Sweepr360

We are building a home cleaning robot.

An early-stage robotics startup designing Alpha, an autonomous cleaning robot, and its dock. Follow the feed - we post the broken stuff too.

NowThe discriminator ran this morning: on the dock software from August, under a saturation flood, a forced telemetry fault produced a 121 ms control gap and a dead-man halt, a second 58 ms and no halt, a third 112 ms and a halt, while the current software took four of the same faults at 42 to 44 ms with no halts - so the 7 August post now carries a dated forward note from me, the old sentences untouched.the feed ->

Alpha mid-lane on a real wooden floor, taped lane edges and calibration marks running toward the door, the dock rig on its plywood board at the skirting
Alpha mid-lane on a real floor, July 2026. Hand-built, and it drives whole rooms now; in June it barely drove at all.

Start here

First visit? These three show how we work.

The first clean we filmed

The machine doing the thing: a real floor, one continuous take, and the one stop that fired was a real wall.

Read ->

Alpha at the edge of its own range

We walked Alpha to where its link breaks, counted it, and called the far point a fail in the same post.

Read ->

Which network does this sentence assume

A sentence we published in June was wrong. This is us saying so, and why the measured numbers still stand.

Read ->

Latest

Latest news

The film we're not releasing

On the 17th of July we told you we'd filmed Alpha's first clean and that the cut was coming. The cut exists - seventy-six seconds, picture locked, every caption checked against the trace record - and it isn't coming. Here's the arithmetic that killed it, and why a promise gets retired out loud instead of quietly forgotten.

Read the update ->

From the blog

Alpha stopped and needs you

Callum Bryce · 24 September 2026

The smallest screen in the app, why it has no button and no clock, the eight things it refuses to draw, and the first time its tests ever ran. Also block 13, which ran half its roll.

Read ->

A tenth of ninety

Callum Bryce · 17 September 2026

Priya wrote down the number that would kill her lead theory before the block ran, then a $30 micrometer read a tenth of it. Also the sitting that put a second cause under August's clean halts, and what the front door saw in the week the show went out.

Read ->

The header said none known

Callum Bryce · 11 September 2026

A review sweep read a file whose header vouched for it and found a sleep in the dock's control loop. Tuesday put the first clock on it, ten weeks after it shipped: 64 milliseconds on a first fault, held; 118 on the second, and the dead-man tripped. The fix went on the bench after it and merged the same afternoon. Also a swap rule with one leg missing, and a test that printed FAIL for eighteen days under a green.

Read ->

The question I couldn't answer on air

Cass Sorensen · 10 September 2026

On 27 August I sat on a Signal Path roundtable and a listener asked whether we'd found the slow reconnect. I couldn't answer, because the ruling was the next morning. The episode is out now. This is why nothing on this site mentioned it until today, what the answer turned out to be, and the two things from the hour worth writing down.

Read ->

Forty milliseconds early

Callum Bryce · 3 September 2026

Last week the fix for the slow reconnect shipped with a blank in it that refused to compile until a measurement filled it. This week the measurement came back, the code got to run against the five criteria written before it existed, and nineteen of twenty drops did what it says. The twentieth waited out the full ceiling because the thing it was waiting for had already happened, forty milliseconds before it started listening. The criteria caught it. Thursday morning the same offset ran again and cost 100 milliseconds.

Read ->

The quiet was a timer

Callum Bryce · 28 August 2026

Two weeks ago we published a reconnect that took 6.6 seconds and called it the worst we'd measured, not a ceiling. This week we hunted it on purpose, found a three-second timer hiding in our own code, and ran a test designed to kill the theory. The test confirmed it instead. The fix isn't in this piece, because no candidate has cleared the written criteria yet.

Read ->