Does a loop reconstruct more?
Compare occupied-voxel completeness under matched layouts and seeds.
Technical summary
What was tested—and why
A road work zone may be straight, tapered, curved, or partially closed. A robot cannot safely treat an unfamiliar layout as an ordinary empty road. This pilot tests one bounded evidence chain: patrol the protected side, fuse repeated 3-D observations, recognize channelizing devices and obstacles, estimate only the device-supported boundary, then label observed cells as traversable, inspect, no-go, or unknown. Unknown remains an explicit outcome when the patrol has not supplied enough evidence.
The study compares a single scan, a one-way patrol, and an out-and-back loop on three illustrative boundary shapes and five fixed seeds. The implementation deliberately uses known poses so that this first test isolates viewpoint coverage, semantic fusion, boundary inference, and operational-map lowering. Localization error is reserved for the next experiment.
3 × 5 × 3layouts × fixed seeds × patrol strategies
4 statestraversable · inspect · no-go · unknown
Truth-isolatedprediction API cannot receive the reference boundary, polygons, or object IDs
Does more coverage improve the map?
Measure three-class actionable mIoU while counting unknown cells as misses.
Does the model invent safe space?
Report the ground-truth no-go cells incorrectly labeled traversable as a separate failure metric.
Does performance depend on layout?
Expose every layout–strategy cell instead of relying on one overall average.
Research gap
The literature supplies components, not the full chain
Work-zone datasets, metric-semantic mapping, traversability estimation, and navigation masks are all relevant—but they answer different questions. The study keeps those layers separate.
| Primary source | What it contributes | What remains open here |
|---|---|---|
| WorkZone3D · WACV 2026 | Multimodal 3-D boxes for small work-zone channelizers and evidence for camera–LiDAR fusion. | Dense robot-centric reconstruction, traversability, and operational authority. |
| ROADWork · ICCV 2025 | Work-zone-specific recognition, scene interpretation, and video pathway prediction. | Dense 3-D ground truth for a mobile inspection patrol. |
| Kimera · ICRA 2020 / Voxblox · IROS 2017 | Metric-semantic meshes and distance-field mapping for planning. | Work-zone semantics do not themselves authorize a corridor. |
| STEP · RSS 2021 / TNS · RSS 2022 | Uncertainty-aware and construction-terrain traversability. | Temporary traffic-control policy near live traffic. |
| USDOT WZDx v4.2 / Nav2 Keepout Filter | Interoperable work-zone records and enforcement of a supplied keep-out mask. | Neither replaces local sensing or decides that an inferred mask is authoritative. |
The downloadable literature matrix contains 13 primary technical and government sources, their contribution to this design, and the claim each source does not support.
Methods
A five-stage, auditable evidence chain
Operational ontology
| Output state | Pilot rule | Deployment interpretation |
|---|---|---|
| TRAVERSABLE | Observed, boundary-supported road cell on the protected-side prior. | A mobility candidate; still subject to robot capability and approved corridor. |
| INSPECT | Observed cell near an inferred TTC boundary or detected device. | A perception target, not permission to manipulate the device. |
| NO-GO | Observed live-side road cell or directly observed obstacle buffer. | Must be fused with the approved TTC plan/supervisor constraints. |
| UNKNOWN | Insufficient observation or an unsupported TTC-boundary interval. | Abstain, collect another view, or request human review. |
Experiment matrix
- Layouts: right-lane closure, shoulder work area, and lane-shift chicane.
- Seeds: 20260808–20260812, reused across patrol strategies.
- Strategies: single pose ⊂ one-way patrol ⊂ loop patrol.
- Representative case fixed before results: right-lane closure, seed 20260808, loop patrol.
Primary metrics
- Completeness: reconstructed reference occupied voxels / all reference occupied voxels.
- Actionable mIoU: unweighted IoU of traversable, inspect, and no-go over the full ROI.
- False-traversable rate: no-go truth cells predicted traversable / all no-go truth cells.
- Unknown predictions reduce mIoU; confidence intervals summarize only the 15 fixed runs per strategy.
Results
Repeated viewpoints changed both coverage and the operational map
All numbers below are read from saved per-run evidence. No run was dropped. The intervals are descriptive 95% Student t-intervals across three layouts × five seeds, not population-level safety uncertainty.
Preregistered comparison. The loop patrol increased mean reconstruction completeness from 7.1% to 78.7% and actionable mIoU from 0.086 to 0.717 relative to a single pose. The third expectation was not supported: loop mean no-go→traversable rate was 0.084%, above the single-pose value of 0.000%. That zero accompanies only 9.9% known coverage—unknown cells are excluded from this failure rate—so it is not evidence that the single pose was safer.
| Patrol strategy | Completeness | Actionable mIoU | No-go → traversable | Known coverage | TTC precision / recall |
|---|---|---|---|---|---|
| Single pose | 7.1% [6.9, 7.3] | 0.086 [0.074, 0.099] | 0.000% [0.000, 0.000] | 9.9% [8.6, 11.2] | 100.0% / 23.6% |
| One-way patrol | 62.7% [61.7, 63.6] | 0.687 [0.667, 0.707] | 0.108% [0.023, 0.193] | 80.4% [78.7, 82.0] | 95.4% / 100.0% |
| Loop patrol | 78.7% [77.9, 79.4] | 0.717 [0.699, 0.735] | 0.084% [0.025, 0.142] | 82.1% [80.3, 83.9] | 98.5% / 100.0% |
Values are means; brackets are descriptive 95% t-intervals across 15 fixed runs, except TTC precision, which is the mean per-run auxiliary precision. Percentages are rounded for display; full precision remains in the downloadable CSV.
Result interpretation
The return pass improved mean actionable mIoU over the one-way patrol in every tested layout: 0.733 for the right-lane closure, 0.679 for the shoulder work area, and 0.739 for the lane-shift chicane. The lower shoulder result remains visible rather than being hidden by the overall mean.
One-way and loop patrols both reached 100% mean TTC recall, but mean precision was 95.4% and 98.5%; some runs over-detected devices. The loop reduced mean boundary MAE from 0.116 m to 0.056 m relative to one-way. The single pose had a smaller 0.029 m MAE only over 14.4% mean boundary support, so the error cannot be read without coverage.
Discussion
What the pilot supports
Supported conclusions
- Across all three illustrative shapes, the loop achieved higher mean completeness, actionable mIoU, and known coverage than the single pose.
- Explicit unknown cells preserved abstention and exposed why false-traversable rate must be interpreted with coverage; the preregistered “loop FTR no higher” expectation was not supported.
- The saved evidence chain can be recomputed from raw observations, fused voxels, aligned BEV rasters, and diagnostic numerators/denominators.
Important limitations
- Known poses isolate mapping; localization drift and loop closure are not evaluated.
- Semantic labels are procedurally corrupted observations, not output from a trained network.
- Three illustrative geometries do not represent the distribution of real work-zone designs.
- No vehicle interaction, weather, terrain dynamics, robot gait, manipulation, or field safety case is modeled.
Why semantics are not authority
A cone boundary can be perceived, but perception alone cannot decide whether the robot is allowed on one side. The pilot freezes a protected-side coordinate prior so the inference can be scored. A real system must merge observed geometry and semantics with an approved temporary traffic-control plan, a site-supervisor corridor, or an equivalent authoritative source. That policy layer should override perception, and the navigation stack should preserve unknown space rather than convert uncertainty into permission.
Next validation gates
| Gate | Input / change | Evidence required before the claim expands |
|---|---|---|
| 1 · CARLA oracle-pose capture | Ordinary LiDAR as algorithm input; semantic/instance sensors only for ground truth. | Registration, device recall, boundary error, map quality, manifests, and synchronized sensor provenance. |
| 2 · Pose uncertainty / SLAM | Injected pose noise, then LIO-SAM or equivalent estimated trajectories. | Absolute/relative trajectory error, map degradation, loop-closure failures, and runtime. |
| 3 · Domain shift | Lighting, weather, device type/layout, sensor placement, and unseen maps. | Calibration, failure taxonomy, abstention/coverage tradeoff, and held-out performance. |
| 4 · Controlled hardware | Robot or quadruped in a closed mock work zone under supervisor-defined keep-out policy. | Repeatable survey quality, safe-stop behavior, intervention rate, and independently reviewed operating procedure. |
Downstream planning is a separate artifact
The natural interface is a versioned SceneContract: coordinate frame, map hash, TTC detections, supported boundary, traversability layers, unknown mask, and authoritative corridor constraints. A model-checked route planner may consume that contract later. The present experiment does not implement natural-language mission planning or validate the reconstruction-to-planner handoff.
Reproducibility
Inspect the evidence, not just the figure
The package preserves the frozen protocol, literature matrix, executable notebook, per-run metrics, aggregate tables, configuration manifest, validation record, and representative arrays.
Executed notebookTop-to-bottom analysis and metric recomputation · IPYNB
45-run metricsPer-run values and diagnostic counts · CSV
Strategy summaryMeans, dispersion, and descriptive intervals · CSV
Frozen protocolDesign, formulas, artifact contract, and claim boundaries · Markdown
Literature reviewPrimary-source synthesis and research gap · Markdown
Evidence bundleCode, configuration, tables, representative arrays, tests, and checksums · ZIP
Software and dependency versions are recorded in the experiment manifest. Procedural artifacts may be reproduced; they are not substitutes for CARLA or field measurements.
Selected references
Primary sources
- Sural, Sahu & Rajkumar. “WorkZone3D: A Multimodal Dataset for 3D Work Zone Perception in Autonomous Driving.” WACV 2026. Open-access paper.
- Ghosh et al. “ROADWork: A Dataset and Benchmark for Learning to Recognize, Observe, Analyze and Drive Through Work Zones.” ICCV 2025. Open-access paper.
- Behley et al. “SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences.” ICCV 2019. Dataset and paper.
- Shan et al. “LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping.” IROS 2020. Paper.
- Rosinol et al. “Kimera: an Open-Source Library for Real-Time Metric-Semantic Localization and Mapping.” ICRA 2020. Paper.
- Oleynikova et al. “Voxblox: Incremental 3D Euclidean Signed Distance Fields for On-Board MAV Planning.” IROS 2017. Paper.
- Fan et al. “STEP: Stochastic Traversability Evaluation and Planning for Risk-Aware Off-road Navigation.” RSS 2021. Paper.
- Guan et al. “TNS: Terrain Traversability Mapping and Navigation System for Autonomous Excavators.” RSS 2022. Paper.
- Wetzel et al. “The Use of Boston Dynamics Spot in Support of LiDAR Scanning on Active Construction Sites.” ISARC 2022. Paper.
- USDOT. “Work Zone Data Exchange (WZDx), version 4.2.” Specification overview.
- CARLA. “Sensors reference.” Documentation.
- Nav2. “Navigating with Keepout Zones.” Documentation.