PaperlayerView exact revision on arXiv

Source-linked AI summary

SceneFactory-3D: Lifting 2D Traffic Scenes into 3D Physical Counterfactuals for Scalable Physically Grounded Safety Evaluation

Yicheng Zhu, Linfeng Tian, Tianmu Zhao, Yang Chen, Fan Zuo, Tao Li, Zilin Bian

arXiv:2610.02874v1cs.ROcs.MA

TL;DR

Prescribed behavioral and kinematic rules leave out tire-road physics that can change vehicle execution and traffic outcomes under adverse conditions. SceneFactory-3D uses wheel-level physics and matched, GPU-batched counterfactual worlds to evaluate controllers across road conditions. A friction drop from 1.0 to 0.18 reduces safe work-zone clearance across learned policies and classical planners, while the tested grades have comparatively limited effects.

  • Problem

    Prescribed behavioral or kinematic execution does not capture how adverse road conditions alter vehicle motion and traffic outcomes, and planar models do not represent wheel-level effects of 3D geometry and gravity.

    Source: §1 · pp. 1–2

  • Method

    SceneFactory-3D couples per-wheel forces to spatially varying friction and 3D terrain, then compares matched worlds that differ only in road conditions while scenarios and controllers remain fixed.

    Source: §1 · p. 2, §3 · p. 5

  • Results

    A friction drop from 1.0 to 0.18 reduces safe work-zone clearance by 6–90 percentage points across learned policies and 18–19 points for classical planners; tested grades had comparatively limited effects.

    Source: §1 · p. 2, §7 · p. 10

  • Takeaways & Limitations

    Matched physical counterfactuals show that road friction can degrade task completion, increase interaction risk, and affect vehicles outside the altered lane.

    Source: §1 · p. 2, §7 · p. 10

    The simulator is verified but not calibrated, and interpreting a contrast for a real road requires validating the model.

    Source: §3 · p. 5, §6 · p. 9

Abstract

from arXiv · show

Scalable driving simulators typically execute vehicle commands using prescribed behavioral or kinematic rules, overlooking the physics of tire-road interfaces, thereby limiting their ability to capture how adverse road and environmental conditions alter vehicle execution and propagate through traffic. To address this limitation, we present SceneFactory-3D, a GPU-batched, physics-grounded multi-agent driving simulator. Vehicles execute acceleration and steering commands via suspension- and friction-limited forces evaluated at each wheel-contact point. Spatially varying friction, per-world 3D heightfields, gravity, and rigid contact consistently govern wheel motion and chassis collisions. Per-world terrain isolation and GPU batching enable SceneFactory-3D to run matched physical counterfactuals in parallel: traffic scenario setup and vehicle controllers remain fixed while only the road condition changes, enabling the resulting closed-loop effects to be evaluated across parallel worlds. To demonstrate the advantage of the SceneFactory-3D-enabled counterfactual evaluation, we conduct an empirical study on vehicle controllers' sensitivity to road conditions. We study three learned-policy families on 1,024 matched 12-vehicle worlds per condition, and two classical planners on a shared 32-world subset, across 21 friction and grade conditions. When friction drops from 1.0 to 0.18, the share of vehicles that clear the work zone safely falls by 6 to 90 percentage points across learned policies (18-19 for classical planners), and near-collision situations become more frequent for every learned policy. Code: https://github.com/SmallWorldLab/SceneFactory_3D

1 Introduction

Prescribed vehicle-motion rules can miss how adverse road conditions alter execution and affect traffic outcomes. SceneFactory-3D combines wheel-level road physics with matched counterfactual worlds to evaluate those effects across controllers.

1 Introduction

  • Prescribed behavioral or kinematic rules fix how vehicle commands are executed, so they do not capture how adverse road conditions alter vehicle motion and traffic outcomes.Planar friction limits also omit how 3D geometry and gravity affect wheel-level contact and available friction.Source: §1 · pp. 1–2
  • SceneFactory-3D compares matched worlds that differ only in road conditions, keeping the traffic scene, initial state, routes, and controller population fixed.This permits control over where conditions occur, whether controllers observe them, and how effects propagate through traffic.Source: §1 · p. 2
  • The simulator couples per-wheel forces and rigid contact to spatially varying friction and 3D terrain, with per-world terrain ownership keeping physical queries consistent.It also provides classical and learned controllers the same geometric route reference.Source: §1 · p. 2
  • A friction drop from 1.0 to 0.18 reduces the share of vehicles clearing the work zone safely by 6–90 percentage points across learned policies and 18–19 points for classical planners.Near-collision situations also become more frequent for every learned policy.Source: §1 · pp. 2–3

2 Related Work

Prior simulators span behavioral, kinematic, planar-dynamic, and physics-engine approaches, with different trade-offs in execution fidelity and traffic scale. SceneFactory-3D emphasizes matched physical interventions across many interacting vehicles and controllers.

2 Related Work

  • SUMO uses behavioral car-following and lane-changing; Waymax and GPUDrive use kinematic dynamics without tire forces or a friction limit, while CommonRoad offers several vehicle models and friction-circle trajectory checks.Source: §2 · p. 3
  • Physics-engine systems arrange traffic in different ways, while Chrono offers detailed vehicle and deformable-terrain models and SynChrono distributes agents across compute nodes.Examples include CARLA, SUMMIT, SMARTS, and MetaDrive [Dosovitskiy et al., 2017; CARLA, n.d.a,n; Cai et al., 2020; Zhou et al., 2021; Li et al., 2023b], and Chrono [Project Chrono Development Team, n.d.; Tasora et al., 2016; Project Chrono, n.d.].Source: §2 · pp. 3–4
  • SceneFactory-3D separates terrain-diverse training from physical preview and matches physical interventions across many interacting vehicles, unlike prior systems with flat roads or single-vehicle focus.SceneFactory-3D builds on SceneFactory [Zhu et al., 2026]; related efforts include µSync-GS [Chen et al., 2026] and MultiDrive [Kaufeld et al., 2025].Source: §2 · p. 4
  • Terrain randomization and physical-context conditioning are established learning techniques; SceneFactory-3D adds control over context placement, interacting populations, and matched execution.The cited techniques include randomized dynamics [Peng et al., 2018], physical-parameter conditioning [Yu et al., 2017; Kumar et al., 2021; Li et al., 2023c], and friction preview for trajectory tracking.Source: §2 · p. 4

3 Problem Setting

The problem setting treats a rollout as the result of a scenario, a controller assignment, and a physical road condition. Its physical contrast holds the scenario and controllers fixed while changing the road.

3 Problem Setting

  • A scenario specifies the planar layout, routes, vehicle models, and initial state; a physical condition specifies the road height field and friction patches, while controller assignment specifies each vehicle’s policy.In the benchmark, all vehicles use the same controller with shared parameters.Source: §3 · p. 4
  • The rollout equation makes the physical condition shape vehicle transitions and, for controllers given physical inputs, their observations.Vehicle and numerical parameters remain fixed, and the evaluation uses deterministic transitions and controllers up to floating-point effects.Source: §3 · p. 4
  • The physical contrast subtracts the reference-condition outcome from the changed-condition outcome while holding the scenario and controller assignment fixed.It measures the total closed-loop effect within the simulator; trajectories and commands may diverge through vehicle feedback and traffic interaction.Source: §3 · pp. 4–5
  • The contrast requires identical scenario and controller assignment on both sides and a transition that depends on the physical condition; the implementation batches worlds that evolve independently.Road forces and contact, controller-independent routes, and the evaluation budget guide implementation choices.Source: §3 · p. 5

4 SceneFactory-3D

SceneFactory-3D combines shared canonical routes, a force-based vehicle model, isolated per-world terrain, and GPU batching to run matched physical counterfactuals at scale.

4.1 Canonical routes

  • Canonical routes are planned from road layout rather than controller behavior, then checked for continuity, drivable bounds, closed-lane clearance, and curvature.The path checks do not establish whether a route can be executed at a given speed and friction level or by several vehicles simultaneously.Source: §4.1 · p. 5

4.2 Vehicle and road model

  • At each wheel contact, suspension and tire forces govern vehicle motion, with combined tire force limited by friction; requested acceleration or steering may not be realized on low-grip roads.The model uses a linear-slip, arctangent-slip-angle approximation with friction-circle saturation, rather than a Magic Formula or brush model.Source: §4.2 · pp. 5–6

4.3 World isolation and matching

  • Each world’s wheel queries, route and preview queries, and chassis collision geometry use its assigned terrain, while matching checks ensure paired runs differ only on the declared intervention axis.Worlds do not collide across boundaries, and matching is established at initialization rather than by branching a running traffic state.Source: §4.3 · p. 6

4.4 GPU batching

  • GPU batching steps the benchmark’s worlds together using 240 Hz physics and 60 Hz actuation, with controllers able to decide less often and hold commands.The benchmark comprises about 259,000 twelve-vehicle world rollouts; the held-out closure task compares reference roads with matched patch, slope, or combined counterfactuals.Source: §4.4 · p. 6, Figure 2 · p. 7

5 Experimental Setup

The benchmark tests five controller regimes on a two-lane closure task, comparing matched worlds across road conditions. Outcomes track valid passage and safety events, with worlds paired across conditions and controllers.

Task and routes

  • Each 12-vehicle world uses a straight two-lane road with a right-lane closure at 150 m; five vehicles merge from the closing lane while seven remain in the open lane.All vehicles execute through the physical vehicle model and continue interacting after crossing the counting line.Source: §5 · p. 7

Controller regimes

  • The five regimes comprise two classical planners and three learned-policy regimes, with paired-seed contrasts for terrain-diverse training and friction-and-grade preview.The Frozen L0 versus L0-T contrast also includes additional training, while L0-T versus L1 is matched in training.Source: §5 · p. 7

Physical conditions

  • The 21 conditions include a reference road, eight friction levels, eight grades, and four grade–friction compositions.Friction changes on the closing lane across the merge; grade tilts the full road width.Source: §5 · p. 8

Evaluation tiers

  • Tier 1 compares all five regimes on 32 held-out worlds, while Tier 2 evaluates the three learned regimes on 1,024 worlds.Each controller is compared with its own reference rollouts on the same worlds.Source: §5 · p. 8

Outcomes and statistics

  • Valid passage requires crossing the counting line while still valid; collision, off-road departure, and incursion are tracked separately, and the world resets after an overturn or road-mesh exit.Worlds are the statistical unit, paired across conditions and controllers by scenario seed.Source: §5 · p. 8

6 Results

The results show that reduced friction lowers valid passage and raises near-collision exposure, while grade effects are generally smaller and training-seed variation complicates policy comparisons. Severe failures and study scope qualify how these outcomes should be interpreted.

Friction and grade

  • 0.06–0.90: valid passage fell by this amount at µsurface = 0.18 for every learned checkpoint, with every interval excluding zero.The patch covers only the closing lane, yet open-lane passage fell from 0.82–0.88 to 0.28–0.62.Source: §6 · p. 8, Figure 3 · p. 10
  • Most passage loss at µsurface = 0.18 came from censoring after a sliding closing-lane vehicle overturned and reset its world.Open-lane collisions rose only from about 1% to 3–6%.Source: §6 · p. 9
  • Grade changed passage by at most 0.011 for seven of twelve checkpoints; the largest loss was 0.15 at −8% for L1 seed 0, with collisions rising from 4.9% to 22.1%.Grade–friction compositions did not depart resolvably from additivity.Source: §6 · p. 9

Interaction safety

  • At every µsurface ≤0.12, the share of vehicles reaching TTC below 1.5 s rose for all twelve checkpoints, despite shorter valid-driving windows.These are associations, not causal effects.Source: §6 · p. 9

Training exposure and preview

  • Across seeds, the L1 versus L0-T passage contrast was unresolved, and the preview produced no anticipation of low friction.L1 entered a −8% slope 1.0–2.4 m/s slower; Frozen L0 versus L0-T also includes additional training.Source: §6 · p. 9

Classical and learned regimes

  • PDM-Closed-17 lost 0.19 passage at µsurface = 0.18 and still passed 0.52–0.54 at µsurface ≤0.08, at least 0.13 more than any other regime.Its loop time was 30 s per world on 48 CPU workers, versus 0.42 s for a learned policy.Source: §6 · p. 9

Computational scaling

  • 27.0 to 28.4 s: loop time grew by this amount from 64 to 1,024 twelve-vehicle worlds on one GPU, at efficiency 0.95.The run reached 1.71 million vehicle-steps/s with 8.7 GB peak memory, without saturating.Source: §6 · p. 9

Discussion and limitations

  • The simulator is verified but not calibrated, and its safety measures are not evidence about real-road safety.The study also uses one closure geometry, one policy architecture, noise-free preview, and four seeds that cannot resolve family differences of the observed size.Source: §6 · pp. 9–10

7 Conclusion

SceneFactory-3D enables matched physical counterfactuals and shows that reduced friction worsens task completion and interaction risk across controllers, while tested grades have comparatively limited effects. The conclusion also reports near-linear scaling to 1,024 worlds on one GPU.

7 Conclusion

  • Reduced friction degraded task completion for every controller, increased interaction risk, and affected vehicles outside the altered lane; tested grades had comparatively limited effects.Throughput scaled nearly linearly to 1,024 worlds on one GPU.Source: §7 · p. 10

AI Use Statement

  • Generative AI tools assisted with implementation, experiments, analysis, figures, references, and manuscript drafting, while the authors directed the research and take responsibility for its content.Reported numbers in Section 6 and the appendices were generated from recorded runs by analysis scripts.Source: AI Use Statement · p. 12

Reproducibility Statement

  • A frozen manifest records benchmark, checkpoint, terrain, scenario-seed, software, and GPU details, and every evaluated cell was checked against it.The code, manifest, checkpoints, per-world outcome tables, and analysis scripts are slated for release with supplementary material.Source: Reproducibility Statement · p. 12

A Abstraction and Intended Use

  • The capability comparison treats simulator design choices as context-dependent rather than a measured ranking, with adequacy depending on the outcome of interest.Table 2 scores systems on six ordinal axes, including force-based execution, spatial road physics, and matched interventions.Source: A Abstraction and Intended Use · p. 16, Table 2 · p. 16

B Vehicle Model and Runtime

SceneFactory-3D computes vehicle forces at wheel contacts, integrates chassis motion and rigid contacts at 240 Hz, and updates actions at 60 Hz. Its benchmark uses a shared vehicle model, while the model omits several tire and material effects and terrain is fixed during an environment’s lifetime.

B Vehicle Model and Runtime

  • At each 240 Hz step, the simulation integrates chassis rigid-body motion and resolves chassis-terrain and chassis-chassis contacts; a custom kernel computes other vehicle forces, while 60 Hz actions are held for four physics steps.All benchmark, scaling, and batch-sensitivity runs used one RTX PRO 6000 Blackwell Max-Q Workstation GPU; benchmark cells ran sequentially.Source: B Vehicle Model and Runtime · pp. 17–18
  • Wheel-level suspension rays query each world’s road mesh, and compression produces a normal load subject to contact and slope limits.Contact is absent when a ray misses or the surface normal is steeper than cos^-1 0.15.Source: B Vehicle Model and Runtime · p. 17
  • When combined tire force exceeds µFz, longitudinal and lateral forces are scaled to that friction-circle bound, with wheel speed updated using the resulting reaction torque.The tire laws use linear slip and an arctangent slip angle; the model lacks a post-peak sliding decline, relaxation length, and load sensitivity beyond the linear bound.Source: B Vehicle Model and Runtime · p. 17
  • Aerodynamic drag acts on the chassis, while requested acceleration and steering pass through command mappings and software limits before vehicle dynamics.The acceleration mapping is open loop, so realized and requested acceleration can differ; tire-force saturation is a separate physical limit.Source: B Vehicle Model and Runtime · pp. 17–18
  • The authored surface coefficient multiplies vehicle grip rather than representing an independently measured tire-road coefficient; tire temperature, anti-lock braking, detailed material response, and deformable terrain are not modeled.The GPU kernel is regression-tested against the platform’s single-vehicle reference implementation.Source: B Vehicle Model and Runtime · p. 18
  • Terrain is fixed for an environment’s lifetime, and partial resets restore affected dynamic state but not every hidden state needed to branch a warmed-up traffic experiment.Worlds share a world-indexed model, with collisions resolved within—but never between—worlds.Source: B Vehicle Model and Runtime · p. 18
  • The frozen benchmark specifies exact framework and software versions because defaults and API changes can affect physical behavior.The tested setup includes Newton 1.5.0, Warp 1.17.0, PyTorch 2.10 with CUDA 12.8, and Python 3.12.Source: B Vehicle Model and Runtime · p. 18

C Physical Verification

Physical verification checks authored conditions, loaded geometry and friction, and controller-independent vehicle probes. The probes confirm that friction and grade interventions reach the dynamics, while their results are internal-consistency checks rather than real-vehicle calibration.

C Physical Verification

  • Verification checks the authored condition, loaded geometry and friction tables, then uses vehicle probes that bypass adaptive controller behavior.Learned-observation checks also compare preview values against physical tables and terrain geometry.Source: C Physical Verification · p. 19
  • A −4.5% descent lengthens the µsurface = 0.18 stop by about a third, whereas a +3.0% climb offsets part of the friction loss.Across conditions, measured slope matched authored grade within 0.05 percentage points, and achieved braking deceleration corresponded to 1.02–1.15 times authored friction values.Source: C Physical Verification · pp. 19–20
  • The probes are internal-consistency checks, not real-vehicle calibration, and their limits apply to the tested entry states rather than bounding task performance.For example, at 16 m/s a full-brake probe stops within the 140 m patch at µsurface = 0.12, but exits it at 7.1 and 11.0 m/s at 0.08 and 0.05.Source: Table 5 · p. 19, C Physical Verification · p. 20
  • Preview friction matched exactly at 10,902 sampled points, and maximum grade discrepancy was 2.0 × 10^-5 under the stated finite-difference convention.Patch-edge discrepancies mean point-query agreement does not guarantee every wheel simultaneously experiences the previewed surface.Source: C Physical Verification · p. 20
  • Terrain-collider checks covered 1, 2, 4, 16, and 64 simultaneous worlds, including up to 64 unique 43,200-triangle terrains.Checks included collider and wheel-query geometry, assigned terrain variants, chassis contact, and partial resets.Source: C Physical Verification · p. 20

D Controller Interfaces and Training

The section specifies how controllers receive observations and issue actions, how their training and benchmark conditions are structured, and how outcomes and uncertainty are measured. It also describes the reset and censoring rules that constrain interpretation of results.

D Controller Interfaces and Training

  • IDM-17 uses route-relative car following and Stanley lateral tracking, while PDM-Closed-17 uses proposal scoring, a kinematic forecast, and an LQR tracker.Their lateral tracking and decision rates differ, so their comparison is not an ablation of proposal search.Source: D Controller Interfaces and Training · p. 20
  • Learned policies share 112 observation channels, while L1 receives 12 physical preview channels; policies act at 10 Hz and each action is held for six 60 Hz steps.Evaluation uses the actor’s distribution mean; the critic receives no privileged input.Source: D Controller Interfaces and Training · p. 21
  • L0-T and L1 share continuation stages, scenes, terrain sampling, seeds, and reward, while L1’s added input weights are zero-initialized to match its parent’s initial function.Continuation uses up to 4.6 × 10^7 vehicle decisions per seed; terrain is sampled independently of the scene generator, with 60% of patches near the maneuver point.Source: D Controller Interfaces and Training · p. 22

E Conditions, Outcomes and Statistics

  • The benchmark uses 21 friction and grade conditions, including combined grade-friction cases, with conditions applied around a lane closure and a 66-second horizon.The benchmark also combines unseen road geometry and task, a larger population, and friction below 0.2 outside the training range; preview is privileged and exact.Source: D Controller Interfaces and Training · p. 22, E Conditions, Outcomes and Statistics · p. 23
  • A world reset truncates scoring after its first fault, so other vehicles’ outcomes may be truncated and censoring may vary by controller and condition.Matched sensitivity estimates use per-world differences, while bootstrap intervals resample worlds; seed-level summaries describe training variability.Source: E Conditions, Outcomes and Statistics · p. 24

F Studies and Preliminary Evidence

The development study evaluates fixed checkpoints on held-out scenes, then probes how preview inputs relate to policy behavior. It reports preliminary controller outcomes and preview analyses, distinct from the final two-tier benchmark.

F Studies and Preliminary Evidence

  • Table 10 presents a preliminary controller-development evaluation, reporting mean ± standard deviation over four training seeds; it is explicitly not a final two-tier benchmark result.Source: Table 10 · p. 25
  • Terrain-trained arms showed less degradation than the nominal policy when rolling ground was combined with reduced friction, while grade alone changed completion little.Some terrain-trained seeds nevertheless had higher collision rates than their parents on the reference road.Source: F Studies and Preliminary Evidence · p. 26

G Preview Analyses

  • Nominalizing L1’s preview lowered success by 0.016–0.038 on selected friction bands and whole-road conditions, while changing nothing resolvable on the lane patch.A late-shifted preview reproduced the nominalized outcome; early-shifted and constant false previews performed about as well as true preview on these patch-bearing scenes.Source: G Preview Analyses · p. 26
  • At the µsurface = 0.25 band, L1 accelerated more than its nominalized version 30 m before the patch, then less at entry, and entered at a speed similar to L0-T.The probe found no anticipatory braking when friction was reduced only ahead of the vehicle.Source: G Preview Analyses · p. 26
  • L1 and L0-T were trained identically except for preview; their development-study success and failure did not differ resolvably on any friction condition.The only resolved safety difference was 0.005 fewer at-fault collision onsets per vehicle for L1 at whole-road µsurface = 0.4; L1 was 0.13–0.33 m/s slower in most conditions.Source: G Preview Analyses · p. 26

H Computational Scaling

The section measures how simulation throughput and memory scale with world count, and reports controller runtimes under production workloads. Throughput rises almost proportionally from 64 to 1,024 worlds, while planner cost can dominate the loop.

H Computational Scaling

  • The scaling workloads use 12 live vehicles per world, production controller L1 seed 0, and three repeats per point on one GPU.The timed loop includes physics, observation, policy inference, actuation, and configured logging or safety diagnostics; it excludes construction, warm-up, resets, evaluation, and rendering.Source: H Computational Scaling · pp. 26–27
  • Throughput grows almost in proportion to world count over 64–1,024 worlds, while loop time remains nearly constant.The measurements show headroom to 1,024 worlds on this GPU but do not locate saturation or vary vehicles per world.Source: H Computational Scaling · p. 27, Table 11 · p. 27, Figure 4 · p. 27
  • PDM-Closed-17 took 30 s per world (1,565 vehicle-steps/s), about 961 times the production learned cost, so it appears only in Tier 1.This is the cost of the published planner in the authors’ adaptation, not the simulator.Source: H Computational Scaling · p. 27
  • IDM-17 took 1.2 s per world in 32-world batches, while a learned policy took 0.032 s per 66-s world at 1,024 worlds.At 64 worlds, the learned policy took 0.42 s per world; PDM-Closed-17 took 30 s per world with 48 CPU planning workers.Source: H Computational Scaling · p. 27

I Production Audit and Supplementary Results

The supplementary section audits the frozen production run and examines how first-event scoring, measurement windows, and batch size affect reported outcomes. It also presents supplementary grade, hazard, and matched-pair analyses.

I Production Audit and Supplementary Results

  • All 252 Tier-2 cells, 252 rich-subset records, and 42 Tier-1 cells were complete, with frozen manifests matching and no cells failed, discarded, or restarted.In every cell, valid passage matched passages before the first reset.Source: I Production Audit and Supplementary… · p. 28
  • For Frozen L0 seed 0 at µsurface = 0.25, aggregate collision rate was 0.33, versus 0.08 for first-episode collision rate.Valid passage is unaffected by resets; the reported outcome classes are recomputed from first-event steps before the first reset.Source: I Production Audit and Supplementary… · p. 28
  • At µsurface = 0.08, time below TTC 1.5 s per driving second was 1.7–2.3 times larger over the whole episode than in the valid-driving window.Post-failure steps inflate whole-episode interaction measures.Source: I Production Audit and Supplementary… · p. 28
  • The batch-size sensitivity criterion was not met; the reference-to-µsurface = 0.18 change ranged from −0.09 to −0.13 across batch sizes.The criterion required cross-batch differences to stay within the same-batch rerun envelope.Source: I Production Audit and Supplementary… · p. 28, Table 12 · p. 28
  • The hazard analysis uses exposure from 1–2 s before failure; a simpler ever-exceeded-capacity comparison has the opposite sign for several slower checkpoints.For L0-T seed 2 at µsurface = 0.08, the simpler comparison gives 0.19 versus 0.58; residual capacity includes µsurface by construction.Source: I Production Audit and Supplementary… · p. 28, Table 17 · p. 31
  • Tables 13–16 and 18 cover selected-condition passage, vehicle outcomes, family and grade–friction contrasts, and Tier-1 classical and learned passage.Table 13 reports Tier-2 passage by checkpoint; Table 18 reports Tier-1 passage on 32 worlds.Source: Table 13 · p. 29, Table 14 · p. 30, Table 15 · p. 30
  • Figure 5 tracks changes in valid passage, TTC-below-1.5-s share, and closing-lane speed; L1 slows before downhill grades.The figure’s panels show changes in these measures over the grade sweep.Source: Figure 5 · p. 29
  • Figure 6’s selected Frozen L0 pair compares the reference road with µsurface = 0.25; it illustrates one mechanism and is not representative.Its panels show the focal vehicle’s speed and lateral position until failure.Source: I Production Audit and Supplementary… · p. 29, Figure 6 · p. 31
Loading 2610.02874v1…