Ground truth: the foundation physical AI is built on
September 28, 2026
Most businesses can answer two questions about their warehouse with confidence: what came in through the door, and what left it.
Everything that happens in between is murkier than most operators would like to admit.
The WMS says 15,000 units are in bay C4. The ERP agrees. But when did anyone last physically verify that?
Was it before someone shuffled the pallets to accommodate a rush delivery last Tuesday? Before the informal workaround a supervisor ran to get around a racking weight limit that nobody logged?
Physical closes this gap - the one between what systems record and what's actually on the floor. And understanding it properly is the difference between a useful deployment and an expensive one.
Physical AI and digital AI are not the same thing
When most people hear "AI in the warehouse," they picture software making smarter decisions on top of existing data. Better forecasting, automated reordering, smarter WMS rules. That's genuinely valuable, but it rests on an assumption that the underlying data is accurate.
Physical AI starts earlier. It starts with no data at all.
Most digital AI projects work with information that already exists somewhere - a transaction log, a CRM, a web dataset. Physical AI faces a completely different challenge. Nobody has ever digitised what it needs to understand.
Nobody has ever digitised a warehouse floor. Nobody has digitised the contents of racking eighteen metres up either. Nor has the exact location of every pallet after a busy Tuesday shift.
None of this exists in any structured dataset. The system has to go out into the physical world and create it, then keep it current as the world keeps changing.
Warehouses are not static. Forklifts move pallets. Cleaning crews shift stock temporarily. Workers set damaged goods aside.
A CAD file shows the day teams installed the racking - not what the floor looks like this morning.
What "ground truth" actually means
Ground truth is what's physically present - as opposed to what any system believes to be present. In an active warehouse, inventory levels and system records can diverge within hours.
What's changed is the scale at which it's now possible to build and maintain that ground truth:
- 10-12,000 pallet locations scanned per hour
- More than one billion warehouse locations scanned to date
- A full-site picture in 24 hours - versus roughly a year for a traditional manual stock count
When customers run their first proper scan, what surfaces tend to be surprising. In one case, around £1.5 million of stock had been quietly written off as lost. Simply because no one had line-of-sight to where it actually sat.
Some sites reporting full capacity turned out to have roughly 10% of space quietly available. That only became visible once teams mapped pallets accurately.
The data gap was always there. What's new is the ability to close it.
More data isn't the answer. More useful data is.
Dexory's own team took time to learn this lesson. The early instinct was to surface everything: give warehouse teams visibility into every discrepancy, every anomaly, every exception the system detected.
It turned out to be less useful than it sounded.
Presenting fifty red alerts doesn't help a team know where to start. It produces paralysis, not decisions.
The real value isn't in showing everything. The real value is surfacing the ten things that genuinely matter today. Ranked by day-to-day and financial consequence, and letting the rest stay in the background.
That distinction - between data and value - separates a platform that teams use from one they work around. The job of the intelligence layer isn't to mirror reality faithfully. The job is to decide, on the operator's behalf, what's worth their attention right now.
Why humanoids aren't the answer to this problem
Physical AI's rising profile has pulled it into the same conversation as humanoid robotics. The difference is worth being direct about.
The argument for humanoid warehouse robots assumes a general-purpose body solves the hard problem. Something that can reach, carry, and navigate like a person. But the binding constraint in most warehouse automation failures isn't the shape of the robot. The blinding constraint is the accuracy of the data the robot is acting on.
A humanoid sent to handle order picking from an empty, blocked, or mispicked bay will fail. Not because the hardware is wrong - but because the warehouse management system (WMS) had bad ground truth.
A humanoid form doesn't fix that. Better perception, constantly updated, does.
Dexory's view: accurate, current knowledge of the physical space is the harder problem. That's the one worth solving first - regardless of what shape the robot eventually takes.
What changes for people
The anxiety about warehouse automation tends to focus on displacement. The pattern across a global customer base is more nuanced. That base spans 3PLs, retailers and manufacturers across Australia, the Middle East, Europe, and North America.
The roles that automation takes on first tend to be the least desirable - repetitive tasks that nobody wants to own. Repetitive manual walkabouts that pull human workers away from higher-value tasks. Scissor-lift inspections of high pallets. Stock counts take weeks and teams complete them already out of date.
What expands is interpretation. Reviewing what the data reveals. Making decisions about stock placement and space use. Building expertise that's much harder to replicate than a manual count.
The nature of the work shifts. It doesn't simply disappear.
Where to start
Physical AI uses computer vision to build and constantly refresh that picture of the physical space. Physical AI is a discipline built on one premise. The physical world changes faster than any system of record can keep up with. Something has to keep checking.
Organisations getting genuine value from it today tend to share a few habits. They start from a specific warehouse problem. They treat ground truth as something that needing continuous refreshing. And they name an internal owner before the rollout starts - not after it stalls.
Read the full white paper for a deeper look at what this shift means in practice. It includes a readiness checklist for operators evaluating physical AI.
