Physical AI Meets Model Management: Why Camera-First Pipelines Will Define Construction's Next Decade

Nicolas Waern WINNIIO AB / Life Atlas ORCID: 0000-0001-7970-2707

Corresponding author: ceo@winniio.io

License: CC-BY-4.0

Submitted: 2026-04-16

Abstract

This paper presents a design concept for camera-first construction AI pipelines, grounded in a 31-minute session on AI-powered construction site capture. The architecture described has not been implemented or evaluated; it is presented as a design contribution intended to frame the model management problem and motivate future empirical work. Construction sites are among the most data-rich environments in the physical world: they change daily, involve dozens of interdependent trades, generate thousands of decisions with quality and safety implications, and produce physical structures whose as-built reality frequently diverges from their design intent. They are also, by the standards of other data-intensive industries, extraordinarily data-poor: most of the decisions made on a construction site are not recorded, most of the as-built conditions are not documented at the time they occur, and the gap between design model and constructed reality is typically discovered at handover or in subsequent maintenance, not during the construction phase when corrections are cheapest. The session revealed an architectural approach — camera-first pipelines combined with cloud-edge hybrid processing and structured model management — that has the potential to transform the as-built documentation problem from a retrospective, expensive, and frequently incomplete process into a continuous, real-time, and automatically structured data stream. This paper analyses the camera-first pipeline through inscription theory: the camera as the primary inscription device, translating the physical construction state into digital representations that can be compared against design models, fed into quality management workflows, and archived as as-built documentation. We examine the model management layer — the software architecture required to manage the continuous stream of model updates that a camera-first capture environment produces — and argue that model management is the underexamined bottleneck in construction AI pipelines. We engage the counter-argument that camera-based site documentation is a mature capability and that the architectural claims in this paper are incremental rather than transformative. Three falsifiable propositions are stated. The paper contributes a design framework for camera-first construction AI pipelines grounded in inscription theory and model management architecture requirements.

Keywords: construction AI, camera-first pipeline, physical AI, model management, site capture, photogrammetry, cloud-edge hybrid, Actor-Network Theory, inscription theory, as-built documentation, construction digitalisation, design science research, SMILE methodology

1. Introduction: The Documentation That Never Gets Done

Walk around a construction site at any point between structural completion and handover. The building that is being built is not the building that was designed. Dimensions have been adjusted to accommodate discovered site conditions. Duct routing has been changed to avoid a structural conflict discovered during installation. Electrical conduit has been rerouted around a revised partition layout. Floor finishes have been substituted because the specified product was unavailable. Each of these changes was made by a competent professional exercising appropriate judgement. Almost none of them were recorded in the design model.

At handover, the facility manager receives a set of as-built drawings that are, at best, the original design drawings with markup annotations made by subcontractors who were contractually required to submit redlines but who did so from memory at the end of a project rather than as a continuous documentation practice. The gap between the as-built drawings and the actual as-built state is, in the facility management industry's common experience, significant and systematically underestimated.

The consequences compound over the building's operational life. Maintenance technicians work from inaccurate drawings. Renovation contractors discover services in locations the drawings do not show. Regulatory compliance records reference installations that were subsequently modified without documentation update. The as-built documentation problem that begins on the construction site propagates into every subsequent phase of the building's life.

Camera-first AI pipelines represent the first technically credible approach to continuous as-built documentation: capturing the physical state of a construction site as it evolves, processing the captures into structured model representations, and maintaining a running as-built model that reflects actual conditions rather than design intent. This paper examines what is required to make this approach work at construction site scale.

2. Theoretical Framework

2.1 Inscription Theory and the Construction Site

Inscription theory [1], as developed in the sociology of technology, addresses the process by which human actions, decisions, and knowledge are encoded into non-human artefacts. In construction, the primary inscription artefacts are drawings: the design team's decisions are inscribed in architectural and engineering drawings; the contractor's decisions are inscribed in shop drawings, RFIs, and site instructions; the as-built state is supposed to be inscribed in as-built documentation.

The as-built documentation failure is, in inscription theory terms, a failure of the inscription artefact: the redline drawing system requires human inscription effort at a point in the project lifecycle when inscription effort is in competition with the immediate demands of finishing construction and achieving practical completion. The inscription device — the drawing markup process — is poorly designed for its purpose.

The camera as inscription device has different properties: it requires minimal human effort (pointing a device at a surface, or mounting a camera that captures continuously), it inscribes at the moment of construction rather than retrospectively, and it produces a continuous record rather than a retrospective summary. The camera is a better inscription device for as-built documentation than the drawing markup system, for a structural reason: it shifts inscription effort from the end of the project to the duration of the project, and from human labour to automated or semi-automated capture.

2.2 ANT and the Construction Site Network

ANT [2][3] is appropriate for analysing construction site data architecture because the construction site is an ANT actor-network in the most direct sense: a temporary assembly of heterogeneous human and non-human actants — general contractor, subcontractors, design team, client representative, regulatory inspector, scaffolding, concrete pumps, BIM models, drawings, contracts, cranes, and weather — whose collective behaviour produces the constructed asset. The data produced by this network is as heterogeneous as its composition: photographs, inspections, time-lapse recordings, survey measurements, delivery notes, change orders, test certificates.

A camera-first pipeline is an enrolment mechanism: it enrols the camera — and through the camera, the physical construction activity — into the data network by translating physical site conditions into digital representations that can be processed, stored, and queried by the project's information management systems.

2.3 The SMILE Methodology Applied to Construction AI

Following Waern (2026a, DOI: 10.5281/zenodo.19587944), the as-is state is construction as-built documentation that is retrospective, incomplete, and disconnected from the design model. The to-be state is continuous, camera-generated as-built documentation that is maintained in real time, structured in a format consistent with the design BIM, and accessible to all project stakeholders without manual compilation. Every architectural choice in the camera-first pipeline is evaluated against this transition.

3. The Camera-First Pipeline: Architecture

3.1 Capture Layer

The capture layer defines how photographic data is collected from the construction site. Three capture modalities are relevant at current technology readiness levels:

Structured walk-through capture uses a human operator with a smartphone or 360-degree camera to systematically photograph all accessible areas of the site at defined intervals (weekly or bi-weekly). The capture produces a photogrammetric record that can be processed into a point cloud and compared against the design model. This modality is low-cost, requires minimal infrastructure, and can be deployed on any site with a single trained operator.

Automated fixed-camera capture uses networks of cameras mounted at fixed positions in the structure, capturing images at regular intervals (hourly or daily). The cameras provide continuous coverage of specific areas — structural connections, MEP service corridors, areas of high modification risk — without requiring operator presence. The modality is higher infrastructure cost but provides continuous coverage of areas that are inaccessible after subsequent construction phases cover them.

Drone photogrammetry uses UAVs to capture the site from above and from positions inaccessible to ground-based cameras. Particularly valuable for roofing, structural steel, and external facade documentation. Requires flight authorisation and weather-dependent scheduling.

The camera-first pipeline integrates all three modalities into a unified processing architecture.

3.2 Processing Layer: Cloud-Edge Hybrid

The processing layer translates raw photographic data into structured geometric and semantic representations. Two processing tasks dominate the computational load:

Photogrammetric reconstruction converts overlapping photographs into point clouds and textured meshes using structure-from-motion (SfM) and multi-view stereo (MVS) algorithms. This process is computationally intensive — a 2000-image site walk-through may require 30-90 minutes of processing on a dedicated GPU. The computational load makes cloud processing the default for photogrammetric reconstruction.

Design model comparison compares the reconstructed point cloud against the design BIM to identify deviations: areas where the as-built geometry differs from the design geometry by more than a specified tolerance. This comparison requires both the point cloud (current site state) and the design model (reference state) to be available simultaneously; in a cloud architecture, both can be held in cloud storage.

The edge layer serves a different function: preprocessing. Raw photographs must be quality-checked (blur rejection, coverage analysis) before upload, and time-stamped metadata (location, orientation, crew context) must be attached before the network connection is required. These operations run on edge devices — the capture operator's smartphone or a site-based compute node — reducing the data volume and improving the quality of what is uploaded to cloud processing.

3.3 Model Management Layer: The Underexamined Bottleneck

The model management layer is the architectural component that the session identified as systematically underinvested. A camera-first pipeline that operates continuously on an active construction site generates model updates at a frequency that manual model management cannot sustain: weekly walk-throughs produce weekly model snapshots; continuous fixed-camera coverage generates continuous model state updates; drone surveys produce additional snapshots on an irregular schedule.

Model management at this update frequency requires:

Version control for 3D models: the ability to track model states over time, compare specific versions, and roll back to prior states when processing errors are discovered. The analogies are to software version control systems (Git), but the data objects (point clouds, meshes, BIM models) are orders of magnitude larger and require specialised diffing algorithms.

Deviation tracking: the ability to identify when and where a deviation from the design model first appeared, correlate it with site activity records (which trades were working in that area at that time?), and escalate it to appropriate quality management personnel.

Model conflict resolution: when multiple capture sources (walk-through, fixed camera, drone) produce overlapping representations of the same physical area, the model management layer must reconcile potentially inconsistent representations into a single authoritative model state.

Access control and audit trail: different project stakeholders (client, design team, contractor, regulatory inspector) need different levels of access to the model and different views of the deviation log. The model management layer must enforce these access controls and maintain an audit trail of all model updates.

None of the current commercial construction AI platforms evaluated in the session provided comprehensive model management capability. The gap between capture capability (mature and well-served by commercial platforms) and model management capability (fragmented, immature, and typically delegated to general-purpose data management tools not designed for 3D model management) is the primary technical barrier to camera-first pipeline deployment at scale.

4. The Counter-Argument: Is Camera-First Construction Documentation Mature?

The counter-argument holds that camera-based construction site documentation is not new: photogrammetry, drone surveys, and site photography have been deployed on construction projects for a decade; the architectural claims in this paper are incremental rather than transformative; and the "camera-first pipeline" framing overstates the novelty of the approach.

This counter-argument correctly identifies that the component technologies — photogrammetry, drone surveys, site cameras — are individually mature. It is wrong in two respects.

First, the integration of these components into a continuous, automated pipeline connected to the design BIM is not mature. Current deployments use these technologies episodically (a photogrammetry survey at foundation completion, a drone survey at structural topping out) and produce outputs that are manually processed and compared against the design model. The camera-first pipeline described in this paper is architecturally continuous: it is designed to run throughout the construction phase, automatically processing captures, automatically comparing against the design model, and automatically escalating deviations. This is architecturally different from episodic survey deployment, not merely faster.

Second, the model management layer — the architectural component that makes continuous camera-based documentation practically manageable — is genuinely immature. The absence of 3D model version control, automated deviation tracking, and conflict resolution capability from current commercial platforms is not a minor gap; it is the bottleneck that prevents camera-first pipelines from delivering on their architectural promise.

5. ANT Analysis: Enrolling the Construction Site Camera

5.1 The General Contractor Translation

The general contractor's interest in camera-first documentation is complicated by the evidentiary implications of continuous site recording. A general contractor who maintains a continuous photographic record of all site activities has a more complete record of everything that happened on site — including things that should not have happened. The same record that documents progressive as-built conditions also documents unsafe work practices, deviations from specification that were rectified before discovery, and the sequence of events that preceded a defect claim.

The translation required: framing camera-first documentation as a risk management tool rather than a surveillance mechanism. A continuous as-built record protects the contractor against false defect claims by documenting the state of work at handover. It provides evidence that a subsequent trade's work caused damage to completed work. It documents the sequence of client-instructed variations that affect programme and cost. The record that feels threatening as surveillance becomes valuable as evidence.

5.2 The Subcontractor Translation

Subcontractors are on site under more constrained commercial relationships than the general contractor: they are typically paid by measured work, not for documentation effort. A subcontractor asked to support camera documentation is being asked to spend time on an activity that does not directly contribute to the measured work they are paid for. The translation required: embedding camera documentation in existing subcontractor workflows (progress photograph requirements that already exist in most subcontracts) rather than adding it as a separate requirement.

5.3 The Regulatory Inspector Translation

Regulatory inspectors — building control officers, structural engineers of record, fire officers — have inspection duties that currently require physical site visits at prescribed stages. A camera-first pipeline that produces structured, queryable, time-stamped model records of inspectable elements could, in principle, reduce the frequency of physical inspections by providing remote access to documented construction states. This translation requires regulatory engagement and potentially regulatory framework revision: most building regulation frameworks specify physical inspection rather than documented model review.

6. Falsifiable Propositions

1. The as-built completeness proposition: Construction projects that deploy camera-first pipelines with weekly walk-through capture and automated design model comparison will deliver as-built documentation with less than 10% geometry deviation coverage gaps (areas of the building where no photographic capture was obtained) at handover, compared to greater than 40% coverage gaps in equivalent projects using conventional redline as-built documentation processes. This proposition can be falsified by a comparative documentation completeness audit of camera-first and conventional as-built documentation processes across matched project types.

2. The deviation detection proposition: Camera-first pipelines with automated design model comparison will identify at least 75% of significant deviations (deviations exceeding specified tolerance thresholds in MEP routing or structural geometry) at least two weeks before the affected area is enclosed by subsequent construction phases, providing the opportunity for correction without demolition. This proposition can be falsified by a study of deviation detection timing showing that camera-first automated detection does not provide a significant timing advantage over conventional inspection for detecting rework-relevant deviations.

3. The model management bottleneck proposition: Construction AI projects that allocate at least 30% of their technical development budget to model management infrastructure (version control, deviation tracking, conflict resolution) will achieve deployment at full intended scale within their planned implementation timeline at a significantly higher rate than projects that allocate less than 15% to model management, reflecting model management as the primary technical bottleneck. This proposition can be falsified by a project portfolio analysis showing no significant relationship between model management investment proportion and on-schedule deployment rate.

7. Limitations and Future Research

This paper is a design concept contribution: the camera-first pipeline architecture described has not been built or evaluated, and the claims about model management gaps are based on gap analysis against described commercial platforms rather than systematic platform evaluation. The session that grounds the paper was 31 minutes in duration, providing a limited evidence base for the architectural claims. The construction industry productivity and digitalisation context is supported by industry-level analysis — McKinsey Global Institute [8] identified construction as one of the least digitised major industries and documented the productivity gap relative to other sectors — but the specific architectural argument of this paper has not been empirically tested.

The ANT analysis of contractor enrollment is based on general industry knowledge; specific contractual and commercial dynamics vary significantly across project delivery models (design-bid-build, design-build, construction management) and regulatory jurisdictions.

The photogrammetric processing performance figures cited in Section 3.2 are based on published benchmarks; actual performance depends on hardware configuration, software implementation, and image capture quality.

Future research should: (1) develop and evaluate a 3D model version control architecture specifically designed for construction AI pipeline requirements, testing it against the model management requirements specified in Section 3.3; (2) conduct a comparative study of as-built documentation completeness and accuracy across camera-first and conventional documentation processes in at least five matched construction projects; (3) develop and test a deviation detection performance benchmark for construction AI pipelines, establishing baseline detection rates for common deviation categories in structural, mechanical, and electrical construction contexts.

8. Conclusion

Construction sites produce and destroy evidence every day. The concrete that goes into a slab covers the reinforcement that should have been documented. The ceiling that is installed covers the MEP services that should have been recorded. The partition that is built covers the electrical conduit that should have been mapped. In each case, the opportunity to document the as-built condition exists for a narrow window — days or weeks at most — before the construction sequence closes it permanently.

Camera-first pipelines are, in essence, an attempt to keep the window open: to continuously document the as-built condition as it is created, rather than retrospectively reconstructing it from a combination of design drawings, contractor markup, and inference. The camera is a better inscription device for as-built documentation than the drawing markup process for a structural reason: it inscribes in real time, without requiring the human effort that is in shortest supply at the end of a construction project.

The technology components — photogrammetry, drone surveys, design model comparison — are individually mature. The model management layer that would make continuous camera-based documentation practically manageable is not. Addressing this bottleneck — building the version control, deviation tracking, and conflict resolution infrastructure that continuous capture pipelines require — is the critical development investment that will determine whether camera-first construction documentation fulfils its architectural promise in the coming decade.

References

[1] Akrich M. The de-scription of technical objects. In: Bijker W, Law J (eds.) Shaping Technology/Building Society. MIT Press, 1992: 205–224.

[2] Callon M. Some elements of a sociology of translation. In: Law J (ed.) Power, Action and Belief. Routledge, 1986: 196–223.

[3] Latour B. Science in Action. Harvard University Press, 1987.

[4] Golparvar-Fard M, Peña-Mora F, Savarese S. D4AR — a 4-dimensional augmented reality model for automating construction progress monitoring data collection, processing and communication. Journal of Information Technology in Construction 2009;14:129–153.

[5] Bosché F, Ahmed M, Turkan Y, Haas CT, Haas R. The value of integrating scan-to-BIM and scan-vs-BIM techniques for construction monitoring using laser scanning and BIM. Automation in Construction 2015;54:1–14.

[6] Macher H, Landes T, Grussenmeyer P. From point clouds to building information models: 3D semi-automatic reconstruction of indoors of existing buildings. Applied Sciences 2017;7(10):1030.

[7] Davila Delgado JM, Oyedele L, Demian P, Beach T. A research agenda for augmented and virtual reality in architecture, engineering and construction. Advanced Engineering Informatics 2020;45:101122.

[8] McKinsey Global Institute. Reinventing Construction: A Route to Higher Productivity. McKinsey & Company, 2017.

[9] Waern N. Designing the Life Programmable Interface: A Modular Architecture for Personal Health AI Sovereignty. Zenodo, 2025. DOI: 10.5281/zenodo.17462962.

[10] Waern N. Edge-Native Biological Digital Twins: Architecture Principles for Local-First Personal Health Intelligence. Zenodo, 2025. DOI: 10.5281/zenodo.17464804.

[11] Waern N. SMILE: A Methodology for Impact-First Digital Twin Deployment in Complex Sociotechnical Systems. Zenodo, 2026. DOI: 10.5281/zenodo.19587944.