Evidence before claims.

From a robot patrol to a semantic work-zone digital twin

45 fixed runs: the loop patrol achieved 78.7% mean reconstruction completeness, 0.717 actionable mIoU, and 82.1% known coverage.

The known-pose procedural study fuses multi-view 3D observations, detects temporary traffic-control devices, infers only supported boundary segments, and preserves unknown space. Its preregistered false-traversable-rate expectation was not supported; the single-pose zero occurred with only 9.9% known coverage. CARLA, SLAM, and hardware remain separate validation gates.

Executed notebook · truth-isolated prediction/audit artifacts · 11/11 validation checks · 11 unit tests

Measured against an explicit acceptance criterion

Median post-restoration error was 4.1 cm; the registered acceptance threshold was 10 cm. Trial-level CSV results and generation code are retained with the proposal evidence package.

Two-pass change detection recovers a 0.84 m displacement to within 15 mm

Displaced device measured at 0.825 m against 0.840 m ground truth; removal flagged; seven undisturbed devices moved 0.000–0.027 m — no false positives at a 0.30 m gate.

Ten channelizing devices in a simulated lane closure were scanned from a four-station patrol, perturbed (one displaced, one removed, one toppled), and rescanned; association between passes is purely geometric. The registered failure is reported, not tuned away: the toppled device shifted its centroid only 0.274 m — below the gate — so centroid-only tests miss topples, motivating a height/orientation feature. Device points come from the simulator's semantic channel, so this validates the matching stage, not the detector.

64-channel simulated LiDAR · 2 cm ranging noise · RTX PRO 4500 · preliminary evidence for a university-led USDOT DOT Bots Challenge Stage I concept paper currently in preparation (not yet submitted; August 2026)

LiDAR-anchored colorized twin from a six-station patrol

118,235 points after 6 cm voxelization; 75,784 colorized from five posed patrol images; devices, barricades, equipment, and the retaining wall all metrically placed.

Patrol LiDAR is aggregated in the world frame at known poses and colorized by z-buffered projection of posed RGB — a deliberately rigid baseline for the twin. This is not joint optimization: the registered next gate upgrades it to LiDAR-anchored Gaussian-splatting reconstruction with joint pose refinement, following recent methods that initialize splats from LiDAR and refine camera poses against it. The staged closure itself was audited from a bird's-eye view for MUTCD plausibility (advance sign, shoulder-to-lane-line merging taper, tangent line, barricaded work area) before capture.

CARLA 0.9.15 · six LiDAR stations inside the protected corridor · five posed cameras · scene manifest and generation scripts retained

Why the payload carries LiDAR and cameras: night is the hard case

The same closure imaged by day and at night in an unlit work zone, plus the 64-channel LiDAR return at the night pose: the camera loses the far devices to darkness while LiDAR geometry is unchanged, since it supplies its own illumination. Cameras remain necessary for colour, legend, and lit-status information that geometry cannot provide. Simulated LiDAR omits retroreflective and wet-surface effects; field validation remains a separate gate.

CARLA 0.9.15 · identical pose day/night · supports unattended night-patrol feasibility, when worker and traffic exposure is lowest