Augmented Reality
Most augmented reality in manufacturing does not pay back. That is not a fashionable thing to write on a vendor site, but it is what the pilot graveyard in this sector says, and pretending otherwise is why so many plant AR programmes stall after the second site.
The ones that do pay back share a single property: the overlay has to land on a specific piece of equipment, not in a general area. Where the task only needs "you are in cell 4", a printed sheet on the machine works, costs nothing and never runs out of battery. Where the task needs "this valve, not the identical one beside it", nothing on paper can do it and the AR case becomes real.
That test decides almost everything else, including how accurate the positioning has to be and therefore what the deployment costs. This post applies it to the four applications plants actually consider.
| Application | Accuracy needed | Pays back? | The condition |
|---|---|---|---|
| Line-side work instructions | Centimetres, anchored to the unit | Yes | High variant mix and units that look alike. Low mix, and the operator has already memorised it |
| Changeover and setup | Centimetres | Yes | Frequent changeovers. The saving is minutes of downtime multiplied by how often it happens |
| Quality and first-article inspection | Centimetres, plus a model to compare against | Yes | Best of the four, because it produces a timestamped record as a by-product rather than just saving time |
| Training and onboarding | Metres is often enough | Sometimes | Only where the equipment is genuinely hard to distinguish. Otherwise a good video is cheaper and easier to keep current |
Three of the four demand centimetre positioning. That is not a coincidence and it is not a specification anyone chose for its own sake. It falls out of the same property: the value appears at the moment the overlay is attached to one specific unit.
This is the case that separates the approaches, and it is worth being concrete rather than abstract about it.
A plant room with forty near-identical pumps. A utilities floor with rows of matching valve skids. A packaging hall with twelve lines built to the same drawing. In each of these, the technician's failure mode is not that they cannot do the job. It is that they do the job to the wrong unit.
Consider what each positioning approach returns in that room.
Markers and QR codes work, at the marker. Someone has to place and survey a target on all forty pumps, keep them legible in an environment of steam, wash-down and repaint, and re-place them after every maintenance event that removes a housing. The maintenance burden scales with the number of things you want to identify, which is exactly the wrong way round.
Beacons put the technician in the room. One to five metres in a plant room with forty pumps identifies the room, which they already knew.
Plant drawings and P&ID tell them which pump it should be, and then they still have to work out which physical object corresponds to that tag, which is the actual problem.
Visual positioning returns full 6-DoF pose: exactly where the camera is and exactly which way it is facing, in the shared coordinate frame of a map of that room. Because the pose is absolute and the map is shared, a label placed on pump 27 stays on pump 27, for every device, next month. That is the entire mechanism, and it is why the identical-equipment case is the one to test a vendor on.
A Fortune 100 industrial customer running this in a private cloud measured 2.5x technician productivity on asset finding and a 4x reduction in mean time to repair. The mechanism was not that the repair got faster. It was that the time spent identifying the right asset collapsed.
The table says three of the four pay back. The mechanism is different in each, and the difference decides which one a plant should start with.
Line-side instructions return error reduction, not time. An experienced operator on a familiar variant is not slow, so the saving is not seconds per cycle. The saving is the rework, scrap and warranty cost of the units built to the wrong variant spec, which concentrates on new starters, agency staff and the first shift after a variant introduction. If the plant runs three variants and a stable crew, the number is small. If it runs forty and staffs peaks with agency labour, the number is not small.
Changeover returns downtime, and downtime has a rate the plant already knows. This is the easiest of the four to build a case for, because the finance team has the per-minute figure on hand and does not have to be persuaded of it. Minutes saved per changeover, multiplied by changeovers per week, at a rate that already exists in the model. Where changeovers are rare the case disappears, and it should.
Inspection is the strongest, and for a reason that has little to do with speed. A first-article check done through a positioned overlay produces a timestamped record of what was checked, from where, against which revision of the model, as a by-product of doing the work. Plants currently pay for that evidence separately, in paperwork the inspector fills in afterwards from memory. Getting it for free changes the arithmetic, and it is also the application least likely to be argued down by the people who have to use it.
Training is where most plant AR programmes start, and it is the weakest of the four. It demos beautifully, the accuracy requirement is loose enough that almost anything works, and that is exactly the problem: if metres are good enough, a video is cheaper and easier to keep current. Start with the application that needs the accuracy, then get training free once the map exists.
Accuracy numbers in this category get quoted loosely, and two genuinely different figures get flattened into one. They are not interchangeable and they answer different questions.
Visual positioning is sub-5 cm. It answers "where am I in this building, and which way am I facing". That is the number that matters for navigating to the right pump, anchoring an instruction to a machine, and keeping content in place across sessions and devices.
Object tracking is sub-millimetre. It answers "where exactly is this part, relative to my camera, right now". That is the number that matters for aligning a CAD model to a physical assembly, checking a fit, or overlaying a tolerance on a feature.
Most plant applications need both, in sequence: positioning gets the technician and the content to the right unit, then object tracking locks the overlay onto the part itself. A vendor quoting a single accuracy figure for both is either simplifying or does not do one of them. Ask which number applies to which step. The object tracking page covers the second half; the VPS page covers the first.
This is where manufacturing AR projects die quietly, usually about six weeks after a successful demo, and it is rarely about the technology.
Network egress. Plenty of plant networks do not allow a call to a third-party cloud API from the OT side, and no amount of demo quality changes that. If the only deployment mode is public cloud, the project ends at the security review.
Validation. In GMP environments, anything that touches a controlled procedure carries a validation obligation, and a vendor that ships a silent model update has just invalidated it. Version pinning and change control are procurement requirements, not features.
Data residency and contract. In this sector the binding constraint is usually a customer contract rather than a regulation. A plant building parts for a defence or aerospace prime frequently cannot let interior imagery of the line leave its own infrastructure, whatever the vendor's security posture.
The practical answer is to settle the deployment mode in the first conversation rather than the last. MultiSet runs in public cloud, private cloud, self-hosted on-premises and fully on-device, and the on-device path exists precisely for the case where nothing leaves the building. Options and terms are on the pricing page.
Most plants have already been scanned, often several times, for brownfield engineering, capital projects or as-built verification. That scan is usually filed as a deliverable and never used again.
Scan-agnostic ingestion means it can become the localization map directly. MultiSet accepts E57 point clouds from Leica, Faro, NavVis, Matterport and XGRIDS, meshes, metric-scaled Gaussian splats, 360 video and phone LiDAR. The full input list is on the 3D mapping page.
For a plant this matters more than it does elsewhere, because scheduling a capture crew into a live production area is not a booking problem, it is a shutdown conversation. Using a scan that already exists removes that conversation entirely.
Ask one question of any plant AR proposal: does the overlay have to land on a specific unit, or just in a general area? If it is a general area, buy a laminated sheet. If it is a specific unit among units that look the same, that is the case AR actually solves, and it needs centimetre positioning to work at all.
Line-side instructions, changeover and inspection pass that test. Training usually does not. Being honest about the fourth is what makes the first three credible.
See the work instructions use case, or send us one scan of a plant room and we will process it.
