"A Warehouse Digital Twin Is Only as Useful as the Site It Represents"
A warehouse layout can show where an autonomous vehicle fits without showing whether it can complete useful work there. Pedestrian activity, staging queues, floor conditions and charging requirements all affect the operating environment. When virtual testing leaves those conditions out, a successful simulated route may answer a narrower question than the warehouse needs resolved.
KION’s description of the information required for Physical AI testing in existing facilities puts that distinction at the centre of deployment planning. The task starts with capturing the site accurately, then connecting its physical features to the work performed within them.
Capture the facility, not just its dimensions
Johannes Hinckeldeyn, Director Research Engineering, KION Group, said in responses provided to The Supply Chainer: “Physical AI relies heavily on understanding physical geometry, obstacles, and constraints.” He identifies point clouds and three-dimensional LiDAR or BIM scans as ways to make an existing warehouse machine-readable. The information required includes floor irregularities, column positions, ceiling clearances, rack tolerances, lighting conditions and dynamic bottlenecks. Those details establish why a general building plan is an incomplete starting point. It may identify an aisle while omitting physical characteristics that affect how equipment operates within it.
The operational consequence is that the model’s scope needs to be explicit. Teams evaluating a simulation need to know which conditions were captured, which were assumed and which were excluded. Otherwise, a promising result can be interpreted as evidence about circumstances the test never represented.
Give the map an operating meaning
Hinckeldeyn also calls for semantic layering: annotating spatial information with functional zones such as picking stations, staging areas, pedestrian walkways, busy intersections and charging locations. That layer connects geometry to purpose. An empty section of floor may be physically traversable while being reserved for staging or regularly occupied by another process. A route can therefore be geometrically possible without being operationally dependable.
For warehouse teams, the mapping exercise creates a handoff between engineering information and operating knowledge. Employees who understand how space is used need a way to contribute information that a physical scan alone cannot establish.
A hypothetical example illustrates the distinction. A staging area that is clear during data capture may fill during a collection peak. If the test assumes permanent availability, it would not establish performance during that peak. This is a model-design consideration, not a condition reported at a KION installation.
Previous learning needs a site-specific test
Hinckeldeyn identifies similarity to previously learned environments as an important success factor for Physical AI in an existing warehouse. That qualification limits what can be inferred from experience elsewhere. A system that has performed in another facility still needs evaluation against the conditions and tasks of the proposed deployment.
Earlier coverage by Warehouse Insider examined how machine information can support operational improvement. Bryan Boatner, Chief Revenue Officer, Ranpak, said: “What's happening on the packaging line can be used to troubleshoot or prevent bottlenecks”.
Boatner was discussing sensor-connected packaging equipment, rather than KION’s virtual testing. The connection is the information requirement: models and analysis become useful when they reflect the process being operated.
For a warehouse using simulation before deployment, the corresponding question is whether observations from the real site can challenge the assumptions in the test. A detailed digital representation is a starting point; its continued relevance depends on how changes in the operation are handled.
Compare the operation before and after deployment
Hinckeldeyn argues that evaluation should extend beyond commissioning and ramp-up time to sustained performance, autonomy, robustness and safety, measured against clear baselines from before AI deployment. His response identifies the dimensions of assessment without specifying numerical targets or reporting a customer result. It therefore supports an evaluation framework rather than a demonstrated performance gain.
For warehouse managers, selecting the baseline is consequential. A faster autonomous movement may provide limited benefit if material still waits at its destination. Fewer manual interventions also need to be considered alongside completed work and the consequences of interruptions. The acceptance decision should connect the virtual test to those operating requirements. Teams need to establish which real conditions the simulation represents and how performance will be checked after deployment. The digital twin earns its place when it helps answer a defined operating question, with clear limits on what its results establish.
