Classical vs. AI-Based Methods in Autonomous Vehicles: Reproducible Experiments from Perception to Field Safety Data

I. Introduction

Road crashes killed 39,254 people in the United States in 2024, a rate of 1.19 fatalities per 100 million vehicle miles travelled (VMT) [1]. Automated driving systems are expected to reduce this toll, and after decades of research that began with neural-network steering [2] and the DARPA challenges [3,4], Level 4 services [5] now operate without a human driver in several cities [6]. These systems depend on AI throughout their software stack, from learned perception [7,8] and probabilistic estimation [9] to prediction, planning and control [10,11], and increasingly on end-to-end networks trained for the driving objective [12,13].

Surveys of the field [7,8,13,14] catalogue methods well, but they cannot answer a practitioner’s quantitative questions: how large is the gain from a given technique, under what conditions does it appear, and is it statistically reliable? Primary studies answer such questions for one component at a time, but they use different datasets, tuning budgets and metrics, and they seldom report uncertainty. Two consequences are well documented. Results can depend more on evaluation protocol than on method, as shown for open-loop planning metrics [15]. And field safety claims can be statistically meaningless without enough exposure [16].

This paper addresses that gap with a controlled, reproducible evaluation of representative classical and AI-based methods at every layer of the stack, conducted under one statistical protocol and released as open code. The study is organised around four research questions (RQs):

• RQ1 (perception): How much do learned spatial context and appearance features improve drivable-area and lane perception over hand-crafted pipelines on real driving images, and how sensitive are classical detectors to image conditions?
• RQ2 (estimation, prediction, planning): How do sensor fusion, multimodal prediction, heuristic search and reinforcement learning (RL) trade accuracy against robustness and computation?
• RQ3 (control): How do geometric, LQR and MPC controllers compare under actuator limits when each is tuned by the same validation procedure?
• RQ4 (fleet learning and field evidence): How much of the benefit of pooled fleet data can federated learning recover under non-IID data, and what do public field data show about AV safety, including the exposure needed to support a claim?

The contributions of this work are the following.

• A common experimental protocol for AV components, with held-out evaluation, validation-based tuning, CIs, and paired Wilcoxon tests with Holm correction (Section III).
• Twelve experiments (E1–E12) with procedures, parameter tables and scripts (Section IV). They include several findings that contradict simple narratives: the gain of multimodal predictors comes from hypothesis count rather than model quality, controller rankings reverse after principled tuning, and a classical lane detector is insensitive to global brightness.
• An independent statistical re-analysis of the Waymo Safety Impact Data Hub (220.6 million rider-only miles) [6,17] with exact CIs for nine crash types and an updated exposure analysis [16].
• An open release of code, transcribed public data and results that reproduces every table and figure.

Section II reviews related work, Section III describes the methodology, Section IV reports the experiments, Section V discusses the answers to the RQs, Section VI analyses threats to validity, and Section VII concludes.

II. Related Work

A. System Architectures and Surveys

SAE J3016 [5] defines the levels of driving automation used throughout this paper. Modular AV stacks decompose driving into perception, localisation, prediction, planning and control, a design inherited from the DARPA Urban Challenge [4] and adopted by open-source platforms such as Autoware [18]. Broad surveys cover the modular stack [8,14], deep learning for driving [7], planning and decision-making [11], and motion planning and control [10,19]. End-to-end driving has been surveyed by Chen et al. [13]. These works synthesise methods but do not provide controlled cross-component experiments, which motivates the present study.

B. Perception

Deep convolutional networks [20,21] displaced hand-engineered features in perception. Detection progressed from two-stage detectors [22] to single-stage [23] and Transformer-based set prediction [24]. Lidar detection became efficient with pillar encodings [25], and multi-camera perception moved into bird’s-eye view (BEV) through depth-based lifting [26], spatiotemporal attention [27], and camera–lidar fusion [28]. Dense segmentation benchmarks such as Cityscapes [29] and BDD100K [30] drove progress in drivable-area estimation. Classical lane detection, which remains common in low-cost driver assistance, combines edge detection [31] with the Hough transform [32], typically implemented in OpenCV [33]. Benchmark datasets for these tasks include KITTI [34], nuScenes [35] and the Waymo Open Dataset [36]. We use comma10k [37] because it is permissively licensed and can be downloaded without registration, which makes exact reproduction possible.

C. State Estimation and Mapping

The Kalman filter [38] and its extended and unscented variants [39] remain the workhorses of multi-sensor fusion, and particle filters handle multimodal beliefs [9]. Simultaneous localisation and mapping (SLAM) [40] builds maps while estimating pose, using feature-based visual pipelines such as ORB-SLAM2 [41], place recognition for loop closure [42], and occupancy grids for planning [43].

D. Motion Prediction

Physics-based predictors are strong short-horizon baselines. Learned predictors encode agents and maps as vectors [44] or use Transformer intention queries [45], and are evaluated with multimodal displacement metrics on large datasets such as the Waymo Open Motion Dataset [46] and Argoverse 2 [47]. Recurrent [48] and attention-based [49] sequence models underpin most of these methods.

E. Planning and Decision-Making

Graph search with Dijkstra’s algorithm [50] and A* [51] underlies route and grid planning, and Hybrid A* extends it to vehicle kinematics [52]. Tactical decisions are often modelled as Markov decision processes and solved with Q-learning [53] or deep RL [54,55]. A survey of deep RL for driving [56] highlights reward design and sim-to-real transfer as open problems. Imitation learning suffers from covariate shift [57], which motivates data augmentation in systems such as ChauffeurNet [58].

F. Motion Control

Geometric path trackers include Pure Pursuit [59] and the Stanley controller [60], both based on the kinematic bicycle model [61]. Optimal control offers the linear quadratic regulator (LQR) [62] and model predictive control (MPC) [63], which is widely used in industry [64] and practical with convex modelling tools [65] and fast QP solvers [66].

G. End-to-End and Foundation Models

End-to-end networks map sensors to actions [67] and have evolved into sensor-fusion Transformers [68], planning-oriented multi-task models [12] and vectorised planners [69]. Generative world models [70], neural scene reconstruction [71] and vision-language driving models [72,73] are emerging. Li et al. [15] showed that open-loop metrics can reward trivial ego-status baselines, underscoring the need for careful protocols. We do not evaluate these large models; our experiments isolate mechanisms, such as context, multimodality and anticipation, that such models exploit.

H. Safety Evidence, Standards and Fleet Learning

Kalra and Paddock [16] quantified the mileage needed to demonstrate AV safety statistically. Waymo has published retrospective crash-rate comparisons based on NHTSA Standing General Order reports [74], human benchmarks [75] and successive analyses [17,76], summarised on a public data hub [6]. California publishes annual testing reports [77]. Exact Poisson intervals [78] are appropriate for such rare-event counts. Safety assurance of AI-based functions draws on ISO 26262 [79], ISO 21448 [80], ISO/PAS 8800 [81] and UL 4600 [82]. Federated learning [83] allows fleets to train shared models without centralising raw data. Uncertainty estimation [84], saliency-based explanation [85] and adversarial robustness [86] are complementary concerns.

III. Methodology

A. Experimental Design

Figure 1 shows the AV software architecture studied: the sensor data, the signals exchanged between modules, and the experiment and research question that address each module. Each experiment compares a small set of methods (the independent variable) on a shared task and data split, and reports task metrics (dependent variables) with uncertainty. Table I summarises the design. Experiment E4, a numerical verification of textbook equations, is reported in Appendix A.

Table I. Experimental Design Matrix

Exp.	RQ	Data	Independent variable	Primary metrics	Replication	Test
E1	RQ1	comma10k, 1,000 masks	Semantic class	Pixel share	1,000 images	Bootstrap CI
E2	RQ1	comma10k, 378 frames	Canny thresholds; brightness	P, R, F1	100 dev / 278 test	Bootstrap CI; Kruskal–Wallis
E3	RQ1	comma10k, 400 frames	Feature set	IoU	150 test × 3 seeds	Wilcoxon + Holm
E5	RQ2	Simulation	Estimator	Position RMSE	50 runs	Wilcoxon + Holm
E6	RQ2	Simulation	Pose source	Map IoU, pose RMSE	10 seeds	Wilcoxon
E7	RQ2	Simulation	Predictor	ADE, FDE, miss rate	5 seeds × 3,000	t CI; Wilcoxon + Holm
E8	RQ2	Simulation	Heuristic, weight	Nodes, time, excess cost	49 maps	Wilcoxon + Holm
E9	RQ2	Simulation	Policy	Crash rate	10 seeds × 500 episodes	Wilcoxon
E10	RQ3	Simulation	Controller; speed	RMS error, steering rate	10 seeds × 2 speeds	Wilcoxon + Holm
E11	RQ4	Waymo hub; CA DMV; NHTSA	Crash type; severity	Rate ratio; miles needed	Published counts	Exact Poisson
E12	RQ4	comma10k, 400 frames	Training regime	IoU; gap recovered	5 seeds	t CI; Wilcoxon

Table I. Experimental Design Matrix

Exp.	RQ	Data	Independent variable	Primary metrics	Replication	Test
E1	RQ1	comma10k, 1,000 masks	Semantic class	Pixel share	1,000 images	Bootstrap CI
E2	RQ1	comma10k, 378 frames	Canny thresholds; brightness	P, R, F1	100 dev / 278 test	Bootstrap CI; Kruskal–Wallis
E3	RQ1	comma10k, 400 frames	Feature set	IoU	150 test × 3 seeds	Wilcoxon + Holm
E5	RQ2	Simulation	Estimator	Position RMSE	50 runs	Wilcoxon + Holm
E6	RQ2	Simulation	Pose source	Map IoU, pose RMSE	10 seeds	Wilcoxon
E7	RQ2	Simulation	Predictor	ADE, FDE, miss rate	5 seeds × 3,000	t CI; Wilcoxon + Holm
E8	RQ2	Simulation	Heuristic, weight	Nodes, time, excess cost	49 maps	Wilcoxon + Holm
E9	RQ2	Simulation	Policy	Crash rate	10 seeds × 500 episodes	Wilcoxon
E10	RQ3	Simulation	Controller; speed	RMS error, steering rate	10 seeds × 2 speeds	Wilcoxon + Holm
E11	RQ4	Waymo hub; CA DMV; NHTSA	Crash type; severity	Rate ratio; miles needed	Published counts	Exact Poisson
E12	RQ4	comma10k, 400 frames	Training regime	IoU; gap recovered	5 seeds	t CI; Wilcoxon

B. Computing Environment

All experiments were run on a single virtual CPU core (Intel Xeon, 2.1 GHz) with 3 GB of RAM and no GPU, under Linux x86-64. Table II lists the software versions. Wall-clock times were measured with Python’s high-resolution performance counter and are reported only for relative comparison, because pure-Python implementations are far slower than production code.

Table II. Software Environment

Component	Version	Use
Python	3.12.3	All experiments
NumPy / SciPy	2.4.4 / 1.17.1	Numerics, statistics, Riccati solver
scikit-learn [87]	1.8.0	Gradient boosting, MLP, k-NN
OpenCV [33]	4.13.0	Image processing, Canny, Hough
CVXPY [65] with OSQP [66]	1.9.2	Model predictive control
pandas / Matplotlib	3.0.2 / 3.10.8	Data handling, figures

Table II. Software Environment

Component	Version	Use
Python	3.12.3	All experiments
NumPy / SciPy	2.4.4 / 1.17.1	Numerics, statistics, Riccati solver
scikit-learn [87]	1.8.0	Gradient boosting, MLP, k-NN
OpenCV [33]	4.13.0	Image processing, Canny, Hough
CVXPY [65] with OSQP [66]	1.9.2	Model predictive control
pandas / Matplotlib	3.0.2 / 3.10.8	Data handling, figures

C. Data Sources and Splits

Real driving images. comma10k [37] contains 9,880 trainable dashcam frames (1164 × 874 pixels) with crowd-sourced masks in five classes: road, lane markings, undrivable, movable objects and ego vehicle. A reproducible sample was drawn with Python’s random.sample (seed 42): 1,000 masks for E1, and the first 400 of these (in sorted, hash-random order) with their images for E2, E3 and E12. For E2, frames whose lane labels intersect the detector’s region of interest form a 100-frame development set and a 278-frame test set. For E3 and E12, the first 250 frames are used for training and the last 150 for testing. No test frame is used for training or tuning.

Public field data. The Waymo Safety Impact Data Hub (data through March 2026) [6] provides rider-only (RO) miles by city, crash rates per million miles (IPMM) with human benchmarks, and crash counts by type. The benchmarks follow [75], and crash data come from NHTSA Standing General Order reports [74]. We transcribed these tables together with California DMV testing-mileage totals [77] and the 2024 U.S. fatality rate [1] on 16 September 2026.

Simulations. E5–E10 use seeded simulations whose parameters are given with each experiment. A new random seed changes every stochastic element (noise, traffic, data generation, model initialisation), and test seeds are never used for tuning.

D. Evaluation Metrics

For binary segmentation with predicted mask P and ground truth G, the intersection over union is

$$\mathrm{IoU}={\displaystyle \frac{|P\cap G|}{|P\cup G|}}$$

(1)

and it is computed per image. For lane detection, a predicted pixel is a true positive if a labelled pixel lies within $\tau$ = 6 pixels, and a labelled pixel is recalled if a prediction lies within $\tau$. Precision (P), recall (R) and the per-image F1 score are

$$P={\displaystyle \frac{T{P}_{\mathrm{pred}}}{N_{\mathrm{pred}}}},R={\displaystyle \frac{T{P}_{\mathrm{label}}}{N_{\mathrm{label}}}},F_1={\displaystyle \frac{2PR}{P+R}}$$

(2)

with F1 set to 0 when there are no detections. Localisation accuracy is the root-mean-square error of the estimated position ${\widehat{p}}_k$ over K time steps:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{K}}\sum _k{∥{\widehat{p}}_k-p_k∥}^2}$$

(3)

For trajectory prediction over T future steps, with K hypotheses ${\widehat{y}}^{\left(k\right)}$ where applicable,

$$\begin{array}{cc}\hfill \mathrm{ADE}& ={\displaystyle \frac{1}{T}}\sum _t\parallel {\widehat{y}}_t-y_t\parallel ,\mathrm{FDE}=\parallel {\widehat{y}}_T-y_T\parallel ,\hfill \\ \hfill {\mathrm{minADE}}_K& =\underset{k}{min}\mathrm{ADE}\left({\widehat{y}}^{\left(k\right)}\right)\hfill \end{array}$$

(4)

and the miss rate is the fraction of cases whose best final error exceeds 2 m. For planning, the excess cost of a path relative to the optimal cost c* is $100·(c$− c*)/c* percent. Decision policies are scored by the crash rate per 200-step episode. Controllers are scored by the RMS lateral error $e_y$ and by the RMS steering rate, a proxy for comfort and actuator wear:

$$\begin{array}{cc}\hfill {\mathrm{RMS}}_e& =\sqrt{\frac{1}{K}{\sum }_k{e}_{y,k}^2},\hfill \\ \hfill {\mathrm{RMS}}_{\dot{\delta }}& =\sqrt{\frac{1}{K}{\sum }_k{\left(({\delta }_k-{\delta }_{k-1})/\Delta t\right)}^2}\hfill \end{array}$$

(5)

Controller parameters are tuned by minimising the scalar objective

$$J={\mathrm{RMS}}_e+\lambda {\mathrm{RMS}}_{\dot{\delta }},\lambda =0.01\mathrm{m}/{(}^{\circ }/\mathrm{s})$$

(6)

on a validation path (Section IV-D).

E. Statistical Analysis

The unit of analysis is the independent replicate: an image, a map, a Monte Carlo run or a training seed. Results are reported as mean [95% CI]. When replicates are seeds or runs, CIs use the t distribution:

$$\overline{x}\pm t_{0.975,n-1}{\displaystyle \frac{s}{\sqrt{n}}}$$

(7)

When replicates are images, CIs are percentile bootstrap intervals with 5,000 resamples. Paired comparisons use the two-sided Wilcoxon signed-rank test, whose effect size is the matched-pairs rank-biserial correlation

$$r_{\mathrm{rb}}={\displaystyle \frac{{\sum }_{d>0}{R}_d-{\sum }_{d<0}{R}_d}{\sum R_d}}$$

(8)

where $R_d$ are the ranks of the absolute paired differences; $r_{\mathrm{rb}}$ ranges from $-1$ to 1. When several comparisons are made within an experiment, p-values are adjusted with the Holm–Bonferroni procedure. Differences across three or more independent groups use the Kruskal–Wallis test. Significance is assessed at $\alpha$ = 0.05, but we emphasise effect sizes and CIs, because very small differences can be significant with many replicates (Section IV-B). For event counts k with benchmark-expected count E (E11), exact Garwood intervals [78] give

$$\begin{array}{cc}\hfill \mathrm{CI}\left(\lambda \right)& =\left[\frac{1}{2}{\chi }_{\alpha /2}^2\left(2k\right),\frac{1}{2}{\chi }_{1-\alpha /2}^2(2k+2)\right],\hfill \\ \hfill \mathrm{CI}\left(\mathrm{ratio}\right)& =\mathrm{CI}\left(\lambda \right)/E\hfill \end{array}$$

(9)

F. Reproducibility Procedure

Figure 2 shows the evaluation pipeline, and Algorithm 1 gives the protocol that every experiment script implements. Seeds, splits and parameter grids are hard-coded in the scripts. Long experiments cache per-seed results under results/raw, so an interrupted run resumes without recomputation. The entire study is reproduced with the commands below; the first command downloads the comma10k sample (about 100 MB).

Algorithm 1 Common experimental protocol

1: Fix the data split and all random seeds; the test split is never inspected during development. 2: For each method with free parameters, select parameters on a development or validation set only (E2, E10). 3: For each replicate (image, map, run or seed), run every method on identical inputs and record all metrics. 4: Store raw per-replicate results (results/raw) and aggregate them into mean [95% CI] (Eq. 7 or bootstrap). 5: Test the pre-specified paired comparisons with the Wilcoxon signed-rank test and report $r_{\mathrm{rb}}$ (Eq. 8); apply Holm correction within the experiment. 6: Export the summary JSON (results/) and figures (figures/) used verbatim in this paper.

IV. Experiments and Results

Each experiment below states its objective and hypothesis, setup, procedure, reproduction script and results. Hypotheses are labelled H1–H12 by experiment number, and their outcomes are summarised in Table XVII (Section V).

A. RQ1: Perception on Real Driving Images

1). E1: Class Composition of Driving Scenes

Objective and hypothesis. To quantify the class imbalance that perception models face. H1: lane markings occupy under 1% of image pixels.

Setup and procedure. Each of the 1,000 comma10k masks is downsampled by a factor of two, and every pixel is assigned the nearest of the five class colours. We compute per-image class shares, their mean with 95% bootstrap CIs over images, the share of images containing each class (more than 0.1% of pixels), and the empirical spatial prior P(drivable | pixel). Script: exp01_comma10k_statistics.py.

Results. Table III and Figure 3 show the results. Lane markings occupy 0.588% [0.561, 0.613] of pixels, so H1 is supported, even though they appear in 83.3% of frames. The largest-to-smallest class ratio is 88.7:1, and lane markings cover only 3.1% of the road surface. The drivable prior is concentrated in a horizontal band above the hood (Figure 3c), a regularity that E3 tests explicitly.

Table III. Pixel Class Composition of 1,000 comma10k Frames (E1)

Class	Mean pixel share (%) [95% CI]	Frames containing class (%)
Road	18.91 [18.61, 19.20]	98.9
Lane markings	0.588 [0.561, 0.613]	83.3
Undrivable	52.12 [51.72, 52.52]	100.0
Movable objects	2.89 [2.61, 3.18]	82.1
Ego vehicle	25.50 [25.20, 25.81]	100.0

Table III. Pixel Class Composition of 1,000 comma10k Frames (E1)

Class	Mean pixel share (%) [95% CI]	Frames containing class (%)
Road	18.91 [18.61, 19.20]	98.9
Lane markings	0.588 [0.561, 0.613]	83.3
Undrivable	52.12 [51.72, 52.52]	100.0
Movable objects	2.89 [2.61, 3.18]	82.1
Ego vehicle	25.50 [25.20, 25.81]	100.0

2). E2: Classical Lane Detection

Objective and hypothesis. To measure how well a classical edge-and-line pipeline [31,32] recovers human-labelled lane markings, and whether its performance depends on image brightness. H2a: held-out F1 exceeds 0.6. H2b: F1 decreases in darker frames.

Setup. The detector applies contrast-limited histogram equalisation (clip limit 2.0, 8 × 8 tiles), a 5 × 5 Gaussian blur, Canny edge detection, a trapezoidal region of interest (ROI) covering 50–78% of the image height, and a probabilistic Hough transform (threshold 40, minimum length 40 px, maximum gap 25 px). Segments with slope magnitude outside [0.3, 3.0] are discarded. The matching tolerance is $\tau$ = 6 px (Eq. 2). Figure 4a shows the complete pipeline.

Algorithm 2 Lane detector threshold selection and testing (E2)

1: Keep frames whose lane labels intersect the ROI (378 frames); split them into development (100) and test (278) sets. 2: For each Canny threshold pair (low/high) ∈ {20/60, 40/120, 60/150, 80/200, 150/300}: detect lanes on each development frame and compute per-image P, R and F1 (Eq. 2). 3: Select the pair with the highest mean development F1. 4: Run the selected detector once on the test set; report mean P, R and F1 with 95% bootstrap CIs. 5: Split test frames into terciles of mean grey level; compare F1 across terciles with the Kruskal–Wallis test.

Results. On the development set, F1 peaks at thresholds 40/120 (0.695); the looser 20/60 (0.675) and tighter 60/150 (0.670) are close, but 150/300 collapses to 0.185 as recall falls to 0.15 (Figure 6b). On the test set (Table IV), precision is 0.734 [0.695, 0.770], recall 0.700 [0.661, 0.736] and F1 0.658 [0.620, 0.694], so H2a is supported. The per-image F1 distribution is bimodal (Figure 6a): the median frame is detected well (median precision 0.862, median recall 0.838), but 20 frames (7.2%) yield no detection and 14.0% have recall below 0.2. H2b is not supported: F1 is statistically indistinguishable across brightness terciles (Kruskal–Wallis p = 0.93). Visual inspection (Figure 5) attributes failures instead to worn or occluded markings, strong shadows, glare, and dense traffic that masks the markings. Global brightness is therefore a poor proxy for difficulty.

Table IV. Classical Lane Detection on 278 Held-Out Frames (E2)

Subset	n	Precision [95% CI]	Recall [95% CI]	F1 [95% CI]
All test frames	278	0.734 [0.695, 0.770]	0.700 [0.661, 0.736]	0.658 [0.620, 0.694]
Darkest third (grey level 7–66)	93	—	—	0.667 [0.601, 0.733]
Middle third (67–86)	92	—	—	0.663 [0.597, 0.725]
Brightest third (87–151)	93	—	—	0.645 [0.577, 0.708]

Table IV. Classical Lane Detection on 278 Held-Out Frames (E2)

Subset	n	Precision [95% CI]	Recall [95% CI]	F1 [95% CI]
All test frames	278	0.734 [0.695, 0.770]	0.700 [0.661, 0.736]	0.658 [0.620, 0.694]
Darkest third (grey level 7–66)	93	—	—	0.667 [0.601, 0.733]
Middle third (67–86)	92	—	—	0.663 [0.597, 0.725]
Brightest third (87–151)	93	—	—	0.645 [0.577, 0.708]

Figure 5. Test-set lane detections at high, median and low recall (E2). Top: detected segments (cyan), human labels (red) and ROI (yellow). Bottom: Canny edges inside the ROI.

Figure 5. Test-set lane detections at high, median and low recall (E2). Top: detected segments (cyan), human labels (red) and ROI (yellow). Bottom: Canny edges inside the ROI.

Figure 6. Lane detection results (E2). (a) Per-image test F1. (b) Precision–recall trade-off of Canny thresholds on the development set; the selected setting is circled. (c) Test F1 by brightness tercile with 95% bootstrap CIs.

Figure 6. Lane detection results (E2). (a) Per-image test F1. (b) Precision–recall trade-off of Canny thresholds on the development set; the selected setting is circled. (c) Test F1 by brightness tercile with 95% bootstrap CIs.

3). E3: Contribution of Context Features to Drivable-Area Segmentation

Objective and hypothesis. To isolate the contribution of spatial position, colour and local context to learned drivable-area segmentation, the mechanisms that deep networks exploit at scale [20,27]. H3: each added feature group significantly increases test IoU.

Setup. Frames are downsampled eightfold to 146 × 109 pixels, and the drivable label is road plus lane markings. Four nested feature sets are compared: position (normalised row, normalised column and horizontal offset from the image centre; 3 features), colour (BGR, HSV and Lab; 9 features), position + colour (12), and all features (26), which add 14 texture and context features: local grey-level mean and standard deviation in 3 × 3, 7 × 7 and 15 × 15 windows, Sobel gradient magnitude and its 9 × 9 average, and colour averaged over 9 × 9 and 25 × 25 windows. The classifier is scikit-learn’s HistGradientBoostingClassifier (200 iterations, learning rate 0.1, 63 leaf nodes) [87], trained on 250,000 pixels sampled from the 250 training frames. As a non-learned reference, the spatial-prior baseline labels a pixel drivable if more than 50% of training masks do. The pipeline is shown in Figure 4b.

Algorithm 3 Feature-set ablation (E3)

1: Extract all feature groups for the 400 frames; fix the 250/150 train/test split. 2: For seed s ∈ {0, 1, 2} and each feature set F: sample 250,000 training pixels with seed s, train the classifier with seed s, and compute per-image IoU (Eq. 1) on all 150 test frames. 3: Average each test frame’s IoU over the three seeds; report the mean with 95% bootstrap CI over frames. 4: Run paired Wilcoxon tests over the 150 test frames for the pre-specified comparisons; apply Holm correction.

Results. Table V and Figure 7Figure 8 show the results. Seed variation is negligible (standard deviation of mean IoU ≤ 0.002 across seeds), so the CIs reflect frame-to-frame variability. Position alone reaches IoU 0.701, which equals the non-learned spatial prior (median paired difference $-0.0001$). Colour alone is much worse (0.521) and the most variable across frames (SD 0.227), because road appearance varies with lighting and surface. Combining the two adds 0.071 over position ($r_{\mathrm{rb}}$ = 0.70), and texture and context add a further 0.022 ($r_{\mathrm{rb}}$ = 0.68); both steps are significant after Holm correction (p < ${10}^{-11}$), supporting H3. Remaining errors concentrate at road boundaries, at low light and where pavement resembles the road (Figure 8). The steadily diminishing but significant gains from wider context are consistent with the success of large-receptive-field CNN and Transformer models, which this small classical model does not attempt to match.

Table V. Drivable-Area Segmentation on 150 Held-Out Frames (E3)

Feature set	Features	Road IoU [95% CI]	Paired comparison	Median $\mathit{\Delta }$	${\mathit{r}}_{\mathbf{rb}}$	p (Holm)
Spatial prior (no learning)	—	0.701 [0.680, 0.721]	—	—	—	—
Position	3	0.701 [0.679, 0.721]	vs spatial prior	$-0.0001$	$-0.22$	0.017
Colour	9	0.521 [0.484, 0.556]	—	—	—	—
Position + colour	12	0.773 [0.753, 0.793]	vs position; vs colour	+0.071; +0.209	0.70; 0.98	$3.1\times {10}^{-13}$; $1.4\times {10}^{-24}$
All (+ texture, context)	26	0.803 [0.785, 0.821]	vs position + colour	+0.022	0.68	$1.1\times {10}^{-12}$

Table V. Drivable-Area Segmentation on 150 Held-Out Frames (E3)

Feature set	Features	Road IoU [95% CI]	Paired comparison	Median $\mathit{\Delta }$	${\mathit{r}}_{\mathbf{rb}}$	p (Holm)
Spatial prior (no learning)	—	0.701 [0.680, 0.721]	—	—	—	—
Position	3	0.701 [0.679, 0.721]	vs spatial prior	$-0.0001$	$-0.22$	0.017
Colour	9	0.521 [0.484, 0.556]	—	—	—	—
Position + colour	12	0.773 [0.753, 0.793]	vs position; vs colour	+0.071; +0.209	0.70; 0.98	$3.1\times {10}^{-13}$; $1.4\times {10}^{-24}$
All (+ texture, context)	26	0.803 [0.785, 0.821]	vs position + colour	+0.022	0.68	$1.1\times {10}^{-12}$

B. RQ2: State Estimation and Mapping

1). E5: GNSS–Odometry Fusion Under a GNSS Outage

Objective and hypothesis. To quantify the benefit of fusing complementary sensors and to compare linearisation strategies. H5a: fusion reduces RMSE relative to GNSS alone, especially during an outage. H5b: EKF and UKF perform equivalently for this mildly nonlinear model.

Setup. A vehicle follows a 120 s curved route at speeds varying between 12 and 28 m/s. Wheel-speed ($\sigma$ = 0.3 m/s) and yaw-rate ($\sigma$ = 0.02 rad/s, bias 0.01 rad/s) odometry arrive at 10 Hz. GNSS fixes ($\sigma$ = 3 m) arrive at 1 Hz, except during a 20 s outage (t = 50–70 s). Five estimators are compared: dead reckoning; GNSS only (holding the last fix); a linear constant-velocity Kalman filter (KF) using GNSS only; and an extended (EKF) and unscented (UKF) [39] Kalman filter that fuse odometry with GNSS. The fusion filters use the state x = [x, y, $\theta$, ${v]}^{\top }$ and the unicycle motion model

$$\begin{array}{cc}\hfill x_{k+1}& =x_k+\Delta t{v}_{k}cos{\theta }_k,\hfill \\ \hfill y_{k+1}& =y_k+\Delta t{v}_{k}sin{\theta }_k,\hfill \\ \hfill {\theta }_{k+1}& ={\theta }_k+\Delta t{\omega }_k\hfill \end{array}$$

(10)

with the standard KF prediction and update [9,38]:

$$\begin{array}{cc}\hfill {\widehat{x}}^{-}& =f(\widehat{x},u),P^{-}=FP{F}^{\top }+Q,\hfill \\ \hfill K& =P^{-}{H}^{\top }{\left(H{P}^{-}{H}^{\top }+R\right)}^{-1},\hfill \\ \hfill \widehat{x}& ={\widehat{x}}^{-}+K\left(z-H{\widehat{x}}^{-}\right),P=(I-KH)P^{-}\hfill \end{array}$$

(11)

where F is the Jacobian of f (EKF), or the unscented transform is used instead (UKF). Figure 9a shows the filter structure and the outage switch.

Algorithm 4 Fusion experiment (E5)

1: For run i = 0, …, 49: generate the true trajectory, odometry and GNSS with seed i. 2: Run all five estimators on identical measurements. 3: Compute the RMSE (Eq. 3) overall, while GNSS is available, during the outage, and the 95th-percentile error. 4: Aggregate with t CIs (Eq. 7); run Wilcoxon tests over the 50 runs for EKF vs GNSS (overall, outage) and EKF vs UKF, with Holm correction.

Results. Table VI and Figure 10 show the results. Dead reckoning diverges (420 m RMSE) because gyro bias integrates into heading error. GNSS alone is unbiased but noisy (11.2 m) and fails during the outage (265.5 m). Fusion reduces RMSE to 2.7 m with GNSS available and to 30.2 m during the outage. The EKF improves on GNSS alone in all 50 runs ($r_{\mathrm{rb}}$ = $-1$, $p_{\mathrm{Holm}}$ = 5.3 ×${10}^{-15}$), supporting H5a. The EKF–UKF comparison illustrates the difference between statistical and practical significance. The overall RMSE difference is significant ($p_{\mathrm{Holm}}$ = 0.0011) because the 50 paired runs are highly correlated, but its median is only 7 mm, 0.06% of the error, and it reverses sign between the nominal (EKF better by 0.07 m) and outage periods. We therefore regard H5b as supported in practical terms.

Table VI. Position RMSE in Metres Over 50 Monte Carlo Runs, Mean [95% CI] (E5)

Estimator	Overall	GNSS available	20 s outage	95th percentile
Odometry only	420.0 [410.3, 429.7]	437.2 [427.2, 447.1]	319.7 [311.1, 328.4]	756.1
GNSS only	108.6 [108.3, 108.9]	11.22 [11.15, 11.29]	265.5 [264.8, 266.3]	321.6
Linear KF (GNSS only)	119.2 [117.3, 121.1]	4.80 [4.72, 4.88]	292.5 [287.9, 297.2]	337.2
EKF (odometry + GNSS)	12.57 [11.85, 13.29]	2.71 [2.66, 2.77]	30.24 [28.42, 32.05]	34.2
UKF (odometry + GNSS)	12.56 [11.84, 13.29]	2.78 [2.73, 2.83]	30.18 [28.36, 32.00]	34.2

Table VI. Position RMSE in Metres Over 50 Monte Carlo Runs, Mean [95% CI] (E5)

Estimator	Overall	GNSS available	20 s outage	95th percentile
Odometry only	420.0 [410.3, 429.7]	437.2 [427.2, 447.1]	319.7 [311.1, 328.4]	756.1
GNSS only	108.6 [108.3, 108.9]	11.22 [11.15, 11.29]	265.5 [264.8, 266.3]	321.6
Linear KF (GNSS only)	119.2 [117.3, 121.1]	4.80 [4.72, 4.88]	292.5 [287.9, 297.2]	337.2
EKF (odometry + GNSS)	12.57 [11.85, 13.29]	2.71 [2.66, 2.77]	30.24 [28.42, 32.05]	34.2
UKF (odometry + GNSS)	12.56 [11.84, 13.29]	2.78 [2.73, 2.83]	30.18 [28.36, 32.00]	34.2

2). E6: Effect of Pose Accuracy on Occupancy-Grid Mapping

Objective and hypothesis. To measure how pose error propagates into map quality. H6: landmark-corrected EKF poses yield significantly better maps than odometry.

Setup. A vehicle traverses a closed 656-pose route (0.1 s steps) in an 80 m × 60 m world with boundary walls and five rectangular obstacles, carrying a ${360}^{\circ }$ lidar (180 beams, 25 m range, $\sigma$ = 3 cm). An occupancy grid (0.25 m cells) is updated every tenth pose in log-odds form [9,43]:

$$\begin{array}{cc}\hfill l_{t,i}& =l_{t-1,i}+log{\displaystyle \frac{p(m_i\mid z_t,x_t)}{1-p(m_i\mid z_t,x_t)}},\hfill \\ \hfill p_{\mathrm{occ}}& =0.7,p_{\mathrm{free}}=0.4\hfill \end{array}$$

(12)

with log-odds clipped to $[-4$, 4]. Three pose sources are compared: ground truth; odometry with a 2% scale error, a 0.004 rad/s yaw-rate bias and noise; and an EKF that corrects the same odometry with range-bearing observations ($\sigma$ = 0.3 m, $2^{\circ }$) of 12 known landmarks, modelled as

$$\begin{array}{cc}\hfill r& =\sqrt{{(x_l-x)}^2+{(y_l-y)}^2},\hfill \\ \hfill \varphi & =atan2(y_l-y,x_l-x)-\theta \hfill \end{array}$$

(13)

The procedure repeats trajectory noise, odometry, observations and mapping for 10 seeds. Map quality is the IoU of occupied cells against the true map, evaluated on cells observed with ground-truth poses. Figure 9b shows the loop. Script: exp06_localization_mapping.py.

Results. Table VII and Figure 11 show the results. Odometry drift (4.61 m RMSE, 10.1 m final error) destroys map structure (IoU 0.086) even though 95% of observed cells are still classified correctly, which shows why cell accuracy is a misleading map metric. The EKF reduces pose RMSE to 0.277 m and raises IoU to 0.502. It improves IoU in all ten seeds (Wilcoxon p = 0.002, $r_{\mathrm{rb}}$ = 1), supporting H6. The remaining gap to perfect poses (0.994) indicates that sub-cell pose accuracy is needed for crisp maps, which motivates joint pose-graph optimisation with loop closure [41,42].

Table VII. Occupancy Mapping Over 10 Seeds, Mean [95% CI] (E6)

Pose source	Pose RMSE (m)	Final pose error (m)	Cell accuracy	Occupied-cell IoU
Ground truth	0	0	1.000	0.994 [0.993, 0.996]
Odometry	4.61 [4.23, 4.98]	10.11 [9.23, 10.99]	0.951 [0.950, 0.952]	0.086 [0.076, 0.096]
EKF with landmarks	0.277 [0.269, 0.285]	0.18 [0.11, 0.25]	0.979 [0.977, 0.980]	0.502 [0.473, 0.531]

Table VII. Occupancy Mapping Over 10 Seeds, Mean [95% CI] (E6)

Pose source	Pose RMSE (m)	Final pose error (m)	Cell accuracy	Occupied-cell IoU
Ground truth	0	0	1.000	0.994 [0.993, 0.996]
Odometry	4.61 [4.23, 4.98]	10.11 [9.23, 10.99]	0.951 [0.950, 0.952]	0.086 [0.076, 0.096]
EKF with landmarks	0.277 [0.269, 0.285]	0.18 [0.11, 0.25]	0.979 [0.977, 0.980]	0.502 [0.473, 0.531]

C. RQ2: Prediction, Planning and Decision-Making

1). E7: Physics-Based, Learned, Unimodal and Multimodal Trajectory Prediction

Objective and hypothesis. To determine whether learned and multimodal predictors improve on physics-based baselines, and why. H7a: the learned MLP has a lower ADE than constant velocity (CV). H7b: the multimodal predictor’s advantage comes from outputting several hypotheses rather than from a better model.

Setup. Trajectories are generated at 10 Hz for seven manoeuvres: keep lane, left and right turn, brake to stop, left and right lane change, and accelerate. Initial speed is uniform in 6–16 m/s, and acceleration ($\sigma$ = 0.15 $\mathrm{m}/{\mathrm{s}}^2$), yaw-rate noise ($\sigma$ = 0.01 rad/s) and position noise ($\sigma$ = 5 cm) are added. Manoeuvre onset is drawn uniformly between 0.5 s before and 0.8 s after the prediction time, so intent is sometimes invisible in the history. Each trajectory is normalised to the agent’s pose at the prediction time; models observe 2 s and predict 3 s. Table VIII lists the models; they are trained on 12,000 and tested on 3,000 trajectories per seed.

Table VIII. Trajectory Predictors Compared in E7

Model	Type	Outputs	Configuration
Constant velocity (CV)	Physics	1	Velocity from the last 0.5 s
CTRV	Physics	1	Speed and yaw rate from the last 1 s
MLP regressor	Learned	1	2 × 256 ReLU layers, Adam, early stopping, standardised inputs and targets
k-NN top-1	Learned (library)	1	Nearest training history (Euclidean, standardised)
k-NN library	Learned (library)	6	Futures of the six nearest histories; scored with $minAD{E}_6$

Table VIII. Trajectory Predictors Compared in E7

Model	Type	Outputs	Configuration
Constant velocity (CV)	Physics	1	Velocity from the last 0.5 s
CTRV	Physics	1	Speed and yaw rate from the last 1 s
MLP regressor	Learned	1	2 × 256 ReLU layers, Adam, early stopping, standardised inputs and targets
k-NN top-1	Learned (library)	1	Nearest training history (Euclidean, standardised)
k-NN library	Learned (library)	6	Futures of the six nearest histories; scored with $minAD{E}_6$

Figure 12 shows how the history is normalised, which predictors produce single or multiple hypotheses, and which metrics apply to each.

Procedure. For seed s ∈ {7, …, 11}, a new dataset is generated, the MLP is trained with seed s, and all models are scored with Eq. (4). Metrics are averaged over test trajectories within a seed, and t CIs are computed over the five seeds. Paired Wilcoxon tests over the 3,000 trajectories of seed 7 compare MLP vs CV, k-NN-6 vs MLP and k-NN-1 vs MLP, with Holm correction. Script: exp07_trajectory_prediction.py.

Results. Table IX and Figure 13 show the results. The MLP has the lowest mean ADE among single-output models (2.49 m versus 2.78 m for CV, with non-overlapping CIs), but its per-trajectory median difference from CV is not significant ($p_{\mathrm{Holm}}$ = 0.72). It also has the highest miss rate (86.6%). Both observations follow from mean regression: under ambiguous intent, a squared-error model predicts the conditional mean, which reduces large errors on average but rarely lands within 2 m of any actual outcome (Figure 13c). H7a is therefore only partly supported. The multimodal library halves ADE and FDE (1.43 m and 3.56 m) and has the lowest miss rate (45.3%). Its single-hypothesis variant, however, is the worst model of all (ADE 3.56 m; worse than MLP with $r_{\mathrm{rb}}$ = 0.49, $p_{\mathrm{Holm}}$ < ${10}^{-116}$). This supports H7b: the benefit comes from covering several futures, not from a better predictor, which is why forecasting benchmarks and planners must consume trajectory distributions [45,46].

Table IX. Trajectory Prediction Over a 3 s Horizon, Mean [95% CI] Over 5 Seeds (E7)

Model	ADE (m)	FDE (m)	Miss rate (%)
Constant velocity	2.782 [2.756, 2.807]	7.60 [7.52, 7.67]	68.9 [67.6, 70.3]
CTRV	2.812 [2.792, 2.833]	7.72 [7.66, 7.79]	79.8 [78.7, 80.8]
MLP (single output)	2.490 [2.458, 2.522]	6.90 [6.79, 7.01]	86.6 [84.0, 89.3]
k-NN top-1	3.563 [3.478, 3.647]	9.74 [9.51, 9.98]	82.4 [81.7, 83.1]
k-NN library, best of 6	1.428 [1.400, 1.455]	3.56 [3.49, 3.63]	45.3 [43.7, 46.9]

Table IX. Trajectory Prediction Over a 3 s Horizon, Mean [95% CI] Over 5 Seeds (E7)

Model	ADE (m)	FDE (m)	Miss rate (%)
Constant velocity	2.782 [2.756, 2.807]	7.60 [7.52, 7.67]	68.9 [67.6, 70.3]
CTRV	2.812 [2.792, 2.833]	7.72 [7.66, 7.79]	79.8 [78.7, 80.8]
MLP (single output)	2.490 [2.458, 2.522]	6.90 [6.79, 7.01]	86.6 [84.0, 89.3]
k-NN top-1	3.563 [3.478, 3.647]	9.74 [9.51, 9.98]	82.4 [81.7, 83.1]
k-NN library, best of 6	1.428 [1.400, 1.455]	3.56 [3.49, 3.63]	45.3 [43.7, 46.9]

2). E8: Heuristic Graph Search

Objective and hypothesis. To quantify the effect of heuristic admissibility and weighting in A* [51]. H8: the tight admissible octile heuristic expands significantly fewer nodes than Dijkstra’s algorithm [50] and the Euclidean heuristic without losing optimality, while inadmissible and weighted variants trade path cost for speed.

Setup and procedure. Fifty 200 × 200 grids are generated with seed 11, each with 45 random rectangular obstacles (sides 4–28 cells), and 49 of them admit a corner-to-corner path. Search is 8-connected with unit and $\surd 2$ step costs and no corner cutting. Five planners expand nodes in order of

$$\begin{array}{cc}\hfill f\left(n\right)& =g\left(n\right)+\epsilon h\left(n\right),\hfill \\ \hfill h_{\mathrm{oct}}\left(n\right)& =max\left(\right|\Delta x|,|\Delta y\left|\right)+(\sqrt{2}-1)min\left(\right|\Delta x|,|\Delta y\left|\right)\hfill \end{array}$$

(14)

with h = 0 (Dijkstra), Manhattan (inadmissible on 8-connected grids), Euclidean, octile ($\epsilon$ = 1), and weighted octile ($\epsilon$ = 2). Every planner solves every map; we record nodes expanded, runtime and excess cost relative to Dijkstra, and run Wilcoxon tests over maps with Holm correction. Script: exp08_path_planning.py. Figure 14 shows the search loop.

Results. Table X and Figure 15 show the results. The octile heuristic expands 3.84 times fewer nodes than Dijkstra (8,044 versus 30,884) and 32% fewer than Euclidean, and it returns optimal paths on all maps ($p_{\mathrm{Holm}}$≤ 1.1 ×${10}^{-9}$, $r_{\mathrm{rb}}$ = $-1$). This supports H8. The inadmissible Manhattan heuristic and weighted A* expand about 15 times fewer nodes than octile A*, at a mean cost increase of 1.9–2.0% and a worst case of 9.5%, far below the theoretical bound of 100% for $\epsilon$ = 2 but material for safety-relevant clearance decisions.

Table X. Grid Planning Over 49 Solvable Maps, Mean [95% CI] (E8)

Planner	Nodes expanded	Runtime (ms)	Excess cost (%)	Worst excess (%)
Dijkstra (h = 0)	30,884	122.2	0 (optimal)	0
A*, Manhattan	539	2.8	1.97 [1.25, 2.70]	9.5
A*, Euclidean	11,853	84.5	0 (optimal)	0
A*, octile	8,044	43.2	0 (optimal)	0
Weighted A*, octile, $\epsilon$ = 2	518	3.2	1.90 [1.21, 2.58]	8.4

Table X. Grid Planning Over 49 Solvable Maps, Mean [95% CI] (E8)

Planner	Nodes expanded	Runtime (ms)	Excess cost (%)	Worst excess (%)
Dijkstra (h = 0)	30,884	122.2	0 (optimal)	0
A*, Manhattan	539	2.8	1.97 [1.25, 2.70]	9.5
A*, Euclidean	11,853	84.5	0 (optimal)	0
A*, octile	8,044	43.2	0 (optimal)	0
Weighted A*, octile, $\epsilon$ = 2	518	3.2	1.90 [1.21, 2.58]	8.4

3). E9: Learned Anticipation Versus Rules for Lane-Change Decisions

Objective and hypothesis. To test whether a learned policy outperforms hand-written rules when both have identical information. H9: Q-learning [53] achieves a significantly lower crash rate than a rule with the same lookahead.

Setup. A three-lane highway is discretised into cells. Slower traffic approaches the ego vehicle by one cell per step, and each lane of a newly spawned row is occupied with probability 0.10 (never all three). Actions are keep lane, move left and move right. The reward is +1 per step, $-0.2$ per lane change, $-0.5$ for an invalid move and $-20$ for a collision, which ends the episode; episodes last at most 200 steps. The state is the ego lane plus the occupancy of the three cells ahead in each lane (1,536 states). The tabular Q-learning update is

$$\begin{array}{cc}\hfill Q(s,a)& \leftarrow Q(s,a)+\alpha \left[r+\gamma \underset{a^{\prime }}{max}Q(s^{\prime },a^{\prime })-Q(s,a)\right],\hfill \\ \hfill \alpha & =0.1,\gamma =0.95\hfill \end{array}$$

(15)

Four policies are compared: uniform random; a reactive rule that changes to a free adjacent lane when the next cell is blocked (one-cell lookahead); a gap-seeking rule that moves to the reachable lane with the longest free gap within three cells (the same information as the learned policy); and Q-learning. Figure 16 shows the training and evaluation loops.

Algorithm 5 Training and evaluation of lane-change policies (E9)

1: For training seed s = 0, …, 9: initialise Q = 0 and train for 6,000 episodes with $\epsilon -greedy$ exploration, $\epsilon$ decaying linearly from 1.0 to 0.02 over the first 3,600 episodes. 2: Evaluate all four policies greedily on 500 episodes whose traffic seed (10,000 + s) is never used in training. 3: Record the crash rate, steps survived and lane changes per 100 steps. 4: Aggregate with t CIs over seeds; compare Q-learning with the gap-seeking rule, and the two rules with each other, using Wilcoxon tests over seeds.

Results. Table XI and Figure 17 show the results. Q-learning crashes in 6.3% [5.5, 7.1] of test episodes, versus 59.8% for the gap-seeking rule; it is better in all ten seeds (p = 0.002, median difference $-53$ percentage points), supporting H9. Extra lookahead alone does not help the rule: the gap-seeking and reactive rules do not differ (p = 0.28). The learned policy visits 834 of the 1,536 states and anticipates configurations that become unescapable two steps later, when two lanes close and the free lane is not adjacent. It moves pre-emptively, making 2.9 times as many lane changes as the rules. This behaviour would require comfort and traffic-law constraints before real deployment.

Table XI. Lane-Change Policies on 500 Held-Out Episodes, Mean [95% CI] Over 10 Seeds (E9)

Policy	Lookahead	Crash rate (%)	Steps survived	Lane changes per 100 steps
Random	—	100.0	12.2 [11.9, 12.4]	66.9 [66.3, 67.5]
Reactive rule	1 cell	60.4 [58.4, 62.3]	133.0 [130.3, 135.7]	9.34 [9.27, 9.42]
Gap-seeking rule	3 cells	59.8 [58.7, 60.9]	133.8 [131.6, 136.0]	9.38 [9.30, 9.46]
Q-learning	3 cells	6.3 [5.5, 7.1]	193.5 [192.6, 194.4]	26.8 [24.0, 29.7]

Table XI. Lane-Change Policies on 500 Held-Out Episodes, Mean [95% CI] Over 10 Seeds (E9)

Policy	Lookahead	Crash rate (%)	Steps survived	Lane changes per 100 steps
Random	—	100.0	12.2 [11.9, 12.4]	66.9 [66.3, 67.5]
Reactive rule	1 cell	60.4 [58.4, 62.3]	133.0 [130.3, 135.7]	9.34 [9.27, 9.42]
Gap-seeking rule	3 cells	59.8 [58.7, 60.9]	133.8 [131.6, 136.0]	9.38 [9.30, 9.46]
Q-learning	3 cells	6.3 [5.5, 7.1]	193.5 [192.6, 194.4]	26.8 [24.0, 29.7]

D. RQ3: Motion Control

1). E10: Path-Tracking Controllers Under a Common Tuning Protocol

Objective and hypothesis. To compare geometric and optimisation-based lateral controllers fairly. H10a: model-based controllers (LQR, MPC) achieve the lowest tracking error at both speeds. H10b: MPC produces the smoothest steering.

Setup. The plant is a kinematic bicycle model [61] with wheelbase L = 2.7 m, a first-order steering actuator (time constant 0.1 s), a $\pm {30}^{\circ }$ steering limit, a ${25}^{\circ }/s$ rate limit and pose-measurement noise ($\sigma$ = 3 cm, $0.{17}^{\circ }$), simulated at 20 Hz:

$$\dot{x}=vcos\psi ,\dot{y}=vsin\psi ,\dot{\psi }={\displaystyle \frac{v}{L}}tan\delta$$

(16)

Pure Pursuit [59] steers toward a look-ahead point at distance $L_d$ = max(2, $k_v·v$ + 2), and Stanley [60] combines heading error ${\psi }_e$ with front-axle cross-track error e:

$${\delta }_{\mathrm{PP}}=arctan{\displaystyle \frac{2Lsin\alpha }{L_d}},{\delta }_{\mathrm{St}}={\psi }_e+arctan{\displaystyle \frac{ke}{1+v}}$$

(17)

The LQR [62] and MPC [63] act on the linearised error dynamics with curvature $\kappa$ and feed-forward ${\delta }_{\mathrm{ff}}$ = $atan\left(L\kappa \right)$. LQR uses the discrete-time Riccati solution with Q = diag(1, 2) and input weight R:

$$\begin{array}{cc}\hfill {\dot{e}}_y& =v{e}_{\psi },{\dot{e}}_{\psi }={\displaystyle \frac{v}{L}}\delta -v\kappa ,\hfill \\ \hfill u& ={\delta }_{\mathrm{ff}}-K{[e_y,e_{\psi }]}^{\top }\hfill \end{array}$$

(18)

MPC solves, at every step, a 20-step (1 s) horizon problem with CVXPY and OSQP [65,66]:

$$\begin{array}{cc}\hfill \underset{{\delta }_0,\dots ,{\delta }_{19}}{min}\sum _k[& e_{y,k+1}^2+2e_{\psi ,k+1}^2\hfill \\ & +2{({\delta }_k-{\delta }_{\mathrm{ff},k})}^2+S{({\delta }_k-{\delta }_{k-1})}^2]\hfill \end{array}$$

(19)

subject to the linearised dynamics of Eq. (18), $|{\delta }_k|$≤${30}^{\circ }$ and $|{\delta }_k$−${\delta }_{k-1}|$≤${25}^{\circ }·\Delta t$.

The test reference (430 m) comprises a 3.5 m lane change and ${90}^{\circ }$ left and right curves (minimum radius 45 m). A separate validation reference (320 m) has a shorter lane change and curves in the opposite order with different radii. The first second of each run is excluded from the metrics. Figure 18 shows the closed loop and the outer tuning loop.

Algorithm 6 Controller tuning and testing (E10)

1: For each controller, define a grid over its key parameter (Table XII). 2: For each grid value, simulate the validation path at 12 m/s with noise seeds 100 and 101; compute J (Eq. 6). Runs that diverge ($|e_y|$ > 5 m) receive J = ∞. 3: Select the grid value with the lowest mean J; check that it is not at the edge of the grid (the grid was extended when it was). 4: Simulate the selected controllers on the test path at 8 and 15 m/s with noise seeds 0–9. 5: Report the RMS lateral error, maximum error, RMS steering rate (Eq. 5) and compute time with t CIs; test MPC against each other controller at 15 m/s with Wilcoxon tests and Holm correction.

Results. Table XII reports the tuning results. The Pure Pursuit and LQR grids initially selected values at their edges, so both grids were extended until the optimum was interior; LQR with R = 1 diverged on the validation path. Table XIII and Figure 19 report the test results. No controller dominates, and H10a is not supported. At 8 m/s, Pure Pursuit is the most accurate (2.2 cm RMS) and Stanley the least (4.6 cm); at 15 m/s the ranking reverses, with Stanley best (1.9 cm) and LQR worst (6.4 cm). Stanley’s effective gain k/(1 + v) roughly doubles as speed halves, which makes it oscillatory at low speed after tuning at 12 m/s. H10b is supported: MPC has the lowest steering rate at both speeds (3.0 and $3.5^{\circ }/s$), is second most accurate at both speeds, and beats LQR in all ten seeds on both error ($-0.8$ cm) and smoothness ($-0.{97}^{\circ }/s$; $p_{\mathrm{Holm}}$ = 0.008). The price of MPC is computation, 3.4 ms per step (99th percentile 4.5 ms) versus under 0.03 ms for the other controllers, although this is well within the 50 ms control period. Importantly, these rankings differ from those in a preliminary run of this experiment that used hand-picked gains, in which LQR appeared best. Unequal tuning is thus a first-order confounder in controller comparisons.

Table XII. Validation Tuning of Controller Parameters at 12 m/s (E10)

Controller	Tuned parameter	Grid	Selected	J at selection
Pure Pursuit	Look-ahead gain $k_v$ (s)	0.1, 0.2, 0.3, 0.45, 0.6, 0.8	0.3	0.110
Stanley	Cross-track gain k	0.5, 1, 2, 3, 5	2.0	0.073
LQR + feed-forward	Input weight R	1 (diverged), 4, 16, 32, 64, 128, 256	32	0.116
Linear MPC	Steering-rate weight S	5, 20, 40, 80, 160	40	0.097

Table XII. Validation Tuning of Controller Parameters at 12 m/s (E10)

Controller	Tuned parameter	Grid	Selected	J at selection
Pure Pursuit	Look-ahead gain $k_v$ (s)	0.1, 0.2, 0.3, 0.45, 0.6, 0.8	0.3	0.110
Stanley	Cross-track gain k	0.5, 1, 2, 3, 5	2.0	0.073
LQR + feed-forward	Input weight R	1 (diverged), 4, 16, 32, 64, 128, 256	32	0.116
Linear MPC	Steering-rate weight S	5, 20, 40, 80, 160	40	0.097

Table XIII. Path Tracking on the 430 m Test Path, Mean [95% CI] Over 10 Noise Seeds (E10)

Controller	RMS error 8 m/s (cm)	RMS error 15 m/s (cm)	Max error 15 m/s (cm)	RMS steering rate 8 / 15 m/s ${(}^{\circ }/\mathit{s})$	Compute (ms)
Pure Pursuit	2.2 [2.1, 2.2]	5.8 [5.7, 5.9]	12.2	6.59 / 3.67	0.01
Stanley	4.6 [4.5, 4.6]	1.9 [1.9, 2.0]	5.7	6.02 / 4.41	0.01
LQR + feed-forward	3.5 [3.5, 3.6]	6.4 [6.3, 6.5]	13.7	4.18 / 4.41	0.01
Linear MPC (N = 20)	2.8 [2.7, 2.8]	5.6 [5.5, 5.7]	12.4	3.00 / 3.46	3.4

Table XIII. Path Tracking on the 430 m Test Path, Mean [95% CI] Over 10 Noise Seeds (E10)

Controller	RMS error 8 m/s (cm)	RMS error 15 m/s (cm)	Max error 15 m/s (cm)	RMS steering rate 8 / 15 m/s ${(}^{\circ }/\mathit{s})$	Compute (ms)
Pure Pursuit	2.2 [2.1, 2.2]	5.8 [5.7, 5.9]	12.2	6.59 / 3.67	0.01
Stanley	4.6 [4.5, 4.6]	1.9 [1.9, 2.0]	5.7	6.02 / 4.41	0.01
LQR + feed-forward	3.5 [3.5, 3.6]	6.4 [6.3, 6.5]	13.7	4.18 / 4.41	0.01
Linear MPC (N = 20)	2.8 [2.7, 2.8]	5.6 [5.5, 5.7]	12.4	3.00 / 3.46	3.4

E. RQ4: Field Evidence and Fleet Learning

1). E11: Statistical Re-Analysis of Public AV Safety Data

Objective and hypothesis. To verify the published safety comparison independently and to quantify the evidence it can support. H11a: rider-only crash rates are significantly below human benchmarks for every crash type. H11b: current exposure is insufficient to demonstrate a fatality-rate reduction.

Setup and procedure. For each of nine crash types with any reported injury, the hub provides the observed Waymo count k and the benchmark-expected count E, which is the human benchmark rate applied to Waymo’s mileage [6,17]. We computed the rate ratio k/E with the exact interval of Eq. (9), treating E as fixed, and aggregated over types. For exposure, a fleet with true rate (1 −$\rho )·r_h$ that observes its expected count k after n miles demonstrates a lower rate than the human rate $r_h$ at 95% confidence when

$${\displaystyle \frac{{\chi }_{0.95}^2(2k+2)}{2n}}<r_h;k=0:n>{\displaystyle \frac{-ln0.05}{r_h}}$$

(20)

We solve Eq. (20) for n numerically, for $\rho$∈ [0.2, 0.95] and three outcomes: fatalities ($r_h$ = 1.19 per 100 million VMT [1]), serious-injury-or-worse crashes (0.23 IPMM) and any-injury-reported crashes (3.91 IPMM) [6]. Script: exp11_public_safety_data.py. Figure 20 traces the provenance of every quantity used.

Results. Through March 2026 the service logged 220.6 million RO miles (Figure 21a), and crash rates are below the benchmark in every city (Figure 21b). Across types, 142 crashes were observed against 834 expected, a ratio of 0.170 [0.143, 0.201] or an 83.0% reduction, consistent with the hub’s 81.9% headline within rounding. Every type’s upper CI limit is below 1 (Table XIV, Figure 21c), supporting H11a. The effect is heterogeneous. Intersection $(-96.2\%)$, single-vehicle $(-96.7\%)$ and pedestrian $(-92.6\%)$ crashes fall most, whereas front-to-rear $(-47.5\%)$ and secondary crashes $(-58.7\%)$ fall least; these categories often involve the AV being struck, which bounds what the AV’s own driving can prevent. For exposure (Table XV, Figure 22b), demonstrating an 80% fatality reduction requires about 388 million miles, and 252 million even with zero fatalities, both more than the current total. Current exposure is ample for injury outcomes (1.2 million miles for an 80% reduction), which supports H11b. California DMV reports (Figure 22a) show testing mileage of 4.5–9.1 million miles per year, two orders of magnitude short of fatality-level evidence.

Table XIV. Any-Injury-Reported Crashes by Type: Observed Versus Benchmark-Expected (E11)

Crash type	Observed k	Expected E	Ratio k/E [exact 95% CI]	Reduction (%)
Vehicle-to-vehicle intersection	13	340	0.038 [0.020, 0.065]	96.2
Single vehicle	2	60	0.033 [0.004, 0.120]	96.7
Pedestrian	6	81	0.074 [0.027, 0.161]	92.6
Motorcycle	6	38	0.158 [0.058, 0.344]	84.2
Cyclist	9	56	0.161 [0.073, 0.305]	83.9
All others	3	17	0.176 [0.036, 0.516]	82.4
Vehicle-to-vehicle lateral	11	57	0.193 [0.096, 0.345]	80.7
Secondary crash	19	46	0.413 [0.249, 0.645]	58.7
Vehicle-to-vehicle front-to-rear	73	139	0.525 [0.412, 0.660]	47.5
All types	142	834	0.170 [0.143, 0.201]	83.0

Table XIV. Any-Injury-Reported Crashes by Type: Observed Versus Benchmark-Expected (E11)

Crash type	Observed k	Expected E	Ratio k/E [exact 95% CI]	Reduction (%)
Vehicle-to-vehicle intersection	13	340	0.038 [0.020, 0.065]	96.2
Single vehicle	2	60	0.033 [0.004, 0.120]	96.7
Pedestrian	6	81	0.074 [0.027, 0.161]	92.6
Motorcycle	6	38	0.158 [0.058, 0.344]	84.2
Cyclist	9	56	0.161 [0.073, 0.305]	83.9
All others	3	17	0.176 [0.036, 0.516]	82.4
Vehicle-to-vehicle lateral	11	57	0.193 [0.096, 0.345]	80.7
Secondary crash	19	46	0.413 [0.249, 0.645]	58.7
Vehicle-to-vehicle front-to-rear	73	139	0.525 [0.412, 0.660]	47.5
All types	142	834	0.170 [0.143, 0.201]	83.0

Table XV. Miles Needed to Demonstrate a Rate Reduction at 95% One-Sided Confidence (E11)

Outcome	Human rate ${\mathit{r}}_{\mathit{h}}$	$\mathit{\rho }$ = 50%	$\mathit{\rho }$ = 80%	Zero events observed
Fatality	1.19 per 100 M VMT	969 M	388 M	252 M
Serious injury or worse	0.23 IPMM	50.2 M	20.1 M	—
Any injury reported	3.91 IPMM	2.95 M	1.18 M	—

Table XV. Miles Needed to Demonstrate a Rate Reduction at 95% One-Sided Confidence (E11)

Outcome	Human rate ${\mathit{r}}_{\mathit{h}}$	$\mathit{\rho }$ = 50%	$\mathit{\rho }$ = 80%	Zero events observed
Fatality	1.19 per 100 M VMT	969 M	388 M	252 M
Serious injury or worse	0.23 IPMM	50.2 M	20.1 M	—
Any injury reported	3.91 IPMM	2.95 M	1.18 M	—

2). E12: Federated Learning Under Non-IID Fleet Data

Objective and hypothesis. To measure how much of the benefit of pooling fleet data federated averaging (FedAvg) [83] recovers when vehicles see systematically different conditions. H12a: FedAvg outperforms isolated local training. H12b: more local epochs improve FedAvg under non-IID data within a fixed number of rounds.

Setup. The 250 training frames of E3 are divided among ten simulated vehicles, 25 frames each. In the non-IID split, frames are sorted by mean brightness, so client means range from 18 to 119 grey levels; an IID random split serves as a reference. Each client trains a per-pixel MLP (26 → 32 → 1, ReLU, 897 parameters) with SGD (learning rate 0.05, batch size 256) on 1,200 sampled pixels per frame, using the E3 features. The server aggregates client models $w^{\left(k\right)}$ holding $n_k$ samples:

$$w_{t+1}=\sum _k{\displaystyle \frac{n_k}{n}}{w}_{t+1}^{\left(k\right)}$$

(21)

Figure 23 shows the architecture.

Algorithm 7 Federated learning experiment (E12)

1: For seed s = 0, …, 4: sample pixels, form the IID split and initialise a common model with seed s. 2: Centralised: train on the pooled data for 40 epochs (the same number of gradient steps per round as FedAvg with E = 1). 3: FedAvg (non-IID with E = 1 and E = 5; IID with E = 1): for 40 rounds, every client trains E local epochs from the global model and the server averages the models (Eq. 21). 4: Local-only: each client trains alone for 40 epochs. 5: After every round, evaluate IoU on the 150 test frames; report final IoU and the share of the local-to-centralised gap recovered, with t CIs over seeds.

Results. Table XVI and Figure 24 show the results. Isolated training generalises poorly (IoU 0.616 [0.601, 0.631]), and the clients with the darkest and brightest data are worst (0.555 and 0.524). FedAvg with one local epoch reaches 0.716 and recovers 54.1% [50.8, 57.4] of the gap to centralised training (0.800), supporting H12a. Five local epochs reach 0.753 and recover 74.4% [72.6, 76.3], nearly matching FedAvg on IID data (77.1%), which supports H12b within the 40-round budget. The ordering holds in all five seeds; with n = 5, the smallest attainable two-sided Wilcoxon p-value is 0.0625, so this conclusion rests on the non-overlapping CIs rather than on p < 0.05. Communication cost is 35.9 kB per round for all ten clients.

Table XVI. Federated Drivable-Area Segmentation After 40 Rounds, Mean [95% CI] Over 5 Seeds (E12)

Regime	Raw data shared	Test IoU	Gap recovered (%)
Local-only (mean of 10 clients)	No	0.616 [0.601, 0.631]	0 (reference)
FedAvg, non-IID, E = 1	No	0.716 [0.708, 0.723]	54.1 [50.8, 57.4]
FedAvg, non-IID, E = 5	No	0.753 [0.750, 0.756]	74.4 [72.6, 76.3]
FedAvg, IID, E = 1 (reference)	No	0.758 [0.751, 0.764]	77.1 [75.0, 79.2]
Centralised (pooled)	Yes	0.800 [0.799, 0.801]	100 (reference)

Table XVI. Federated Drivable-Area Segmentation After 40 Rounds, Mean [95% CI] Over 5 Seeds (E12)

Regime	Raw data shared	Test IoU	Gap recovered (%)
Local-only (mean of 10 clients)	No	0.616 [0.601, 0.631]	0 (reference)
FedAvg, non-IID, E = 1	No	0.716 [0.708, 0.723]	54.1 [50.8, 57.4]
FedAvg, non-IID, E = 5	No	0.753 [0.750, 0.756]	74.4 [72.6, 76.3]
FedAvg, IID, E = 1 (reference)	No	0.758 [0.751, 0.764]	77.1 [75.0, 79.2]
Centralised (pooled)	Yes	0.800 [0.799, 0.801]	100 (reference)

V. Discussion

Figure 25 maps each research question to its hypotheses, experiments and outcomes, and Table XVII lists the supporting evidence. We now answer the research questions and draw practical implications.

Table XVII. Summary of Hypotheses and Outcomes

Hypothesis	Outcome	Key evidence
H1: lane markings occupy under 1% of pixels	Supported	0.588% [0.561, 0.613]
H2a: held-out lane F1 exceeds 0.6	Supported	F1 = 0.658 [0.620, 0.694]
H2b: lane F1 decreases in darker frames	Not supported	Kruskal–Wallis p = 0.93
H3: each feature group raises segmentation IoU	Supported	+0.071 and +0.022; $p_{\mathrm{Holm}}$ < ${10}^{-11}$
H5a: fusion beats GNSS, especially in outages	Supported	30.2 m vs 265.5 m; $p_{\mathrm{Holm}}$ = 5.3 × ${10}^{-15}$
H5b: EKF and UKF are equivalent	Supported (practically)	Median $\Delta$ = 7 mm, significant but negligible
H6: EKF poses improve maps over odometry	Supported	IoU 0.502 vs 0.086; 10/10 seeds
H7a: learned MLP beats constant velocity	Partly supported	Lower mean ADE; median $\Delta$ n.s.; highest miss rate
H7b: multimodal gain comes from hypothesis count	Supported	k-NN-6 best; k-NN-1 worst
H8: octile A* is faster and still optimal	Supported	$3.84\times$ fewer nodes than Dijkstra; 0% excess
H9: Q-learning beats a rule with identical inputs	Supported	6.3% vs 59.8% crashes; p = 0.002
H10a: LQR and MPC are most accurate at both speeds	Not supported	Rankings reverse with speed
H10b: MPC produces the smoothest steering	Supported	Lowest steering rate at both speeds
H11a: rates below benchmark for all crash types	Supported	All 9 upper CI limits < 1
H11b: exposure insufficient for fatality claims	Supported	388 M miles needed vs 220.6 M driven
H12a: FedAvg beats local training	Supported	0.716 vs 0.616; CIs disjoint
H12b: more local epochs help under non-IID data	Supported	74.4% vs 54.1% of gap recovered

Table XVII. Summary of Hypotheses and Outcomes

Hypothesis	Outcome	Key evidence
H1: lane markings occupy under 1% of pixels	Supported	0.588% [0.561, 0.613]
H2a: held-out lane F1 exceeds 0.6	Supported	F1 = 0.658 [0.620, 0.694]
H2b: lane F1 decreases in darker frames	Not supported	Kruskal–Wallis p = 0.93
H3: each feature group raises segmentation IoU	Supported	+0.071 and +0.022; $p_{\mathrm{Holm}}$ < ${10}^{-11}$
H5a: fusion beats GNSS, especially in outages	Supported	30.2 m vs 265.5 m; $p_{\mathrm{Holm}}$ = 5.3 × ${10}^{-15}$
H5b: EKF and UKF are equivalent	Supported (practically)	Median $\Delta$ = 7 mm, significant but negligible
H6: EKF poses improve maps over odometry	Supported	IoU 0.502 vs 0.086; 10/10 seeds
H7a: learned MLP beats constant velocity	Partly supported	Lower mean ADE; median $\Delta$ n.s.; highest miss rate
H7b: multimodal gain comes from hypothesis count	Supported	k-NN-6 best; k-NN-1 worst
H8: octile A* is faster and still optimal	Supported	$3.84\times$ fewer nodes than Dijkstra; 0% excess
H9: Q-learning beats a rule with identical inputs	Supported	6.3% vs 59.8% crashes; p = 0.002
H10a: LQR and MPC are most accurate at both speeds	Not supported	Rankings reverse with speed
H10b: MPC produces the smoothest steering	Supported	Lowest steering rate at both speeds
H11a: rates below benchmark for all crash types	Supported	All 9 upper CI limits < 1
H11b: exposure insufficient for fatality claims	Supported	388 M miles needed vs 220.6 M driven
H12a: FedAvg beats local training	Supported	0.716 vs 0.616; CIs disjoint
H12b: more local epochs help under non-IID data	Supported	74.4% vs 54.1% of gap recovered

RQ1: perception. Learned models gain from spatial context in a graded way. A position prior alone explains most drivable-area performance (0.701 IoU), appearance adds 0.07, and local context a further 0.02. Each step is statistically robust but smaller than the last. Thin structures are extremely under-represented (lane markings are 0.6% of pixels), so loss weighting and instance-level metrics matter. The classical lane detector is adequate on median frames but fails abruptly on a sizeable minority (7.2% no detection), and its failures are not explained by global brightness. Robustness evaluations should therefore stratify by the physical causes of failure, such as marking wear, occlusion and glare, rather than by simple image statistics.

RQ2: estimation, prediction and planning. Sensor fusion provides the largest effect in the study, an order-of-magnitude error reduction during outages, whereas the choice between EKF and UKF is immaterial for mildly nonlinear vehicle models. In prediction, the key variable is output representation rather than model capacity. A single-output regressor trained on ambiguous futures averages them, and a multimodal output with six hypotheses beats every single-output model even when built from the worst single-output component. In planning, tight admissible heuristics deliver speed without sacrificing optimality, while inadmissible shortcuts carry bounded but non-zero cost. In decision-making, learning discovered an anticipatory strategy that hand-written rules with the same inputs missed, but at the cost of 2.9 times more lane changes, a behaviour that reward design must regulate.

RQ3: control. The central finding is methodological. When each controller was tuned by the same validation procedure, accuracy rankings reversed between 8 and 15 m/s, and they differed from rankings obtained with hand-picked gains. The only consistent advantage was MPC’s smoothness, obtained at about 3.4 ms per step. Controller comparisons should therefore report their tuning protocol and evaluate across the operating envelope; gain scheduling of geometric controllers with speed would be a natural next step.

RQ4: fleet learning and field evidence. Federated learning recovered up to three quarters of the accuracy lost by isolated training without sharing images, and additional local computation partly compensated for heterogeneity. The public field data support a large and statistically robust reduction in injury crashes for one Level 4 service, strongest for intersection and pedestrian crashes. Yet even 220 million miles cannot demonstrate a fatality-rate reduction, which confirms the argument of [16] with current data. Safety cases for rare, severe outcomes must combine field data with simulation, scenario-based testing and standards such as ISO 21448 [80], ISO/PAS 8800 [81] and UL 4600 [82].

Relation to end-to-end and foundation models. The mechanisms isolated here are the same ones that modern end-to-end and vision-language driving models exploit implicitly: context aggregation (E3), multimodal output (E7), anticipation learned from reward or data (E9), and learning from distributed fleets (E12) [12,13,73]. Our results suggest that evaluating such models requires the same safeguards used here: held-out and validation-tuned protocols, multimodal metrics, closed-loop evaluation [15], and effect sizes with uncertainty.

VI. Threats to Validity

Internal validity. Implementations are our own and simplified, so implementation choices could affect rankings. We mitigated this by using identical inputs for all methods, validation-based tuning in E2 and E10, extending tuning grids whose optimum lay at an edge, and releasing code for inspection. Tuning in E10 covered one parameter per controller at one speed, and rankings may change with richer tuning or gain scheduling. The E10 error metric is measured at the rear axle for all controllers, whereas Stanley regulates the front axle, which may favour or penalise it in curves.

External validity. E5–E10 use simplified simulations (kinematic models, grid worlds, synthetic trajectories) that isolate mechanisms but omit real vehicle dynamics, sensor artefacts and interactive traffic. The magnitudes reported here should not be transferred to production systems. comma10k frames come from one camera type and fleet, mostly on U.S. highways. The perception models are deliberately small classical learners rather than deep networks, because no GPU was available, so absolute IoU values understate the state of the art. E11 concerns one operator in five cities with little snow.

Construct validity. IoU at 1/8 resolution, pixel-tolerance lane matching, displacement-based prediction metrics, and RMS error with a steering-rate penalty ($\lambda$ = 0.01) are proxies for driving quality. Different weights would change E10’s selected parameters. The E11 benchmarks cannot be perfectly matched on time of day, trip mix or reporting completeness, and our re-analysis ignores uncertainty in the benchmark itself.

Conclusion validity. We pre-specified comparisons within each experiment and applied Holm correction, but we did not correct across experiments. With many replicates, negligible differences become significant (E5), so we report effect sizes and CIs. With few seeds (E12, n = 5), exact tests cannot reach p < 0.05, and conclusions rest on CIs. The E6 and E9 tests use 10 seeds, which provides limited power for small effects.

VII. Conclusion and Future Work

This paper presented a reproducible, statistically grounded evaluation of classical and AI-based methods across the AV software stack, using public driving images, seeded simulations and public field safety data under one protocol. Of seventeen pre-stated hypotheses, thirteen were supported, one was supported only in practical terms, one partly, and two were not supported. Three findings are especially relevant to practitioners. The advantage of multimodal predictors lies in their output representation. Controller rankings depend on tuning protocol and speed. Finally, current field exposure supports strong injury-reduction claims but not fatality-reduction claims.

Limitations. These findings should be read within the scope of the study. Six of the twelve experiments (E5–E10) use simplified, seeded simulations with kinematic vehicle models, grid worlds and synthetic trajectories. These isolate how the methods behave, but they do not capture real vehicle dynamics, sensor artefacts or interactive traffic, so the reported magnitudes should not be transferred to production systems. The learned perception models (E3, E12) are deliberately small classical learners trained on 250 frames from a single camera type, and the lane detector (E2) is a classical pipeline; no GPU-based deep networks were used. Absolute IoU and F1 values therefore understate the state of the art; the relative effects of context and multimodality are the intended contribution. Controller tuning covered one parameter per controller at a single validation speed (12 m/s), and other tuning budgets or gain scheduling could change the rankings. The field safety analysis concerns one operator in five cities, treats the published benchmarks as exact, and cannot address fatality outcomes with current exposure. In addition, large end-to-end and vision-language driving models were not evaluated directly. The protocol is designed to extend to them, but whether our conclusions hold at that scale remains an open question. Section VI discusses these threats in more detail.

Future work will replace the classical perception learners with deep networks on GPU hardware, evaluate on large multimodal datasets such as nuScenes [35] and the Waymo Open Motion Dataset [46], move planning and control experiments into closed-loop simulation such as CARLA [88] with interactive agents, and extend the protocol to end-to-end and vision-language driving models.