How it works · for the curious

The robot mower you can read.

Mowlark is built in simulation first, measured with real telemetry, and documented all the way to a parts list for your own yard. No cloud, no closed boxes. Every bug we find goes in the devlog.

See the parts list Read the devlog

RTK FIXED blade 300 rad/s cut 0% sim t 0.0 s

Design concept · M1

Liquid chrome. It floats.

One seamless mirror shell over a dark skirt tucked so far in that the body hovers. A light seam all the way round tells you what it's doing from across the lawn.

Render of the M1 concept: a seamless chrome shell with haunches over the rear wheels, a dark skirt and a cyan light seam, on a striped lawn

GNSS eyeMetal blocks satellite signals, so the RTK antenna sits under a dark glass dome at the crown.
HaunchesTwo swells hide 240 mm wheels under a 245 mm shell.
Light seamCyan: positioned and allowed to mow. Amber: blade on. Red: fault.
Stop buttonRed, raised, top-rear. Reachable without putting a hand near the blade.

The M1 at night, its light seam glowing cyan — Night

The M1 from the rear three-quarter view, showing the haunches and stop button — Rear three-quarter

Top view of the M1: reverse-teardrop outline, GNSS eye and red stop button — Plan view

A design concept rendered from the same procedural model the engineering drawings use. Not a product you can buy.

Found in simulation, before it cost money

A robot that lies about where it is

The mower steers by wheel odometry, the way a real one does. Gazebo knows where it really went. On the first drive the two disagreed by more than a metre, and the robot had no idea.

1.18 m1.1 cm

Wheels skidded in every turn

Odometry counted 90° and the robot really turned about 99°. Cylinder-shaped wheel contacts were unstable in the physics engine. Sphere contacts fixed it.

6.7°0°

The blade twisted the robot

Spinning the blade from 0 to 300 rad/s in one step pushed back on the body. A 3-second spin-up before driving, as real mowers do, removed it.

48 cm1.3 cm

Our own metric hid an error

A rotated square still ends where it started. The report now checks the worst error along the whole path, not only the endpoint.

Three plots of a 3 m square drive. Cylinder wheels: the true path ends 118 cm from odometry. Sphere wheels with instant blade: a clean square rotated about 7 degrees. Sphere wheels with blade spin-up: the true path covers odometry exactly. — Orange: where the robot thinks it went. Blue: where it really went. Same drive, three robots. From `analysis/plot_paths.py` over the recorded MCAP bags.

Phase 4 in simulation

It mows the lawn, and goes home to charge

Headland laps round the edge, stripes across the middle, a ring round every flowerbed and tree. When the battery runs low it drives back to the dock, charges, and carries on where it stopped. Every number below comes from a pass/fail report over a recorded run.

95.5%

of the reachable lawn cut

Measured from the blade's real path in Gazebo, on a 5 cm grid. No keep-out incursions; stays within 15 cm of the planned line while mowing.

1.1 cm

docking error, after a recharge

It backs off, lines up 1 m out on the dock's axis and creeps in at 0.15 m/s. Contacts must stay closed for 0.15 s before it counts.

~576 m²

per 155 Wh charge

208 mWh per square metre, from a battery model with a real voltage curve and internal resistance. It never dropped below 13% charge.

Battery voltage and state of charge over a mowing run: discharge while mowing, a return to the dock, a charge, and mowing resumed — Mow, return, charge, resume: `analysis/power_report.py`.

The plan

From simulation to your yard, in stages

The parts most likely to hurt someone are the blade guard, the body and the battery. A cheap commercial mower has already solved those, so the first real robot is a brain transplant, and the blade stays unplugged until the safety checklist passes.

H0
Simulation
Gazebo yard, ROS 2 Jazzy, telemetry into Elasticsearch, pass/fail reports. It mows, avoids obstacles, docks and charges, and it has found every bug on this page before any hardware existed.
Running
H1
Bench and push-cart
Raspberry Pi 5, RTK GNSS and an IMU on a garden cart. Survey the lawn boundary to centimetres, with no motors at all.
Next
H2
Brain transplant, blade off
A used YardForce Classic 500 with the open-source OpenMower mainboard and our ROS 2 stack. Drive, map, dock.
Planned
H3
Blade on
Only after every item on the safety checklist passes and is logged. Supervised mowing first.
Gated
H4
Custom chassis
Our own frame and motor controllers, matched to the sim. Only if the commercial body holds us back.
Optional

Overhead render of the simulated yard: fenced lawn, house, tree, flowerbed, and the red mower — The simulated yard, rendered on the GPU with no window open.

The mower's front camera view: lawn, wooden fence, tree trunk — What the mower's own camera sees.

Engineering, transparently

How it's put together

The drawings are generated from the code. The software diagram checks every arrow against the running robot's message graph, so it can't drift.

Swipe a diagram sideways, or tap it to open it full size.

The Pi 5 thinks, the Pico 2 keeps it safe, the motor controllers move it.

What to buy · prices checked 8 Nov 2026

Stage H1 parts list

Start with the sensors. Localization is the hard part, and you can test it by pushing a cart around the lawn.

Part	Why	Price
Raspberry Pi 5, 8 GB	Runs the same ROS 2 container as the simulation	$175
simpleRTK2B starter kit (u-blox ZED-F9P)	Centimetre-level RTK GNSS, antenna included	≈ $235
BNO085 9-DoF IMU	Heading and tilt; later the tilt-safety input	$29.50
RTK corrections	A free state network where available, or your own base station	$0–195
PSU, cooler, NVMe SSD	Power, heat in the sun, room for recorded runs	~$40–70
Stage H1 total, depending on what you already own		≈ $500–700

US dollars, rounded; euro prices converted at an approximate $1.12 per € (Nov 2026). Prices move. Raspberry Pi raised prices several times in 2025–26. The full list, with sources and the H2 robot parts, is in hardware/bom.md.

Safety

Software decides when to mow. Hardware decides when it may.

The blade motor's enable line runs through a chain of normally-closed contacts. Any one of them stops the blade, and a cut wire counts as unsafe.

Two buttons, no app needed. STOP latches: in the sim the wheels stop in 0.2 s and the blade is told to stop at once, and releasing it leaves the robot paused, never moving on its own. HOME/GO: a short press pauses or resumes, a long press sends it to the dock. The light ring shows what it's doing.

This is a hobby project, not a certified product. The checklist in hardware/safety.md is a minimum, and the build manual's blade-on gate must pass first.

Emergency stoplatching switch that cuts power
Lift sensorseither side lifted
Tiltpast 25–30°
Watchdogpulse every < 100 ms, or the blade stops
ROS requestmission logic
Blade enableactive braking to a stop

The code

Everything in one repository

src/            ROS 2 packages: robot, sim, telemetry
analysis/       pass/fail reports over recorded runs
hardware/       parts list, electrical, mechanical, safety
blog/           the devlog and the build manual (this site)
docs/           telemetry contracts, business, open source
website/        this page
observability/  Elasticsearch and Kibana as code

Every experiment runs with no hardware:

make build up
make drive-run

It drives a square in Gazebo, records the run, ships the telemetry and prints a pass/fail report.

The source repository is private for now. The robot is built to be extended through documented plugin interfaces, and parts of it may be published later; nothing is promised yet. The devlog shows the work as it happens. We credit and build on OpenMower.