Euro NCAP 2026 · ISO 21448 SOTIF · AI-Augmented Three-Phase Pipeline
Your AEB system passes every NCAP homologation test. You ship it. Six months later a customer has a collision in heavy fog. The fog scenario was never in your test plan: because your previous software never failed there.
How would you have found it before shipping?
This framework answers that question.
Three phases. Each solves a different problem test engineers face today.
Phase 1: Know your failure space
2,000 parametric scenarios across Euro NCAP 2026 families
Speed × weather × overlap × time-of-day × road surface
Every result labelled: NCAP score (0–4) + collision flag
Phase 2: Cut your regression cost
GBT criticality model learns which parameter combinations cause failures
Ranks all 2,000 scenarios by predicted risk
Top 441 scenarios → 90% defect coverage → 4.5× cheaper than exhaustive
Phase 3: Find what you don't know to look for
UCB bandit probes the new software release
Warm-started from Phase 2 model: exploits known risks, explores unknown ones
Discovers failure modes introduced by regression: never seen in training data
Directly targets ISO 21448 SOTIF: unknown-unknown coverage
| SUT | Known defect | Failures / 2,000 | Failure rate |
|---|---|---|---|
| v1 | Wet-road braking compensation bug | 491 | 24.6% |
| v2 | Fog-dense sensor regression (introduced while fixing v1) | 293 | 14.6% |
| v3 | Clean release | 0 | 0.0% |
| Approach | Scenarios | Failures found | Coverage |
|---|---|---|---|
| Exhaustive sweep | 2,000 | 491 | 100% |
| NCAP homologation fixed points (dry only) | 178 | 0 | 0% |
| GBT-smart (AI-ranked) | 441 | 441 | 89.8% |
NCAP homologation found zero failures on a buggy system. The AI-ranked suite found 441 in less than a quarter of the scenarios.
| SUT probed | GBT trained on | Scenarios | Known failures | Novel arms |
|---|---|---|---|---|
| v1 | v1 | 500 | 441 | 1 |
| v2 | v1 | 500 | 61 | 2 |
| v3 | v2 | 300 | 0 | 0 |
The 441→61 drop in known failures proves v1's wet-road bug was fixed in v2. The 2 novel arms prove a new fog-dense regression was introduced: regions the v1 model never predicted as dangerous. Found by the bandit, not by any systematic test plan. v3 produces silence. The framework correctly finds nothing on a clean release.
Established tools (Foretellix, AVL SCENIUS, dSPACE, IPG CarMaker) are strong at high-fidelity simulation, sensor modelling, and regulation compliance workflows. This framework is not a replacement for those — it is designed to sit alongside them as an AI intelligence layer for test selection and regression analysis.
The specific combination not publicly documented in existing tools:
| Capability | Status in market | This framework |
|---|---|---|
| Parametric scenario variation + NCAP labelling | Available in most tools | ✓ |
| ML criticality ranking to reduce test budget | Not publicly documented | ✓ |
| Cross-version regression via shared ML prior | Not publicly documented | ✓ |
| Adaptive bandit for SOTIF unknown-unknown search | Research stage only | ✓ (POC) |
| SHAP root-cause feature importance | Not in test tools | ✓ |
NCAP 2026 Catalog ──► Variation Engine ──► 2,000 ConcreteScenarios
│
┌─────────────────────────┘
▼
SUMO Simulator + SUT Controller
(TraCI loop, physics-accurate AEB)
│
▼
NCAP Evaluator ──► 0–4 pts + label_critical
│
▼
Results DB (SQLite, exportable)
│
┌───────────┴───────────┐
▼ ▼
Phase 2: GBT Phase 3: UCB Bandit
Criticality Model warm-started from GBT
+ SmartSelector + one-shot novelty reward
+ SHAP explainability → ISO 21448 artefact
SUT integration point: SUTController.step(): override to connect your V-ECU
via FMI 2.0 co-simulation. No other changes required.
Simulator agnostic: SUMO used in this POC. Runner interface designed for drop-in replacement with CarMaker, VEOS, CARLA, or any TraCI-compatible backend.
| Standard | Status |
|---|---|
| Euro NCAP 2026 AEB C2C (CCRs/m/b, CPNCO, CPNA, CPFA, CBNAc) | Active |
| UNECE R157 ALKS | Catalog stub ready |
| ISO 21448 SOTIF | Phase 3 bandit generates documented evidence |
| ISO 26262 | FMI 2.0 integration point pre-wired |
config/ NCAP 2026 + R157 scenario catalogs (YAML)
src/
catalog/ Catalog loader
generator/ Parametric variation engine + SUMO writer
simulation/
evaluator.py NCAP scoring (0–4 points)
runner.py SUMO TraCI batch runner [interface only]
sut_controller.py AEB SUT: FMI 2.0 integration point [interface only]
ml/
criticality_model.py GBT criticality classifier [interface only]
selector.py Coverage-aware smart selector [interface only]
rl_agent.py UCB bandit + novelty reward [interface only]
dashboard/ Streamlit result dashboard (fully functional)
data/
results/ Learning curve CSV, summary data
models/ SHAP importance CSVs, metrics JSON
docs/charts/ Result charts (PNG)
Files marked [interface only] expose the full class/method signature and
docstring. The implementation is available under a collaboration agreement.
No simulation required: pre-populated results included.
pip install streamlit plotly pandas sqlalchemy scikit-learn shap
streamlit run src/dashboard/app.py
The current POC validates the three-phase pipeline concept. The natural next step is replacing the synthetic weather dimension with real kinematic parameters that SUMO simulates natively with full physics:
| Current (POC) | Next version |
|---|---|
| Weather as friction multiplier (manual) | Target deceleration profile (physics-accurate) |
| 5 weather condition buckets | CCRs / CCRm / CCRb / Cut-in / Curved-road arms |
| fog_dense as SUT bug proxy | Reaction latency regression or deceleration authority cap |
| Discrete arm, random sampling within | Bayesian Optimisation within each arm (exact failure threshold) |
The pipeline architecture (Phase 1 / Phase 2 / Phase 3) does not change. Only the parameter space and arm definitions are updated.
This demonstrates a methodology, not a production tool.
| # | Limitation | Impact |
|---|---|---|
| 1 | Weather is a friction multiplier, not real sensor physics | Failure boundary is real; root cause is synthetic |
| 2 | SUT is parametric rule-based, not a real ECU | FMI 2.0 interface pre-wired but untested with physical hardware |
| 3 | 25-arm bandit identifies which region fails, not the exact threshold | Bayesian Optimisation within each arm is the planned next step |
| 4 | Single ego + one target on a straight road, no traffic | Cut-ins, curves, and traffic density not yet in catalog |
| 5 | No sensor model (camera / radar / LiDAR) | SUT receives ground-truth distance from SUMO, no noise or occlusion |
| 6 | SIL only, no HIL validation | Pipeline not tested against a real ECU on a test bench |
| 7 | NCAP scoring simplified | Partial-credit and full VRU scoring not completely implemented |
| 8 | Fixed seed=42 across all SUT versions | Valid for regression comparison; tails of parameter space undersampled |
Open an issue: describe your use case and what simulator / ECU interface you work with.
Responses typically within 48 hours.
Euro NCAP 2026 scenario definitions used under the public regulation specification. SUMO simulator used under EPL 2.0. Core ML pipeline and SUT implementation proprietary.