What a high-speed sauce filling line asks of the bottle and pack

Quick answer: The published evidence supports a staged build. Speed up the filling, capping and labelling block first, then add vision inspection and end-of-line robots sized from your own case rates. Pack engineers have most influence at the first stage, because bottle, neck and label consistency decide whether a fast line runs or jams.

This piece rests mainly on statements from Autopack (a line integrator), robotics and vision suppliers, and PMMI survey data; independent test data is noted where it exists, and there is very little of it.

What P&B (Foods) actually installed

Packaging News reported on 1 November 2026 that Autopack has installed a second sauce packaging line for P&B (Foods), aimed at raising capacity and cutting manual intervention. Autopack says the line runs at up to 200 bottles per minute. Sustainable Packaging News lists the contents: an unscrambler, rotary net-weight filler, capper block and rotary wraparound labeller.

The commercial story is the staging. The partnership began in 2023 and has delivered a series of automation and production upgrades across the firm’s sauce and edible-oil operations. Autopack describes the work as a staged approach to automation, letting P&B (Foods) add capacity as requirements developed.

The public coverage is silent on bottle material, size, closure type, label format and measured efficiency. It also doesn’t mention inspection or palletising on this line. Treat the install as a worked example of the filling block, not a proof of AI.

On the filler: Food Pack Automation’s write-up says a rotary net-weight filler uses load cells under each bottle pedestal on the rotating carousel. Filling by mass rather than volume suits viscous products where temperature changes density. Separately, Packaging News reported on 11 August 2026 that Autopack signed a UK partnership with Italian maker Weightpack for net-weight filling and capping systems. Neither source says whose filler is on the P&B line.

What the pack has to allow at that speed

The coverage doesn’t describe P&B’s bottles, so what follows are general questions to put to an integrator, not details of that line.

  • Unscrambling: empty bottles are fed in bulk and orientated. Light, tall or oddly shaped containers tip and jam, so bottle weight, base and neck shape matter.
  • Weighing: with a load cell under every pedestal, ask how bottle-to-bottle weight variation is handled, and whether headspace stays acceptable (Free tool: fill and headspace calculator, to check the fill volume against the bottle).
  • Capping: neck finish and cap tolerances need to be tight and stable from batch to batch.
  • Wraparound labelling: consistent bottle diameter and label web quality, because the label wraps while the bottle moves.

A bottle change that looks minor on a drawing can change feed behaviour on a line like this. Ask for a trial on production-representative bottles before you commit to a new format.

Vision inspection: what AI changes and who says so

Packaging Europe’s machine vision feature says vision is already widely used for label inspection, code verification, fill-level and completeness inspection and reject handling. Deep learning is the newer layer. Instead of a fixed template, a model is trained on images of good and bad packs and is meant to tolerate normal variation.

The stated benefit is fewer false rejects, meaning good packs thrown away. The German trade title Packaging Journal carries Heuft’s account of its inspection AI, which it describes as a non-generative system used on empty and filled containers. It gives no false-reject figure.

Numbers are easier to find on vendor sites, and they don’t agree. One iFactory page quotes a false positive rate below 0.1 percent for its camera, while an Oxmaint page says AI brings false positives from 30 percent to under 3 percent. No common test method or product is given, so neither figure is a benchmark.

Oxmaint also says SKU changeovers are the most common cause of false positives. That’s a vendor’s view, but it points at something pack and artwork teams control: label variants, gloss and film clarity all change what a camera sees.

End of line: do the case maths before choosing a robot

Palletising comparisons mostly come from integrators and cobot suppliers, so expect some tilt. The robotics association Automate gives typical ranges by type, shown below.

Typical palletiser cycle rates by robot type. Low end of typical range: Cobot 8, Industrial robot 12, Gantry 10; High end of typical range: Cobot 12, Industrial robot 20, Gantry 15 (cycles per minute)
Typical palletiser cycle rates by robot type
Source: Automate (A3) robotic palletizing glossary; typical ranges, not specifications

Cobots handle payloads of 3 to 35 kg at 8 to 12 cycles a minute, industrial robots 12 to 20, and gantry systems 10 to 15. Suppliers draw the cobot line in different places: Remtec suggests a cobot fits when median demand is under 6 cases a minute and loads are under 15 kg, while integrator DNC Automation puts the limit at under 8 cycles a minute and under 20 kg.

Part of the gap is mechanical. Robotiq, a cobot supplier, notes that a large industrial robot can lift a whole layer per cycle, while a cobot lifts one box at a time.

Our calculation: assume a 12-bottle case (our assumption; no source gives P&B’s case size). A line at 200 bottles a minute then makes 200 ÷ 12 = 16.7 cases a minute, or 60 ÷ 16.7 = 3.6 seconds per case. That sits above the cobot range in all three sources and inside the industrial range, before pallet changes and mixed patterns. Free tool: pallet load calculator, for cases per pallet and truck fill once the case size is fixed.

Options compared

Option What it does What it needs from pack and line Evidence in this article
Unscrambler, rotary net-weight filler, capper, wraparound labeller Orients, fills by weight, closes and labels Stable bottle, consistent neck and label Integrator and trade press
AI vision inspection Checks labels, fill, closures, codes; rejects bad packs Consistent lighting, clean image of the pack, trained on your SKUs Mostly vendor claims
Cobot palletiser Stacks one case per cycle beside operators Cases under about 15 to 35 kg, modest case rate Supplier and association ranges
Industrial robot palletiser Stacks cases or layers behind guarding Floor space, guarding, engineering time Supplier and association ranges
Predictive maintenance Flags wear from sensor data before a stop Sensors on drives and bearings, maintenance records Vendor case studies only

The last row covers the efficiency claim most often attached to AI. Vendor pages cite large OEE gains for a Unilever plant or network, but two of them give different before-and-after figures, 72 to 84 percent on one Oxmaint page and 71.4 to 83.6 percent on an iFactory page citing Oxmaint. I found no confirmation from Unilever, so treat them as unverified.

How widely it is adopted, and what is still unproven

Adoption data is better for robots than for AI. PMMI’s 2026 robotics report, produced with Interact Analysis, found the US robotics market serving packaging and processing was worth more than $440 million in 2025 and is projected to reach about $800 million by 2031, a 10.3 percent compound annual growth rate. The report page says 72 percent of surveyed end users currently use robotics, based on 173 US professionals surveyed in May 2026. It’s a US sample and says nothing about UK sauce lines.

Still unproven in public sources: independent, like-for-like false-reject data for AI against rule-based vision on food packs; audited OEE gains on FMCG filling lines; and the effect of AI inspection on a pack’s design freedom. PMMI also offers a separate report on AI adoption in packaging equipment, which I haven’t seen.

What this means for you

  • Packaging engineers: before a format change, ask the integrator how the unscrambler, filler, capper and labeller each cope with your bottle’s weight, neck and diameter tolerances. Insist on trials with production bottles.
  • Procurement: ask for a quote with line speed stated as bottles per minute and cases per minute. Don’t rely on “up to” speeds without the pack format behind them.
  • Quality teams: if a vendor quotes a false-reject rate, ask for the test set, the SKU mix and the reject definition. Run any AI system in shadow mode against your current inspection first.
  • Design teams: flag artwork and label changes to whoever trains the vision model, since changeovers are where vendors say false positives start.
  • Operations: work out cases per minute before choosing between a cobot and an industrial robot, as in the sum above.

Sources

Header image: illustration, AI-generated.