You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Container images for every release on GitHub's registry (ghcr.io/jpittelkow/groundskeeper), for amd64
and arm64 (Raspberry Pi), with a ready docker-compose.yml and .env.example.
Demo mode (DEMO=1): a made-up robot mows a made-up yard (on a public park's lawn, with real aerial imagery
under it) with a week of history, so the panel can be tried without a Yarbo; it never contacts a robot, the Yarbo
cloud or any integration. DEMO_SEED=1 now means the same. scripts/screenshots.py makes the README screenshots
from it.
Sign-in. The panel now asks for a password before it shows anything; only /health and the sign-in page are
open. On the first visit you set the password with a one-time setup code from the panel's log
(docker logs ...), so nobody else on the network can claim it first. Settings → Access changes the
password, signs out this or every browser, and makes named API tokens for programs (Authorization: Bearer).
Forgot it: python -m app.auth reset in the container. PANEL_AUTH=off turns sign-in off for a panel behind a
reverse proxy that signs people in itself. AI assistants keep their own token. Upgrading: after this update
the panel asks for the setup code from its log once.
/health says whether the robot is mid-run (busy), so the deploy script's "not while mowing" check works
without signing in.
Dead ends in Map Studio: draw one (+ Dead end, starting inside an area), reshape, rename or delete it and send
it to the robot like a pathway; select one for its settings (speed, blades, trimmer), read back and checked. A
replaced dead end keeps its settings. From the app's own commands, recorded 2026-10-11.
Changed
The container no longer runs as root. It runs as PUID:PGID (1000:1000 by default; the Unraid example uses
99:100) and hands an existing data folder to that user on its first start. The libraries it downloads at build
time are checked against their hashes, and the image has its own health check.
A pathway's transit offset range goes to ±39.4 in (1.0 m), the Yarbo app's own limit (it was held at 27.6 in
until the limit was known).
Removed
The Terrain map tool. The 3D map shows the ground instead, with the robot's own heights (the flat height and
slope image no longer lined up with the photo).
Fixed
Planned vs driven compares the whole plan. The robot reports each area's plan as separate passes (the edge
around the outline and the fill stripes), and the panel kept only the first of them, so a run could be compared
with a fifth of its plan and look mostly off plan. Every pass is kept now, each from where the run took it up
(a resumed plan leaves out what the robot had already driven, and passes it lists as finished), and the passes kept
with earlier runs of the same plan fill in what a run is missing. Runs recorded before this are worked out from the robot's whole plan report kept with them.
A plan resumed after a long pause stays one run. A run paused for more than 3 hours closes, and its resume
was only joined back within 3 hours of the run's last movement, which such a run can never meet, so a plan
paused in the evening and resumed the next morning became two runs. A resume of the same plan session now joins
up to 36 hours later (never on a report that the plan is done: a new run's first report can be the last plan's). Runs split before this stay apart; Join with previous joins them.
Coverage of a resumed run counts the earlier runs of its plan. What they mowed is neither a gap nor part of
this run's share (it used to show as gaps "next to RTK loss"); the run page says how much
that was.
No stuck alert while the robot waits for an RTK fix. Standing still on a float fix with the blades stopped
is the robot waiting for its fix: the alert now says "Yarbo is waiting for an RTK fix" (and Live shows it), and the
stuck alert comes only if it still does not move 3 minutes after the fix is back.
One odd firmware reading is not a firmware change. A new version counts once the robot has reported it for a
minute; a chassis version that flickered for a few seconds was logged as two firmware changes.
Planned vs driven uses the robot's own plan when it was kept during the run: the robot reports each area's
path once it starts the area, and the panel now keeps it (only the run's first report was kept before, which has
no path yet). Runs without it are compared with a fresh preview, as before, and the result says which.
A Wi-Fi survey cut short by a restart that never rejoined its run still gets its report in Events.
Run distances are right, and survive panel restarts. A run's distance, points, RTK share and battery are now
worked out from its stored track when it ends, is joined or is closed after a restart. Before, the running totals
started again at every restart, and they counted positions that arrived out of place after a weak-signal gap as
driving. Every stored run is worked out again on upgrade; on a test mow the result came within about 4 % of the
wheel odometers.
Positions in the robot's own order. The panel now stores the robot's own time for each position. Distance and
coverage use that order, a position that is out of place and back is not counted as driving, and "hard going"
times wheel travel by it, so a burst of messages after a weak-signal gap no longer looks like jumps.
Wi-Fi survey across restarts: its counters are kept with the run, so the verdict covers the whole mow.
The dock is no longer a weak-positioning spot in Problem areas (time on the charger is left out).
A plan is not marked finished too early when the robot briefly lists an area as done while still mowing it.
Leaving a charging break no longer ends the run for the few seconds the robot shows no plan.
Late confirmations: a command marked unconfirmed is still watched for 3 minutes; if the robot carries it out
late, the log says "confirmed late". Nothing is sent again.
Guard notices are not repeated after a restart for the same detection (the protective pause is unchanged).
Changes to run rules and guards are logged in Events (old → new, and whether the panel or an assistant).
Home Assistant: a value that is not known yet is sent as "unknown", and one refused sensor no longer stops the
others from updating.
Secrets in the robot's replies are masked when kept, and stored copies are masked when the panel starts.
Claude Desktop setup fixed. Settings → AI assistant (and the setup guide) now give a Claude Desktop entry that
runs the mcp-remote bridge on your computer (needs Node.js). The old entry put an http server straight into
Claude Desktop's config, which only starts local programs. The page also says not to use Add custom connector: it
needs https and connects from Anthropic's cloud, which cannot reach a home network.
Secrets from other apps' commands are no longer kept. The panel logs and captures commands other clients send
the robot (to learn their formats); the Yarbo app's "join this Wi-Fi" command carried the Wi-Fi password, which ended
up in Events and Captured writes. Passwords, tokens and email addresses are now masked before anything is stored,
and copies already stored are masked when the panel starts.
Added
Follow the robot in 3D. With the 3D map on, Follow flies the camera behind the robot and 45° to one side and
keeps it there, turning with the robot (at most 40° a second, so the turn at the end of a stripe is a swing, not
a spin). Right-drag (two fingers on a phone) moves the camera round the robot and it keeps that angle; zoom and
tilt stay as you set them; a drag or Fit yard lets go. Works on Live and in a run's replay.
The robot in 3D. With the 3D map on, the robot is a small 3D model that faces its heading, leans with its own
pitch and roll, and glides between the panel's updates (it keeps going at its speed for a moment, then catches
up) instead of jumping. In Runs it drives the replay. The status now carries the robot's compass heading.
Replay draws the track as it goes. Press Play and the run's track grows behind the replay point (the slider
shows it up to where you drag); before Play the whole run shows. Play is now a green button with a play icon.
GET /api/runs/{id}/plan-feedback: the robot's plan reply kept for a run, by shape (or ?raw=true), for
comparing a run with the robot's own plan.
3D map. The new 3D button among the map tools tilts the map and raises the ground under it: the robot's
own RTK heights over the lawns, the 2015 lidar around them, eased together so there is no step where they meet.
The photo, zones, track, run replay and the robot all lie on it, on Live and in Runs. The menu next to it (1× to 3×)
draws the ground taller; right-drag (two fingers on a phone) turns and tilts. Map Studio and the camera
layout's small map stay flat.
Notifications: pick the Home Assistant notify service from a list of the ones your Home Assistant has, the
companion app's phones first, instead of typing its name. "Type a name…" is still there for anything missing.
Map Studio: move a group of corners in small steps. Shift-drag a box over a selected shape's corners (or
Shift-click them), then move them north, south, east or west with the arrows or the arrow keys by a step you set
(0.5 to 100 cm). A burst of moves is one draft edit, so one Undo takes it back.
Wi-Fi survey (Settings → Robot: Off / Next run only / Every run). While a run is driving, the panel asks the
robot every 30 s which Wi-Fi networks it hears and stores each one's signal where the robot was. Each ask is a ~4 s
scan on the robot, so Next run only surveys one run (at least 10 minutes of driving), reports in Events
(answers, reply time, pauses in the robot's messages during a scan and otherwise) and switches itself off.
Signal → Wi-Fi (survey): one network's signal across the yard in 4 m cells, with a list of the networks heard
(a saved one first, else the one heard best).
Imagery for every US state and DC: 91 new free sources, statewide where a state publishes one (18 states and
DC now have it at 15 cm or sharper) and the largest counties and cities elsewhere (2.5–15 cm). Alaska and Hawaii get imagery for
the first time (30 cm statewide). Browse the catalog lists each country's sources under a heading per state,
statewide first; searching a state by name or code (Texas, TX) shows everything that serves it, national and
worldwide sources included. On a phone the list is tall enough to read.
No-vision zones can be sent to the robot from Map Studio: switched on or off, reshaped and renamed, in place
(same id), read back and undone on a mismatch. Switching one on, growing it or moving it (more ground where the
robot ignores its cameras) asks for an extra OK. A zone's obstacle bypass is shown with the area's obstacle
avoidance labels (Moderate bypass). Adding or deleting one still happens in the app.
Pathway mow settings: Blades on and the transit path offset (on/off, range and step, in inches like an area's)
can be changed. Any offset change is checked as a whole: the range must stay within 27.6 in each side (the widest
the app has been seen to send) and the step within 3.9-19.7 in, until the app's own limits are known.
Changed
The panel is now Groundskeeper, with the new logo: a GK monogram over a strip of lawn, the robot as a dot
in the grass. It shows in the nav, browser tabs, the favicon and a home-screen icon (added to a phone's home
screen it opens as Groundskeeper). Home Assistant entity ids (sensor.yarbo_panel_*) and the install stay as
they were; only the link sensor's display name changes. New MCP snippets name the server groundskeeper;
existing assistant configs keep working.
Live follow-cam is live video: Frigate's live stream at 5 frames a second, under a second behind, instead of
a still every 3 seconds; in the Camera layout and the Follow-cam card. It stops while the tab is hidden or you are
on another page, and falls back to stills if the stream fails.
Problem areas: a weak-positioning spot names a weak HaLow link (−77 dBm or lower) or late corrections when
there was one, so a link problem can be told from a poor sky view. The spot's details show its HaLow average.
Top bar: the battery sits at the far right with an icon that fills with the charge, red when low through amber
to green when full; it stays in view on a phone. The local broker chip sits beside the cloud chip.
Map imagery: Browse the catalog and Fetch again are green buttons, easier to spot.
Live → Camera layout: the small map sits right under the emergency buttons, above the robot's state and
Control, instead of at the end of the panel.
Top bar, regrouped: problems first, tinted in their colour; the switches (Control, Map writing) as one pill;
the lawn; then at the right the connections as one pill (Cloud, Broker, RTK, shortened to the word and its dot on
narrower screens) and the robot's state with its battery at the far right, which keeps its row. On a phone the bar
is two rows: the title, layout switch and theme on top, then the scrolling chips with the state and battery always
in view at the end.
The name reads "Groundskeeper for Yarbo" in the nav.
Live → Camera layout: the small map always shows the whole yard. It refits when its card changes size (before,
it could open zoomed in, or stay wherever it had been left) and cannot be panned or zoomed; Map layout gives the
full map.
Live → Camera layout follows the robot: when no camera's coverage holds it, the big slot shows the map again,
and the camera comes back when the robot drives into view (each change waits 10 s). A pinned camera stays up.