3D scanning → agent-usable house data — options & plan

Distilled from the 2026-06 discussion. Raw scanner notes were in scanning.txt (now folded in here). House ground truth: HISTORY.md.

Why scan at all

The schematic was assembled without accurate ground truth: the floor plan is explicitly “not to scale”, and the survey is a blurry JPEG good only for the footprint envelope. Because the levels are registered to each other, an error in the basement propagates through the whole house. A scan would give an independent geometric reference — something to trace against and overlay-check, rather than a replacement for the schematic. It does not replace the model — it becomes a reference you trace against and overlay-check: a tool does the math, a vision-capable agent does the interpretation.

Structured output vs. mesh

Forget “which scanner” first. The real fork is what the output IS, because it changes the whole downstream pipeline:

The irony to internalize: the cheaper dedicated scanners give better geometry but less agent-ready structure; the phone/tablet gives weaker geometry but instant structure. The “scan a house, then ask the agent questions” dream is specifically the RoomPlan path. For this house right now, mesh is plenty, because the survey plus a small extraction step already supply the “walls” understanding.

A third category — photo-based (photogrammetry). Most Android 3D-scan apps (and any phone without a LiDAR sensor) reconstruct geometry from ordinary photos rather than a depth sensor. They’re the most accessible option — any camera works — but they produce an unscaled mesh: there is no metric reference, so scale must be set by hand from a known dimension, and fidelity trails both LiDAR and dedicated scanners, especially on blank/featureless walls. Useful for rough massing or a textured visual to eyeball, not for trustworthy wall dimensions. The clean way to place the three: RoomPlan = scaled structure, LiDAR / dedicated scanners = scaled geometry, photogrammetry = unscaled geometry. For our pipeline, a photogrammetry mesh can still be dropped in as a rough mock-up, but a survey/known-dimension scale step is mandatory before any measurement is trusted.

Accuracy

• iPad/iPhone LiDAR: ~±1–3 cm horizontal in good conditions, but peer-reviewed testing shows most points land ~10 cm off the reference on dynamic (walking) scans, and accuracy degrades as the area grows and pose changes accumulate. Great for one small room held still; shaky as one coherent whole-house capture.
• RoomPlan specifically reduces the room to rectangles — sloped/non-flat walls become primitive boxes, ceilings and height variation are ignored, and it does not give wall thickness. So the 2nd floor is out (angled knee-wall/roof geometry is exactly what it can’t represent), and wall thickness you’d still tape-measure (easy, and already being done).
• Reflective surfaces (mirrors, glass) distort LiDAR badly — mind the bathroom.
• Dedicated structured-light (Einstar/Revopoint) give higher raw fidelity and open meshes, but are optimized for objects; whole-room coherence depends on good tracking over featureless walls — verify room-scale tracking before trusting a big sweep.

Scanner options — updated 2026 prices, openness, reputation

(Prices move; treat as ballpark and re-check at purchase time. Sources at bottom.)

The agent pipeline

A scan becomes agent-usable through one small step: ingest the scan with a point-cloud library (Open3D in Python; CloudCompare for interactive cleanup), then turn it into something the model can act on. Three uses, cheapest-first:

1. Slice → traceable floor plan. Cut the point cloud at ~1.2 m height to get a true-to-scale 2D plan image, then drop it into the tracing workflow as a background, replacing the not-to-scale floor plan. Highest value, lowest effort — this is an established research pipeline (slice → iterative-RANSAC line detection → floor plan, >90% line accuracy).
2. Plane-fit → wall coordinates. RANSAC plane-fit the vertical surfaces to get exact wall positions, then overlay-check the model against reality and flag where it diverges.
3. Structured scan → walls directly. RoomPlan output becomes walls with minimal translation. Powerful but rectangle-limited; best for clean single-story rooms.

This is the literal substrate for the “scan a house with a tablet, then ask the agent questions” idea: the scan is the perception input, a deterministic script does the geometry, and the agent reasons over the result.

Recommendation

For this house, right now: don’t buy anything on the critical path. The survey already anchors the footprint core, and room-by-room is enough — the only unmeasurable spot left is the crawlspace under the kitchen (a deferred TODO).

• If anyone can borrow a RoomPlan-capable Apple device for an afternoon: scan just the crawlspace, export both the structured model and a point cloud/.obj. Then build the ingest→slice→plane-fit tool and prove the whole pipeline on the one area that can’t be tape-measured. Best ROI, zero spend, unblocks a real TODO.
• If nobody has Apple gear: the crawlspace is tape-measurable with patience, so no purchase blocks this house. Don’t buy a $450 iPad for one crawlspace.

For the “scan-a-house-and-ask-the-agent” idea: this is the RoomPlan path (structured output is the magic), so an Apple device is on that roadmap — but pair it with the extraction pipeline to escape RoomPlan’s rectangle/no-thickness limits. If it grows toward whole-house coherent scans, that’s where a Revopoint Miraco (open, standalone, ~$1.6k) or a used Faro/BLK360 earns its keep. Make “open formats” a hard requirement evaluated per-vendor, not a nice-to-have.

The surprise to weigh: the Einstar Vega at ~$500–759 undercuts the refurb iPad and is open + standalone — so if the goal is “best open geometry per dollar” rather than “RoomPlan’s structured walls,” it may be the better first hardware buy. The question deciding iPad-vs-Vega is structured-output (agent-ready walls) vs. best-open-mesh-per-dollar.

Open questions

• Does the friend already own a RoomPlan-capable Apple device? (Changes the right-now answer entirely — borrow beats buy.)
• Is whole-house coherence needed soon, or is room-by-room fine for the foreseeable future? (Forks “phone/Vega is fine” vs. “tripod LiDAR.”)
• Is “open formats” a hard requirement for the business, or negotiable? (Rules vendors in/out — especially the Leica’s Cyclone-register step.)
• Android: still no mainstream phone with comparable LiDAR plus a RoomPlan-equivalent — worth a periodic re-check, but not a 2026 option. The photogrammetry apps (above) are the Android fallback: accessible but unscaled and lower-fidelity, so they don’t change the recommendation.

Sources

• Revopoint Miraco — 3DTechValley review, 3DSourced, Digital Camera World
• Einstar Vega — 3DTechValley review, Dynamism, 3DPrinting.com sale; Einstar lineup — 3DTechValley Einstar 2
• Apple RoomPlan limits — Volpis overview, it-jim “Awful and Great”, ScanManifold
• iPad Pro LiDAR accuracy — ScienceDirect indoor-mapping eval, MDPI vs industrial scanner, Tandfonline vs TLS
• Point-cloud → floor-plan pipeline — ScienceDirect iterative-RANSAC, MDPI auto-2D-CAD, HPI 2D/3D floor plans
• Leica BLK360 G1 used — Machinio listings