Skip to content
WINGMANAUTOMATION
Measured results

18 points of efficiency, 18 minutes after it took over.

A live 45-minute autonomous control session on a 14-head machine running a 170 g pack, at a Canadian snack-food manufacturer. Four machines ran concurrently from one process: one autonomous, two operated the way they always had been, one in standby. Every figure below is counter-reconciled from the raw production databases.
01The headline

The machine wrote its own new record. Twice.

The system keeps a best-known-good record per recipe and only overwrites it when it beats its own previous best. During the session it beat the stored record twice: first 18 minutes after the controller took over, then again at 29 minutes.

No operator intervened at any point.

The 61.65 % starting point was this machine's own best on this recipe under manual operation, held down by the same feed problem described further down this page. Machine 2 was running a different recipe, so its efficiency is not directly comparable to that record.

61.65%
best previously achieved on this recipe under manual operation
Measured session, 23 Jul 2026The stored best-known-good for this recipe before the run, reflecting accumulated prior native operation.
80.30%
locked in under autonomous control, same shift
Measured session, 23 Jul 2026New recipe record written at 10:34:11, beating the 79.60 % written at 10:23:32, itself beating the stored 61.65 %.
The recipe record, before and during
60%70%80%ARMED61.65 %79.60 %80.30 %10:23:3210:34:11TIME OF DAYprevious best
Swipe sideways to see the whole chart
The stored best for this recipe, written by the machine itself. It went up twice in one session and could not have gone down: a record is only overwritten by one that beat it.Measured session, 23 Jul 2026Recipe record events from the controller's event log. The stored best-known-good for this recipe is only overwritten when the new figure beats it. Both writes were machine-initiated; no operator intervened.
02Like for like

Same shift. Same infeed. Same pack weight.

Machines 1 and 2 ran the same 170 g pack on the same line at the same time, from the same software process: one autonomous, one operated the way it always had been. Machine 4 ran a different 142 g product on a well-fed infeed, so it is not a like-for-like comparison. It is here because it shows the system reporting a healthy machine as healthy.

MachineModePackEfficiencyGiveaway / packGood packs
Machine 1autonomous170 g75.26 %1.06 g1,168
Machine 2manual170 g71.99 %1.56 g784
Machine 3standby————
Machine 4manual142 g99.86 %1.38 g1,436

Swipe the table sideways to see every column

Measured session, 23 Jul 2026Counter-differenced per machine from the weigher's own good-pack, missed-cycle, overweight and total-weight counters. Machine 3 was in standby all session and produced no production data.
32%
less product given away per pack (about a third) on the machine the software was driving
Measured session, 23 Jul 20261.06 g against 1.56 g on the manual machine, both running a 170 g pack on the same line, same shift. The weight counter advances in 100 g steps, which puts the reduction between roughly 20 % and 42 %.
1.06g
giveaway per 170 g pack under autonomous control, 0.62 % of pack
Measured session, 23 Jul 2026Counter-derived: 199.8 kg dropped over 1,168 good packs, mean pack weight 171.06 g against a 170 g target.
1.24kg
total giveaway across 1,168 packs, six-tenths of one percent of output
Measured session, 23 Jul 2026
03Through the session

What it looked like as it happened.

Efficiency through the session
406080100ARMEDRECORDRECORDMODEL010203040MINUTES · EFFICIENCY %
Swipe sideways to see the whole chart
Rolling efficiency across the whole session, derived from the machine's own counters rather than estimated. The short grey stretch is the machine before the controller took over; green is under autonomous control. Gaps are periods when the line was stopped or settling and no measurement was taken. RECORD marks a new recipe best; MODEL marks the first learned model going live. The machine was starved on four of fourteen heads throughout.Measured session, 23 Jul 2026Rolling window, counter-differenced from the weigher's good-pack, missed-cycle and overweight counters. Counters are ground truth; logged per-cycle scalars are not used.
84.7%
peak sustained efficiency, five-minute mean
Measured session, 23 Jul 2026
87.4%
peak instantaneous efficiency
Measured session, 23 Jul 2026
18min
from arming the controller to beating the recipe's previous best
Measured session, 23 Jul 2026
04It learned the machine, live

It learned your product in the time it takes to fill a pallet.

Mid-run, on a recipe it had no model for yet, the system trained and deployed its first model for that product without pausing control.

production samples used1,197
training time, in the background while control continued29 s
average error predicting how much each head fills (mean absolute error), on a 170 g pack0.904 g
heads characterised individually14 / 14
05Read this before you use these numbers

What this session was, and what it wasn't.

The caveat: this machine had a feed fault
The machine under autonomous control had a mechanical feed fault for the entire session: four of its fourteen heads were being starved by a product-distribution problem it had just discovered. Every figure on this page was achieved while starved.

This was a measured 45-minute session, not sustained production experience. It is a strong proof of concept and we describe it as one.

The run prioritised efficiency over giveaway by design. Giveaway is trimmed further in a later tuning phase, so the giveaway figures are not the best this system does.

Efficiency and giveaway are derived by counter differencing (the machine's own counters) rather than from logged per-cycle averages, because counters are ground truth. Where the two disagree, we publish the counter figure even when it is less flattering.

No customer, site, equipment manufacturer, model or product brand is named anywhere on this site, by policy.

We'd rather measure your line than estimate it.

That is what a line assessment is. You get your own version of this page, from your own data.

Book a line assessmentCalculate your giveaway