8.5kg·cm coreless micro-servo

Feetech · 8.5 kg·cm

Front-facing rat captured by the rig's IR camera under the deck

Act II · first trial

What's up, doc?

Why it didn't trigger

01 Training set was Reolink shots of rats from behind. Deployment rat arrived head-on.

02 Barely any IR low-light frames in training. The rig sees everything in IR.

03 Classic data distribution ≠ deployment distribution.

Training image: rear-facing rat in Reolink IR footage

The detector had mostly seen this angle. The first real rat gave it the opposite view.

Act II · fixing the dataset

More data.

Gemini Nano Banana

Stock & open web

My own legs

Empty scenes

Act III · mechanical path

Two detours leading to SO-101.

01 · overbuilt bearing

Fusion 360 thrust-bearing assembly cross-section

Assumed I needed a thrust bearing. Lots of CAD, zero testing.

02 · lighter weapon

Custom pitch harness designed to rotate an air pistol

CO₂ pistol: lighter, shorter, lower moment arm.

03 · SO-101 unlock

Custom pitch attachment that bolts onto the SO-101 base

Jetson

Open-source arm + custom pitch mount, Jetson on top. 1 Hz → 20 Hz.

Act III · operator interface

A real world FPS.

Camera Tracker Control web UI — live camera view with crosshair, servo sliders, trigger button

Live camera feed, crosshair overlay, yaw/pitch sliders, keyboard nudges, and a trigger button — useful for calibration and couch testing.

One simple operator loop

/status/stream-frame/set-position/trigger-servo

Act III · rt_200.py architecture

One tracker. Small, testable parts.

Act IV · nights 01–02

The turret moved. Then it engaged.

IR night capture — rat under the deck, camera has panned to track it

night 01 · 02:07 · auto-pan

Frame 01 · tracked

A rat in frame, right where the camera pointed itself. First real unassisted detection.

night 02 · 02:32 · shots fired

Frame 02 · engagement

Detector locked on. Turret slewed. Daisy fired. Likely hit; the spotted rat never returned on Reolink.

First real detection First engagement Calibration still mattered

Act IV · calibration

Camera Sees ≠ barrel points.

The procedure

01 Test-fire at a sheet of paper at known distance.

02 Measure pixel offset between camera crosshair and hit.

03 Shift crosshair down in software until it overlays the hit.

Foreshadowing

This calibration is only valid at one distance. Change how far the rat is, the offset shifts. Later: a whole failure mode.

Act IV · stereo fix

A fix.

Working stereo geometry

The code shifts the crosshair relative to the zeroed distance.

Math from the code

# stereo depth
# B is measured lens-center baseline
Z_mm = fx_px * 52.5 / disparity_px

# ratbot/robot/aiming.py
adjustment = focal_y_px * laser_vertical_offset_mm * (
    (1.0 / depth_mm) -
    (1.0 / reference_distance_mm)
)

adjustment = clamp(adjustment,
    -max_adjust_px, +max_adjust_px)

crosshair_y = target_y + adjustment

closer than 1m

move down

at 1m

0 px

farther than 1m

move up

        rig/live · detect → aim → fire · loop
      

Act IV · terminator moment

The turret tracks. The turret fires.

First clean end-to-end run. Detector locks on, SO-101 slews to target, pitch drops, Daisy goes off.

It is genuinely unsettling to watch a weapon track something on its own — even on an iPad.

End-to-end working Detector + servos + trigger Autonomous

Act V · the verdict

Was it worth it?

Option A · Exterminator

$300

time One week.

effort One phone call.

scars Zero.

output Dead rats.

Option B · Mine

~$830 + ~400 hrs

Hardware bill

7× Feetech servos$140
Raspberry Pi$60
Jetson$300
2× IR cameras$80
CO₂ pistol + pellets$30
Reolink camera$80
Bearings$10
Heat-set inserts$20
Other tools~$40
Filament + prints~$70

Months · several scars · 1 dead rat (unconfirmed) · everything I learned.

The obviously correct answer.

Act V · legal-ish, not legal advice

The Law

What is a machine gun?

NFA definition needs a firearm.

A pellet gun is air-powered, not explosive-propelled. The notes say: no NFA / machine-gun exposure for the air gun itself.

Booby trap problem

Autonomous firing is the red line.

Human-in-the-loop click while watching video is the safer distinction. “Rat detected → fire” by itself is spring-gun / tort territory.

Pest control comparison

Killing rodents ≠ license to discharge anywhere.

Rodent dispatch, animal-cruelty rules, and projectile-discharge ordinances are separate layers from firearms classification.

Ground turret vs flying turret

Mounted / crawlspace

Quadcopter / airborne

NC state: permissive for air guns on private property.

Local: city ordinances often ban outdoor discharge; enclosed/indoor containment is the safer read.

crawlspace: plausibly contained

FAA: dangerous weapons attached to drones are illegal without specific FAA authorization.

State/local: weaponized-drone and privacy rules stack on top.

yard + airborne: different regime

Practical read from the notes: federal air-gun classification is the easy part; human approval, local discharge rules, and “does it fly?” are the risk multipliers.

fin.

Q&A

Ask me anything — rats, servos, YOLO, why I didn't just call the exterminator.

Audience member raising a hand at the rat-detection turret talk

artist's impression

fin. · your turn

Go solve a $300 problem the $830 way.

Every skill in this talk — CAD, YOLO, servos, stereo vision — was a first attempt, with AI riding shotgun.

The barrier between “I wonder” and “it fires pellets” has never been thinner.

Find your rats. It's early.

Appendix

Appendix: the trigger arm kept breaking.

Trigger arm v2 — thicker, curved, re-oriented print

            trigger-arm · v2
          

2–3

shots before snap

40+

shots before snap

Iterations

01

Rotated the print, thickened the arm, added a curved shoulder.

Layer lines now cross the stress axis instead of parallel to it. 20× the shot life.

02

All the test-firing surfaced the bore-sight problem.

Hundreds of shots to prove the arm held up also proved the crosshair lied — same issue from earlier.

03

The harness holding the gun developed play.

Repeated firing loosened the mount that held the pistol itself. Force on the trigger dropped over time — a slow-motion failure.

04

Started designing a bolt mechanism to replace the snap-on.

Kinematically-locked instead of friction-held. Not finished.

Appendix · open source

Appendix: the depth camera, open-sourced.

Live disparity heat map (warm = near) beside the camera view ranging a Red Bull can with a metric distance readout

live · disparity map (warm = near) · click to measure

— Two Pi cameras + a Jetson → metric depth, ~800 lines of Python.

— Browser calibration UI — your laptop screen is the checkerboard.

— Dense disparity maps drive the crosshair's distance readout.

github.com/base698/stereo-csi ↗

Appendix · very meta

Appendix: this deck presents itself.

Presenter view of this very deck: current slide, next-slide preview, speaker notes, and the Open audience window button

presenter.html · showing this deck · showing this slide's neighbor

— The deck is plain HTML — one file, agent-built, arrow keys, print-to-PDF.

— Presenter mode: notes, next slide, timer — one click opens a synced audience window.

— Every animation and video you saw was mirrored to the screen the same way.

Why I Can’tWork From Home