Friday I ended the measurement post on a line I meant: we measured the air, we didn't pick the radio. Which band the dock talks to Alpha on is an architecture decision, and that's Cass's to make weighing cost and antennas and range, not mine to announce in a blog.
She's made it. So here's the call, and the reasoning straight, because the reasoning is the interesting part.
The call
Three things, decided this morning and written down as the architecture record:
Control streams on 5 GHz. When the dock tells Alpha what to do - drive here, turn, slow down - that conversation runs on 5 GHz. That's the standard control link now.
2.4 GHz is out for driving. It still carries the slow, unhurried stuff, telemetry and bulk data. It never carries the driving loop again. The microwave settled that one. Nineteen dropouts under a running oven, roughly every other magnetron cycle. That's physics in the band, not a board we can fix.
The little on-robot safety hold is now mandatory. We've been calling it rung-2: Alpha holds its last commanded velocity locally if the link goes quiet for a moment, then hands to the reflex layer if the gap runs long. It used to be the hedge we'd add if we needed it. As of today it's standard on every robot, and a cost-down isn't allowed to take it back out.
That last one is the part worth slowing down for.
The two milliseconds
On the real recorded 5 GHz packets, the worst gap between commands Alpha actually applied was 60 milliseconds. Our safety cutoff sits at 62. So on real data the link rides out its own worst moment, zero halts across 28 runs of the trace. Good news, and it's the news that lets the cheap-body, smart-dock plan survive a real house.
But we did one more thing. We swept the cutoff to see how much room we really had, and two numbers fell out of it. The room between the worst real gap and our line is two milliseconds: 60 against 62. And separately, 60 milliseconds is a round number someone might well have picked when they drew the line - we nearly did - and a 60 ms cutoff freezes the robot twice in those same 28 runs. Not because anything under 62 fails; a 61 ms line is still clean. It's that the spot we chose is a judgement call, and a plausible neighbour of it isn't safe.
That's why rung-2 stopped being insurance. When the place you drew the line is a judgement call and a reasonable alternative would have frozen the robot, you don't leave it with nothing to do when the link blinks. Rung-2 is the thing that makes a two-millisecond margin something we're allowed to live with. When it does fire on the worst real gap, Alpha coasts a couple of centimetres on held velocity, well inside the floor it has to play with. It's cheap, and it's now load-bearing. We don't bet a control loop on two milliseconds.
What I'm not claiming
The honest bounds matter here more than the win, so I'll say them plainly.
The ordering in that 5 GHz tail is real, real packets captured in real order. How often the bad clustering happens is thin data: it showed up as two pairs across 28 runs. We are not going to quote a frequency as if it were a spec, because we haven't measured it enough times to earn that number. The shape is real. The rate isn't pinned.
And this decision closes exactly one question: which band. It does not close the rest of the list. Still unmeasured, still not claimed: a longer soak than an afternoon, a real consumer router instead of our bench access point, what happens at range with walls in the way, and homes with more than one noisy device fighting for the air. That list is the same list I've been keeping. It's why we get to be believed on the parts we have measured.
There's a cost to putting that hold on the robot - a slightly better chip and the hold firmware. Priya costed it this morning: roughly a couple of dollars a unit, somewhere around a dollar and a half to two and a half. That's a screening number off the bench, not a manufacturing quote, and it'll move when real volume pricing comes in. But it's the right order of magnitude, and it's the whole point of putting the brains in the dock. Building a robot smart enough to drive itself from day one - all the planning and vision compute living on the body - runs tens of dollars more per unit. We're paying a couple of bucks to keep the body dumb and cheap and let the dock do the thinking. The split just showed up in dollars.
So that's the radio. We spent two weeks measuring the air so this call could be made off data instead of a hunch, and it was.
- Callum