Quick answer: Holding a model in correct alignment with a partially built site is a measurable engineering problem, not a setting. Achieving a vendor's documented registration accuracy inside its validated envelope is generally competent professional integration. A core R&D activity may sit where a stated positional tolerance, under stated site conditions, cannot be known in advance and only a systematic progression of work based on established science can determine it. You self-assess.
18 August 2026 — this article describes the current rules. The 2026–27 Federal Budget announced proposed R&DTI reforms for income years starting on or after 1 July 2028. Until any amendments take effect, the R&DTI continues to be administered under the current legislation.
Ask a site engineer why the AR overlay is "wrong" and the answer is a single number: the insert is out by 40 mm. Ask where the 40 mm came from and the question stops being about augmented reality. It becomes a question about survey control transfer, transformation residuals, inertial bias, heading error over a lever arm, and whether the model matches the concrete that was poured. That decomposition is where the research question lives.
Scope: This article covers only registration and alignment accuracy, not the wider immersive site workflow (device rollout, content authoring, digital-twin synchronisation, VR design review, crew training), which our companion Insight on AR and digital twins in construction deals with on the Insights index.
The Statutory Test the Question Has to Pass
Core R&D activities are experimental activities whose outcome cannot be known or determined in advance on the basis of current knowledge, information or experience, but can only be determined by applying a systematic progression of work that is based on principles of established science and proceeds from hypothesis to experiment, observation and evaluation, and leads to logical conclusions; and that are conducted for the purpose of generating new knowledge, including new knowledge in the form of new or improved materials, products, devices, processes or services (s 355-25(1), Income Tax Assessment Act 1997; business.gov.au).
Registration work is unusual among software-adjacent activities in that the established-science limb is both easy to satisfy and easy to fail — and teams rarely notice which they have done.
Registration Accuracy Is an Error Budget, Not a Setting
The observable failure — the overlay is in the wrong place — can reflect several contributing error terms. Separating them turns "AR doesn't work here" into something with a hypothesis attached.
Error term
What it actually is
Where it is normally solved
Control transfer
Getting the site frame into the space, and the uncertainty the network adjustment carries
Established survey practice
Alignment residual
The fit between device frame and site frame — a rigid-body transformation from a few correspondences
Established estimation theory
Tracking drift
Visual-inertial error over time and distance: scale error, gyro bias, angular random walk
Vendor-characterised, inside the vendor's environment
Reference error
The model is right and the building is not — a set-down poured low, a wall 25 mm off line
Survey and as-built, not tracking
A fifth term sits outside the device: the pose can be correct and the operator still sees it wrong, through camera-to-display calibration, parallax and viewing angle. An activity aimed at "improving AR accuracy" without knowing which term dominates has no hypothesis, because it cannot state what would falsify it. And the terms interact in ways the vendor's envelope does not describe: a third of a degree of heading error is invisible at arm's length and is roughly 70 mm of lateral displacement at 12 m.
Site Conditions Are What Put the Problem Outside the Validated Envelope
Commercial spatial-computing systems may have documented performance assumptions or validation conditions that do not fully cover the changing conditions encountered on a construction site. A building under construction is none of those, and each departure is statable:
No GNSS: Below ground there is no absolute fix, so accuracy is carried in from surveyed control and held by relative tracking.
Degenerate geometry: Scan-to-model relocalisation in a long, featureless corridor is under-constrained along the corridor axis — the fit is free to slide, and it does.
Changing geometry: Walls, bulkheads and services go in between passes, so the reference itself is a moving variable.
Feature collapse and interference: Screeded floors and stripped soffits give feature-based tracking little to hold; reflective insulation, wet slabs and direct sun corrupt depth returns.
Competent Professional Integration Versus an Outcome That Cannot Be Determined in Advance
The line is not the difficulty of the work or its novelty to the team. AusIndustry's guidance on AI-related activities makes the point in a neighbouring field: "Using an AI model or technique that is new to you does not, by itself, mean the activity is eligible for the program" (business.gov.au) — guidance that applies directly where a registration stack uses learned components such as a depth estimator or a relocalisation network.
Work generally unlikely to be core R&D on these facts:
• Integrating an AR SDK against its documented interfaces, and following its registration procedure — marker placement, control-point alignment, re-anchoring intervals — to reach the accuracy the vendor documents in conditions the vendor validated.
• Tuning parameters where it is known the issue can be solved by parameter changes and the effect of the change is well understood.
• Routine checks that the rig still meets a stated accuracy where the outcome is already known or determinable (testing undertaken to comply with a statutory requirement or standard falls under the s 355-25(2)(f) exclusion).
• Reproducing an existing commercial registration method (may raise the reproduction exclusion in s 355-25(2)(g) where directed to reproducing a commercial product or process from specifications or public information).
The guidance describes the other side in terms that transfer cleanly: activities may meet the requirements of a core R&D activity "where a technical hurdle exists and an expert in the field considers that only experimentation will determine if a proposed solution, or the way to develop a solution, can resolve the technical hurdle." On registration, that hurdle is a number and a condition: whether a given tolerance is holdable, at a given range, for a given duration, in an environment outside what the field's published knowledge covers.
What "Based on Principles of Established Science" Means Here, Concretely
This limb is where immersive-technology claims most often fail. A registration experiment rests on principles of established science when it draws on bodies of knowledge such as:
Least-squares adjustment and error propagation: From geodesy and survey — how observation uncertainty combines, and what a residual means.
Absolute orientation and rigid-body transformation estimation: Methods for fitting one coordinate frame to another, and their known degeneracies.
Point-set registration convergence: Why iterative-closest-point style fits fail on under-constrained scenes.
Inertial sensor error models: Bias instability, angular random walk, scale factor error, and how each propagates.
Optics and photogrammetry: Lens distortion, exposure and motion blur as measurable inputs to feature quality.
A team that varies settings, watches whether the overlay looks better and repeats has run a disciplined loop with no basis in established science, and does not satisfy the limb however many iterations it logs. A team that predicts from an inertial error model that heading error will dominate beyond 8 m, then measures residuals to test that prediction, is doing something the limb recognises — and still has to satisfy the unknown-outcome and new-knowledge requirements. The equivalent boundary in software sits on our R&D for software and AI page.
A Hypothetical Worked Example
Illustrative and hypothetical only. It is not a ruling, is not based on any client, and the facts below do not establish that anything would be eligible.
An Adelaide mechanical services contractor is setting out hanger inserts from a federated services model on AR tablets, across two basement carpark levels and a plant room of a mid-rise building. No GNSS below ground; roughly 90 m of corridor between the two surveyed control marks; block walls going up between passes; 60–150 lux on temporary work lights; and a plant room dense with reflective ductwork.
Baseline: The SDK documents positional drift below 0.5% of distance travelled in a "well-featured, static indoor environment", with relocalisation validated on a static test set. Spot checks on the previous level found 3D error at the point of use anywhere between 18 mm and 70 mm, with no stated dependence on anything.
Target, fixed on 2 March before any trial: 3D positional error ≤ 15 mm at 95% of check points, at up to 12 m from the nearest anchor, sustained across a 45-minute session, under the stated lighting and geometry conditions. It came from the ±20 mm insert set-out tolerance less an allowance for hand placement.
Ground truth & control: 40 check prisms observed by total station from resected stations on four adjusted control marks — uncertainty about 2 mm, small enough against 15 mm to test the target. Held constant: Tablet unit, SDK version, model version, control network adjustment, operator, walking-speed protocol. Varied one at a time: Anchoring method, session duration, elapsed geometry change since the last wall placement; lighting in a separate block of runs.
Approach A — single surveyed anchor at stair core (Failed): Median 3D error was 9 mm within 6 m of the anchor, but the 95th percentile reached 61 mm at 12 m. Residuals showed growth was almost entirely angular: ~0.35° heading error, roughly 70 mm lateral displacement at 12 m. Eliminated: The hypothesis that drift was translation-scale; ruled out single-anchor registration beyond 6 m here.
Approach B — two anchors 14 m apart + coarse-scan relocalisation: 95th percentile fell to 24 mm and held ~20 min. Once block walls went up, relocalisation failed to converge in the corridor and slid along the axis by up to 90 mm on two runs. Eliminated: Scan-based relocalisation against corridors whose wall geometry changes between passes.
Approach C — two anchors + column targets + bias-corrected inertial model: 95th percentile reached 13 mm at 12 m across 45 minutes in the carpark. In the reflective plant room, it sat at 19 mm.
Result actually reached: Target met in carpark conditions, not met in the reflective plant room (left unresolved). The knowledge generated was the error decomposition itself: heading error dominates at range, and column targets constrain it where scan relocalisation fails.
Activity boundary: For this example, the candidate experimental activity is documented from the hypothesis formulation through to trials and evaluation (~240 hours logged against ~1,900 project hours). Using Approach C as the standard set-out procedure on remaining levels is ordinary delivery, not experiment.
Supporting Activities Around a Registration Experiment
Activities that are not themselves core may be supporting R&D activities where directly related to core R&D activities. Under s 355-30(2), where the activity is of a kind referred to in s 355-25(2), or produces goods or services, or is directly related to producing goods or services, it is a supporting R&D activity only if undertaken for the dominant purpose of supporting core R&D activities (business.gov.au).
The awkward item is the total station work. Weekly control observations required for ordinary production set-out may be directly related to producing services, depending on the particular activity, and should be assessed on their particular facts. The fact that the same observations are also used as ground truth does not, by itself, determine their character for R&DTI purposes. What is distinguishable is the incremental work — check prisms observed only to test the target, a second pass at another time of day to isolate the lighting variable. How the expenditure is treated is for the company and its registered tax agent. Ineligible work generally is covered on our what does not qualify page, and the vertical picture on our property and construction page.
Where an RSP Fits
AusIndustry describes Research Service Providers as scientific or technical service providers, registered in specific fields, that a company can engage to conduct R&D activities on its behalf (business.gov.au). Qualifying expenditure incurred to a non-associate RSP may still be taken into account in determining R&D tax offset entitlement where total notional deductions are below the usual $20,000 threshold, provided the services are within a research field for which the RSP is registered — see claiming R&D under $20,000 and what an RSP is — and using an RSP does not guarantee eligibility — you still self-assess. Offset rates and how they apply are covered on our refundable vs non-refundable offset page.
Ignition Research works with construction and construction-technology teams on the research side of a registration problem — decomposing the error budget, and fixing the measure and the conditions before any trial runs. As a Registered Research Service Provider at Lot Fourteen in Adelaide we supply and structure research capability; we are not a registered tax agent, and your company self-assesses and remains responsible for its own claim. Get in touch.
Frequently Asked Questions
Q: Is improving AR registration accuracy on a construction site R&D?
A: Not automatically. Reaching the accuracy an SDK documents, in conditions its vendor validated, is generally competent professional integration and is generally unlikely to be a core R&D activity on those facts, subject to the activity's own facts and the statutory tests. A core R&D activity may sit where whether a stated tolerance is holdable, at a stated range and duration, in stated site conditions outside the validated envelope, cannot be known or determined in advance — and is pursued through a systematic progression of work based on principles of established science, for the purpose of generating new knowledge. You self-assess.
Q: What causes AR drift and misalignment on building sites?
A: It normally reflects several contributing error sources rather than a single fault: uncertainty carried in from survey control, the residual of the rigid-body fit between the device frame and the site frame, visual-inertial drift over time and distance, divergence between the model and the as-built structure, and camera-to-display calibration and parallax. Angular error is often the term that dominates at working range, because a fraction of a degree of heading error becomes tens of millimetres over 10 metres.
Q: Why does AR alignment fail indoors and below ground?
A: There is no GNSS fix, so absolute position has to be carried in from surveyed control and then held by relative tracking, which accumulates error. Long corridors are geometrically under-constrained, so scan-based relocalisation can slide along the corridor axis; stripped soffits and screeded floors give feature-based tracking little to hold; and reflective or wet surfaces corrupt depth returns.
Q: What does "based on principles of established science" mean for AR tracking work?
A: It means the systematic progression of work rests on an established body of scientific knowledge — survey error propagation and least-squares adjustment, rigid-body transformation estimation and its known degeneracies, inertial sensor error models, optics and photogrammetry — rather than on iterative trial and observation alone. It is one limb of the core R&D activity definition in s 355-25(1), alongside proceeding from hypothesis to experiment, observation and evaluation and leading to logical conclusions, the outcome being unable to be known or determined in advance, and the purpose of generating new knowledge.
Sources & Further Reading
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. ss 355-25 and 355-30
Related: property and construction · what does not qualify · software and AI · what an RSP is · under $20,000 · offset rates · more Insights
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
Thinking about a project like this?
If you're weighing up an AI, software or technical improvement project and can't tell yet whether it's implementation or research, start with a quick read on where it sits.