1. 01

   Capture

   Walk the lot once with the handheld scanner. It records lidar, camera and centimeter GPS together.

   Unitree L2 lidar · ELP AR0234 camera · Jetson Orin Nano Super · Unicore UM980 and UM982 RTK

   • A base on a tripod sends GPS corrections over a 915 MHz radio to the rover on the rig.
   • Hold still 10–20 s in the dock, walk the lot, come back and hold still again.
   • The 5-inch screen shows a live map and the GPS fix while it records.
   • Each walk is saved as a ROS 2 bag with its camera calibration and GPS log.
2. 02

   Process

   The app stitches the walk into one 3D scan, and its AI picks out the pavement, paint, curbs and symbols.

   Pavement Pro for Windows · FAST-LIO · SAM 3

   • One command pulls the walk off the rig over Wi-Fi.
   • It runs on the estimator’s own machine, not in the cloud.
   • On three real lots, each walk closed to within 0.7–3.8 cm of its start once pinned to the GPS.
   • The steps inside are the data pipeline, below.
3. 03

   Verify

   The estimator checks what was found in 2D and 3D, and fixes or adds anything missed.

   Plan view · 3D views: LiDAR, LiDAR + camera, Mesh

   • Every item sits in a takeoff list by bid item, laid out like the Bluebeam markups estimators already use.
   • Slope, point density and height spread show where the scan is thin or the ground uneven.
   • Patching, potholes and crack sealing are drawn by hand for now.
4. 04

   Export

   The quantities come out as a spreadsheet: square feet, linear feet and counts, by bid item.

   CSV

   • Sealcoat in SF.
   • Striping in LF and SF, by line width.
   • Stalls counted by length; symbols by type.
   • Paving and patching in SF.
5. 05

   Price

   Planned

   The contractor’s own rates turn the quantities into the bid.

   The contractor’s unit rates

   • Today the CSV is priced in the estimator’s own spreadsheet or bid software.
   • Next: rates by bid item in the app, and a priced bid PDF with a repair plan.

1. 01

   Upload

   Bring the walk into the app and make it one cloud, pinned to the GPS.

   A ROS 2 bag from the rig, or LAS and LAZ from any scanner

   • Every file is checked against the rig’s SHA-256 checksums, and the capture’s health is reported.
   • The SLAM path is fitted to the RTK fixes, finding the clock offset and heading, and the cloud is leveled.
   • Returns past a 6–8 m range limit are dropped, where the lidar’s noise grows.
2. 02

   Color

   Paint every lidar point with the color the camera saw there.

   The calibrated camera · its measured position on the rig · a 0.6 s frame lag

   • Each point is projected into the camera frames that saw it and takes their color.
   • The lens, the camera’s position against the lidar and the frame timing were each measured on the rig first.
   • On real lots, 91–96% of points are seen, by about 11 frames each; the rest stay an exact gray.
3. 03

   Segment

   SAM 3 labels the camera images, and the labels are laid over the scan.

   SAM 3, fine-tuned on GeoSynth’s simulated lots (examples)

   • SAM 3 masks every frame, plus a top view of the whole lot: pavement, lines, curbs, symbols, cars and trees.
   • The masks are voted onto the lidar points and checked by geometry: paint lies on the ground, poles stand up.
   • Geometry tidies the result, straightening stall lines to 4 in, and it lies over the scan as a 3D mesh, one object a class.
4. 04

   Edit

   Turn the labels into takeoff items, then fix what is wrong or missing.

   Find features · the plan view and 3D scene

   • Find features measures paved area, lines by width, stall lines by length, symbols and hatching, inside a site boundary.
   • Add all puts them in the takeoff, each one listed by bid item to click and inspect.
   • The estimator fixes what is wrong and draws what was missed, on the plan or in 3D.
5. 05

   Export

   Write the bid quantities out as a spreadsheet.

   CSV

   • The Bid tab sums the takeoff by bid item, in SF, LF and EA.
   • Export writes that table as a CSV.

Render Mask

Render Mask

Render Mask

The cars are the “Generic passenger car pack” by Comrade1280, licensed under CC BY 4.0. GeoSynth split them into single cars, rescaled and repainted them.

1. May – Jun 2026

   Parts on a bench

   No enclosure yet: a Jetson Orin Nano, a Unitree L2 lidar with its own IMU, a global-shutter camera, a 140 W power bank and a 5-inch touchscreen, running ROS 2, FAST-LIO and a kiosk that records each walk.

   What changed: the lidar moved from USB-serial to Ethernet, after serial dropped frames whenever the power dipped; an external IMU was dropped, since the L2’s own is already in step with its points.

2. Jun 2026

   The concept sketch

   The lidar tilted forward 60° to see the ground ahead, the camera below it, the Jetson and screen behind, a handle on an aluminum extrusion and the battery at the bottom. One rule: neither sensor may see the other.

   

3. Summer 2026

   The first shells, v40 and v41

   A printed shroud on a 150 mm rail, modeled by hand: the lidar on a 60° face, the camera below, the Jetson and screen back to back, the battery in a shoe under the rail.

   Why it was replaced: both sensors hung off a printed box with a screen cut-out in the load path, so any flex would move the camera against the lidar and smear the color. The model had no parameters or joints, so a rebuild was cheaper than a repair.

   

4. Sep 19 – 21, 2026

   v2 and v3: rebuilt in code, on a rail

   A new design built only by scripts, with a 250 mm aluminum rail as the datum: the lidar and camera on one rigid member, the battery under the rail. Placing the makers’ own 3D models found clashes the part checks had missed, such as the rail arm passing through the Jetson’s board. v3 then rebuilt it into an empty document from the scripts alone, every part matching its reference volume exactly.

   

5. Sep 22 – 25, 2026

   v4: one printed module

   One shroud around a rail sleeve instead of four parts, the lidar back at 60° to see the ground from about 0.7 m ahead.

   What printing it taught: the layout worked, but the Jetson’s ports sat about 80 mm down inside the module, hard to plug, and the fastener sizes needed fixing.

6. Oct 4 – 16, 2026

   v5: built like a cordless drill

   No rail. The handle and battery are one unit under the head, and it stands on the battery when set down. The tray is the door, with the Jetson turned so its ports face a plug bay. Remade from parts already on hand and assembled by 10-14: nine printed parts, about 585 g as printed. “Everything fits into place and works well.”

   Proof it can be rebuilt: on a second machine the scripts rebuilt v5 with 63 of 63 checks the same and all nine print files byte-identical (E-055).

   

7. Oct 16 – 26, 2026

   The dock, then RTK on the rig

   A flat printed dock marks every capture: start still in it, walk, return to it, so any drift shows between the two still moments. The RTK rover was zip-tied to the tray for the first fixed captures on 10-25; the next day the tray was redrawn with the rover’s mount built in.

   

The RTK pods

A base and a rover for centimeter-grade positions, so the SLAM path can be pinned to the ground. The base is a Unicore UM982 with a helical antenna on a tripod; the rover is a UM980 with its puck screwed straight on, riding the rig. A pair of 915 MHz LoRa radios broadcasts the base's RTCM corrections (GPS, Galileo and BeiDou observations every second).

Pod v3 is plug-and-play: each receiver is powered and read over its own USB-C (the Jetson on the rig, a power bank on the tripod), and the radio runs from the receiver's 8-pin over one pre-crimped link. No soldering, no stripping.

The kiosk logs every fix into the capture; Process fits the walk to them, finding the clock offset and heading. The first fix on the rig, on 10-25, read horizontal sigmas of 1.0–1.4 cm and vertical 2.7–3.6 cm.

Date	Version	Why it changed
10-18	Receivers bench-checked	First data off both
10-22	v1 draft: receiver on edge, bank inside, a jumper to the puck	From measured parts
10-24	v2: the puck screws onto the UM980, one flat body 35–40% smaller, no supports	The jumper's threads could not join puck and receiver
10-24	v3: no power wiring inside, 86.8 × 48.5 × 32.8 mm, 44% of v2's footprint	28–30 AWG soldering was the hard part; the goal was under 20 min a pod
10-25	Base as one body on the tripod screw; RTK fixed on the rig	Simpler to set up

The pods in 3D