💻 computer science

Safety-Centered Scenario Generation for Autonomous Vehicles

This paper presents a safety-centered scenario generation framework that creates diverse, parametrized, and critical driving situations to systematically validate autonomous vehicle safety features through simulation, ensuring traceability to ISO 26262 requirements and providing quantitative metrics for accelerated, cost-effective regulatory compliance.

Original authors: Kiruthiga Chandra Shekar, Aliasghar Moj Arab

Published 2026-03-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kiruthiga Chandra Shekar, Aliasghar Moj Arab

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a brand new robot to drive a car. You can't just let it loose on a busy highway to "learn by doing"—that would be dangerous and chaotic. Instead, you need a super-powered driving school where you can safely simulate every possible nightmare scenario without anyone getting hurt.

This paper is about building that super-powered school. Here is the breakdown of how the authors did it, using some everyday analogies.

1. The Problem: The "Needle in a Haystack" Issue

Driving is full of surprises. A child might run into the street, a car might cut you off, or it might start raining heavily. These are called "edge cases"—rare, dangerous moments that happen very rarely in real life.

The Analogy: Imagine trying to learn how to swim by only practicing in a calm, empty pool. You'd be great at calm water, but if you got thrown into a stormy ocean with a shark, you'd panic.
The Issue: Real-world data (logs from actual cars) is like a library of calm days. It rarely has the "stormy ocean" data. If you only train your robot on calm days, it will fail when the storm hits.

2. The Solution: The "AI Scenario Chef"

The authors created a system that acts like a creative chef who doesn't just follow a recipe, but invents new, dangerous dishes to test the robot's stomach.

How it works:
1. The Menu (HARA): First, they use a safety checklist called HARA (Hazard Analysis and Risk Assessment). Think of this as a list of "What could go wrong?" (e.g., "What if a pedestrian jumps out?").
2. The Chef (AI): They feed this list to an Artificial Intelligence. The AI acts like a chef who takes the idea "Pedestrian jumps out" and instantly writes the code to create that exact scene in a video game.
3. The Kitchen (CARLA): The AI writes the code for a simulator called CARLA (a high-tech driving video game). This is the kitchen where the cooking happens.

3. The Process: The "Million-Dollar Test Drive"

Instead of testing the car once, the system runs over 1,000 different versions of the same scary situation.

The Analogy: Imagine you are testing a parachute. You wouldn't just jump once. You would jump 1,000 times, changing the wind speed, the height, the weight of the jumper, and the color of the parachute each time.
What they changed: In the simulation, they tweaked five things for every test:
- How fast the car was going.
- How fast the pedestrian was running.
- The angle they were approaching from.
- How far away they started.
- When they appeared.

They covered 100% of the possible combinations for these factors. It's like checking every single square on a chessboard to make sure the robot knows how to move in any position.

4. The Results: The "Report Card"

After running the simulations, the system gave the robot a report card. It didn't just say "Pass" or "Fail." It gave specific metrics:

Time-to-Collision (TTC): How many seconds did the robot have to react before hitting something?
Minimum Distance: How close did it get to the "crash" before stopping?
Braking: Did it slam on the brakes too hard, or was it smooth?

The Findings:

The robot was generally good at stopping.
However, it struggled more when things happened fast (high speed) or when the angles were tricky (like a cyclist coming from a weird angle).
The system found that in about 33% of the tests, the robot got into a "critical zone" where it had less than 1.2 seconds to react. This tells engineers exactly where to fix the software.

5. Why This Matters: The "Speed and Safety" Balance

The paper compares their new AI method to the old ways of testing:

Old Way (Manual): A human expert writes the code for a test. It takes 45 hours to make 1,000 tests. It's slow and misses many weird scenarios.
New Way (AI): The AI makes 1,000 tests in just 2.3 hours. It found 23 unique "edge cases" that the humans missed.

The Big Picture

Think of this framework as a time machine for safety. Instead of waiting for a real accident to happen to learn a lesson, this system creates thousands of "what-if" accidents in a computer.

It allows engineers to say, "We know this robot can handle a rainy day, a running child, and a sudden stop, because we've already tested it 1,000 times in a virtual world." This makes the road safer for everyone before the car ever leaves the factory.

1. Problem Statement

The deployment of Autonomous Vehicles (AVs) faces significant hurdles in safety validation due to the vast complexity of real-world driving environments. Key challenges include:

Scenario Complexity: AVs must navigate diverse weather, road geometries, and dynamic interactions with other road users.
Edge Case Scarcity: Rare, high-risk "edge cases" (e.g., sudden pedestrian crossings, aggressive cut-ins) occur infrequently in real-world data, making them difficult to capture for training and testing.
Testing Limitations: Physical testing is risky, expensive, and computationally inefficient. Existing simulation methods often suffer from limited scenario coverage, lack of realism, and a failure to systematically link test cases to formal safety goals (such as ISO 26262).
Manual Bottlenecks: Creating test scenarios manually is time-consuming, prone to human error, and difficult to scale.

The paper addresses the need for a scalable, automated, and safety-centric framework that generates diverse, parametrized, and critical driving scenarios to rigorously validate AV safety features.

2. Methodology

The authors propose a framework that integrates Hazard Analysis and Risk Assessment (HARA) with AI-driven code generation within the CARLA simulation environment.

A. Architecture Overview

The workflow follows a three-stage pipeline:

HARA Integration: Safety-critical scenarios are first defined using natural language descriptions derived from HARA. These descriptions identify hazards, safety goals, and risk parameters.
AI-Based Code Generation: A Large Language Model (LLM), fine-tuned for scenario generation, converts the natural language HARA descriptions into executable Python code compatible with CARLA.
- Input: Structured templates including Scenario Type, Actors (ego, pedestrians, etc.), Environmental Conditions, Trigger Events, and Parameter Ranges.
- Output: Python scripts that instantiate actors, script behaviors, configure sensors, and set up logging.
Simulation & Analysis: The generated code runs in CARLA (v0.9.10) in "headless mode" (no graphical rendering) to maximize computational efficiency. The system logs quantitative data for post-processing.

B. Scenario Parameterization & Coverage

To ensure rigorous testing, the framework defines a Scenario Coverage Index (SCI) based on a 5-dimensional parameter space:

Parameters: Ego velocity ( $v_v$ ), Pedestrian/Cyclist velocity ( $v_p$ ), Approach angle ( $\theta$ ), Initial distance ( $d$ ), and Time of appearance ( $t$ ).
Discretization: Each parameter is divided into discrete bins (e.g., 7 velocity bins, 9 distance bins).
Metric Formulation: The SCI calculates the ratio of tested bins to total defined bins, weighted by the parameter's importance to collision risk.
- Formula: $SCI = \sum w_j \cdot \frac{N_{tested}(p_j)}{N_{bins}(p_j)}$
- Goal: Achieve an SCI close to 1.0, indicating full exploration of the safety-critical parameter space.

C. Safety Metrics

The framework computes quantitative metrics to evaluate AV performance:

Time-to-Collision (TTC): Temporal proximity to a crash.
Minimum Distance (Dmin): Spatial proximity.
Post-Encroachment Time (PET): Time separation between the ego vehicle and an obstacle passing a conflict point.
Threshold Violations: Instances where TTC drops below critical safety limits (e.g., < 1.2s).
Collision Severity: Impact velocity and collision type.

3. Key Contributions

Automated Scenario Generation: A novel pipeline using LLMs to translate HARA-derived safety requirements directly into simulation code, reducing manual effort by orders of magnitude.
Traceability to Safety Standards: The framework explicitly maps generated scenarios to ISO 26262 functional safety requirements and HARA-derived safety goals, ensuring regulatory traceability.
Quantitative Coverage Metric: The introduction of the Scenario Coverage Index (SCI) provides a mathematical method to measure the completeness of safety testing across a defined parameter space.
Scalable Validation: By utilizing headless simulation and parallel processing, the system can execute thousands of scenarios efficiently, enabling statistical reliability in safety assessment.

4. Results

The framework was evaluated using the CARLA autopilot module across 1,024 executed scenarios (including Vehicle-to-Vehicle and Vulnerable Road User scenarios).

Coverage Performance:
- The framework achieved 100% parameter space coverage (SCI = 1.0) across all defined bins for velocity, angle, distance, and time.
- 87% of identified critical high-risk parameter combinations were successfully tested.
Safety Metrics:
- Collision Rate: 12.4% (127 collisions out of 1,024 runs).
- Mean TTC: 3.42 seconds.
- Threshold Violations: 33.4% of scenarios had TTC < 1.2s, indicating critical safety situations requiring evasive action.
- Risk Correlation: Higher ego speeds (>30 km/h) and acute crossing angles (30°–45°) correlated strongly with reduced safety margins and increased collision likelihood.
Baseline Comparison (AI vs. Manual vs. Rule-Based):
- Efficiency: The AI method generated 1,000 scenarios in 2.3 hours, compared to 45 hours (manual) and 8.5 hours (rule-based). This represents a 19.6x speedup over manual generation.
- Diversity: AI achieved SCI = 1.0, outperforming manual (0.68) and rule-based (0.82) methods.
- Edge Cases: The AI identified 23 unique edge cases, significantly higher than manual (8) or rule-based (12) methods.
- Code Quality: AI-generated code achieved 94.2% syntax correctness, comparable to human-reviewed code.

5. Significance

This paper presents a transformative approach to AV safety validation by bridging the gap between functional safety engineering (HARA) and automated simulation.

Accelerated Validation: The massive reduction in time required to generate and run diverse scenarios allows for faster iteration cycles in AV development.
Systematic Risk Assessment: By moving beyond random testing to a structured, parameter-driven approach, the framework ensures that safety claims are backed by comprehensive data covering the entire Operational Design Domain (ODD).
Regulatory Confidence: The ability to trace test cases back to specific safety goals and provide quantitative metrics (TTC, SCI) offers robust evidence for regulatory bodies and stakeholders.
Future-Proofing: The modular design supports future integration with Reinforcement Learning (RL) agents, Hardware-in-the-Loop (HIL) testing, and advanced AI safety analysis tools.

In conclusion, the proposed framework demonstrates that AI-driven scenario generation is not only feasible but superior to traditional methods in terms of speed, coverage, and the ability to uncover critical edge cases, thereby significantly enhancing the safety assurance of autonomous vehicles.