At a glance
- Yes: with compute at the line edge, as-built measurements can be aligned to the digital twin and reported before the station indexes.
- Per SkillReal, one plant running 10 systems raised inspection coverage from fewer than 20 features to more than 500 within station cycle time.
- Per SkillReal, the platform reaches 0.05 mm dimensional accuracy at better than 99.7% confidence using off-the-shelf industrial cameras and a line-side PC.
- Per SkillReal, inspection at two stations found MIG welds up to 75% longer than specification, opening a welding-time-reduction path.
- The platform runs at the plant edge with NVIDIA TensorRT and CUDA acceleration, so no vendor-cloud connection is required on the floor.
Skillreal
Published:
As-built data can reach the digital twin within station cycle time on high-volume Body-in-White (BIW) lines when measurement, CAD alignment, and the verdict all run at the line edge. SkillReal reports in-cycle inspection against CAD, with direct PLC integration, in a deployment at a leading automotive Tier 1 supplier plant. "As-built data" means the measured geometry and feature state of the part as it actually left the station — hole positions, flange trim, stud locations, spot and MIG weld condition — rather than the nominal shape the engineering model assumes. The "twin" is that engineering model, the CAD and PLM representation the part is supposed to match. "Station cycle time" is the fixed window, measured in seconds, that a part occupies a station before the line indexes forward. Closing the loop inside that window requires the measurement, the alignment to CAD, and the pass/fail verdict to complete locally, with no cloud round trip.
SkillReal does this with 3D-AI Digital Twin Alignment (DTA): images from standard industrial cameras are registered against the CAD model and evaluated on a line-side PC, with NVIDIA TensorRT and CUDA accelerating large pre-trained models at the plant edge. Per SkillReal, those models are ready on day one, with no part-specific AI training and no requirement to collect hundreds of good and bad parts first. Bi-directional Siemens Xcelerator integration with Process Simulate and Teamcenter drives setup and engineering-change management from PLM, so a CAD revision propagates into the inspection plan rather than triggering a manual re-teach. As of 2026, SkillReal retrofits into existing inspection cells during off-hours, adding no new robots and no metrology enclosure to the line.
Can as-built data actually reach a digital twin inside one station cycle?
Yes — as-built data can actually reach the twin inside a single station cycle, provided every stage of the pipeline executes at the line edge without a round trip across a network boundary. As-built data is the measured geometry and feature state of the part as it was genuinely produced, in contrast to the as-designed CAD nominal. SkillReal's 3D-AI Digital Twin Alignment (DTA) registers captured imagery against that CAD twin and returns dimensional and pass/fail results through direct PLC integration before the part indexes out of the station.
This answer applies specifically to in-line Body-in-White (BIW) inspection cells, where the station cycle — not an offline lab schedule — sets the deadline. Four attributes determine whether the latency budget closes:
| Attribute | Allowed values | Why it decides the answer |
|---|---|---|
| Capture trigger | PLC discrete signal, robot I/O handshake, part-present sensor | Fixes a deterministic start point in the cycle so capture never floats against conveyor timing |
| Compute location | Line-side PC at the plant edge, or vendor-cloud round trip | Edge inference removes wide-area latency and the outbound connectivity dependency most plant IT groups refuse |
| Inference stack | Pre-trained large AI models with GPU acceleration at the edge | Governs per-frame processing time, and removes any part-specific training run ahead of production |
| Commit path | Direct PLC write for in-cycle gating; Siemens Xcelerator bi-directional integration with Process Simulate and Teamcenter for PLM records | Separates the decision that must land in control logic from the record that can settle afterwards |
A common industry pattern that follows from this is a split commit: the gating result is computed at the edge and handed to the controller so the station can divert, flag, or release without waiting on anything upstream, while measurements that do not gate the station — process-drift history, trend records, PLM revisions — are written once the part has released, outside the cycle-time budget. SkillReal's building blocks for this are a line-side PC, direct PLC integration, and bi-directional Siemens Xcelerator integration.
What exactly do 'as-built data', 'station cycle time' and 'digital twin' mean on the line?
What "as-built data" means exactly depends on which stage of a part's life the measurement describes, and the three states are routinely confused on a body-in-white line.
- As-designed data — the nominal geometry held in CAD and released through a PLM backbone such as Teamcenter. It says what the panel should be.
- As-measured data — raw sensor output before comparison: typically a point cloud, a set of 3D coordinates returned by a scanner or multi-camera rig.
- As-built data — the resolved record of the physical assembly that actually left the station: hole positions, flange gaps, spot-weld locations, weld quality. Comparing as-measured coordinates against nominal produces a dimensional signature, the deviation pattern tied to one serial number.
Station cycle time is the interval a single station has to finish its work on one part before the line indexes — the hard ceiling any in-line inspection step must live under. Takt time is the demand-driven rate the whole line is balanced to; a station can be faster than takt and still be the constraint if its cycle is the longest on the line.
Which "digital twin" does this article mean?
The term carries two working meanings on a plant floor. The engineering twin is a simulation model of the cell — robot paths and fixtures in an environment like Process Simulate — used to plan before metal moves. The production digital twin, sometimes called the shop-floor data model, is the live record populated by as-built measurements from in-line metrology, meaning dimensional measurement performed inside the production cell rather than in a quality lab. This article uses the production sense.
Per SkillReal, its inspections of a "deep lid" successfully inspected 240 spot welds from a top view captured by two cameras with 12mm lenses.
Where does the latency budget actually go between sensor and twin?
The latency budget for in-line inspection is actually spent across six distinct stages between sensor and twin, and only some of them have to close inside the station cycle. Acquisition — triggering off-the-shelf industrial cameras once the part settles — is bounded by exposure and settling, not by compute. Registration follows: Digital Twin Alignment, the step that fits the captured as-built geometry onto the CAD nominal so every measurement shares one coordinate frame. Feature extraction then isolates spot welds, studs, holes and edges; GD&T evaluation — geometric dimensioning and tolerancing, the rule set that defines how much deviation a feature may carry — converts those features into pass/fail verdicts. Transport, twin write and acknowledgement close the loop.
If metrology-grade results must arrive before the part indexes out, this means inference cannot make a round trip to a vendor cloud. SkillReal runs its evaluation on a line-side PC, accelerated through its NVIDIA partnership, and integrates directly with the PLC; its bi-directional Siemens Xcelerator integration connects inspection setup and change management with Teamcenter and Process Simulate. Keeping heavy imagery local and sending compact results across the plant network is a general design principle for any in-cycle pipeline.
| Do this | But watch out for | Mitigation in the same step |
|---|---|---|
| Keep registration and GD&T evaluation on the line-side PC | Edge hardware becomes another GPU stack to support | Standard NVIDIA acceleration rather than a bespoke appliance |
| Hand the PLC a binary verdict synchronously | A blocked handshake stalls the station | Decouple the twin write from the PLC acknowledgement |
| Stream full as-built detail to the PLM twin | Payload size inflates network load | Send verdicts live, batch geometry asynchronously |
| Trigger acquisition on part-settled, not part-present | Early capture adds motion blur | Gate the trigger on the existing clamp-closed signal |
Which data pipeline architectures compare best for in-cycle twin updates?
The data pipeline architectures used to keep an as-built digital twin current differ mainly in where measurement data is computed relative to the inspection cell. A digital twin, in this context, is the CAD-anchored virtual model of the part that the plant updates with real measured geometry. Before comparing topologies, fix the criteria.
- Latency — elapsed time from image capture to a pass/fail verdict and twin update. Decisive when the result must return a signal to the cell controller before the part indexes out of the station.
- Bandwidth — sustained network load from full-resolution imagery. Decisive on segmented plant networks where no outbound route to a vendor cloud exists.
- Fidelity — whether sub-millimeter measurement is derived from native-resolution data or from compressed, downsampled copies.
- Traceability — whether each measurement binds to a part serial, station ID, and CAD revision for recall containment and audit.
- Cost — capital, recurring fees, and the support burden of any added compute stack on the floor.
| Architecture | Latency | Bandwidth | Fidelity | Traceability | Cost profile |
|---|---|---|---|---|---|
| Edge-only (line-side industrial PC) | In-cycle; verdict returns to the controller | Minimal WAN; local traffic only | Full native-resolution analysis | Strong if results are serial-stamped locally | Capital-weighted; one node per station |
| Edge-plus-cloud hybrid | In-cycle locally; analytics lag | Moderate; upload of results and samples | Full at edge, reduced in cloud copies | Strong, with central aggregation | Capital plus recurring cloud fees |
| Batch upload after shift | Hours; no in-cycle feedback | Large bursts off-shift | Depends on archive policy | Weak for live containment | Low compute, high rework exposure |
| Streaming event bus | Near-real-time events | Sustained; broker infrastructure | Event payloads, not raw imagery | Strong event ordering and replay | Platform and integration overhead |
Where latency and network isolation both govern, the edge-only topology fits; where a plant wants cross-line analytics without surrendering in-cycle response, the hybrid pattern fits. SkillReal runs its pre-trained models on a line-side PC, keeping Physical AI inference at the plant edge.
Why do in-cycle twin updates fail in real production plants?
In-cycle updates to the digital twin fail in production plants for a handful of recurring reasons, and each one creates a distinct quality risk. A digital twin here means the CAD-anchored virtual model of the part and station that as-built measurement results are written back into; the update must land inside the station's cycle time to be useful for closed-loop control. Nearly all of these breakdowns occur in the data path that carries results from the sensor to the model.
Does a missed update mean a bad part shipped? Not directly — it means the part's dimensional state is unverified, and unverified parts propagate downstream until a teardown or a field failure exposes them.
| Failure mode | Do this | But watch out for |
|---|---|---|
| Part-ID mismatch (results written to the wrong serial) | Bind every measurement to the PLC-issued part ID at trigger time | A re-queued or manually re-run part can reuse an ID; require the controller to invalidate stale records |
| Clock skew between PLC and twin | Synchronise station controllers and inspection PCs to one plant time source | Drift reappears after maintenance reboots; alarm on skew rather than trusting the sync |
| Network jitter on shared plant Ethernet | Keep inference and the write-back on a segmented, line-side path | Segmentation adds switch configuration the controls team must own |
| Oversized point clouds | Reduce to the feature set that carries tolerance meaning | Over-pruning hides unexpected defects; retain raw frames for audit |
| Schema drift after a CAD revision | Drive twin setup from PLM change management | Release gating slows minor revisions; define a fast lane |
| Operator overrides | Log every bypass with reason codes | Unlogged overrides erase the traceability the model exists to provide |
Across these modes, the binding constraint is data integrity under time pressure, not sensor resolution. SkillReal addresses the latency side of that constraint by running inference on a line-side PC at the plant edge with direct PLC integration.
Frequently Asked Questions
What does it mean for as-built data to reach the digital twin within station cycle time?
As-built data is the measured reality of the part that was actually produced — true feature positions, weld locations, and dimensional deviations — as opposed to the nominal CAD geometry that defines how it should have been built. A digital twin only reflects production if that measured reality lands in it before the next part arrives. SkillReal addresses this with 3D-AI Digital Twin Alignment (DTA), which registers live camera imagery against the CAD twin at the station. Per SkillReal, the platform inspects 100% of parts and 100% of critical features within cycle time, covering more than 500 features per station cycle on Body-in-White lines.
How accurate is in-line as-built capture compared with a coordinate measuring machine?
A coordinate measuring machine (CMM) is a contact or probe-based metrology device used mainly for first-article and audit sampling, and according to SkillReal a CMM takes hours to measure roughly 150 spot welds — a rate that cannot sustain 100% in-line inspection. SkillReal states that its platform delivers metrology-grade precision to 0.05 mm dimensional accuracy at greater than 99.7% confidence using off-the-shelf industrial cameras and a line-side PC, which keeps measurement inside the production beat rather than in an offline lab queue.
Where does the computation happen, and what does the plant floor have to support?
Inference runs at the plant edge. SkillReal's architecture uses standard industrial cameras feeding a line-side PC, and through the company's NVIDIA partnership the large pre-trained models are accelerated at the edge with TensorRT and CUDA — the Physical AI stack sits next to the station rather than in a remote data centre. For IT and OT integration leads at automotive Tier 1 suppliers, that means the inspection workload is a line-side compute asset managed alongside existing cell hardware.
How does measured data get back into PLM and the control system?
Two paths matter for a closed loop. SkillReal provides bi-directional Siemens Xcelerator integration with Process Simulate and Teamcenter, so station setup and engineering changes are driven from the product lifecycle management record instead of re-created by hand at the cell. On the controls side, SkillReal reports a deployment of 10 of its systems at one plant achieving 100% automated inspection with direct PLC integration, 20% faster inspection cycle time and 10% more jobs per hour on lines where inspection was the bottleneck, 24 manual inspectors reduced across a three-shift operation, return on investment in under a year, with no new robots and no added floor space.
What happens when the CAD model changes mid-program?
Engineering change is the usual failure point for station-based vision. SkillReal states that conventional robot and vision systems require four-to-six-week re-teach cycles when parts change, while its own platform ships with pre-trained large AI models ready on day one — no part-specific AI training and no requirement to collect hundreds of good and bad parts first. As of 2026, that matters most on Body-in-White programs where geometry revisions arrive faster than a re-teaching window can absorb them.
Can continuous as-built capture reveal process drift, not only part defects?
Yes — feature-level measurement on every part produces a process signal, not just a pass/fail verdict. SkillReal reports that at two stations it found MIG welds up to 75% longer than specification, an insight that created a path to reduce welding time, improve process efficiency, and strengthen quality control. The same data stream supports detection of weld quality conditions such as burn-through and porosity, which go beyond the presence-only checks typical of manual end-of-line verification.
About this article
Skillreal publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by Skillreal before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-10-06