S-RAF: A Simulation-Based Robustness Assessment Framework for Responsible Autonomous Driving

Daniel Omeiza, Pratik Somaiya, Jo-Ann Pattison, Carolyn Ten-Holter, Marina Jirotka, Jack Stilgoe, Lars Kunze

Abstract

As artificial intelligence (AI) technology advances, ensuring the robustness and safety of AI-driven systems has become paramount. However, varying perceptions of robustness among AI developers create misaligned evaluation metrics, complicating the assessment and certification of safety-critical and complex AI systems such as autonomous driving (AD) agents. To address this challenge, we introduce Simulation-Based Robustness Assessment Framework (S-RAF) for autonomous driving. S-RAF leverages the CARLA Driving simulator to rigorously assess AD agents across diverse conditions, including faulty sensors, environmental changes, and complex traffic situations. By quantifying robustness and its relationship with other safety-critical factors, such as carbon emissions, S-RAF aids developers and stakeholders in building safe and responsible driving agents, and streamlining safety certification processes. Furthermore, S-RAF offers significant advantages, such as reduced testing costs, and the ability to explore edge cases that may be unsafe to test in the real world.

Introduction

The rapid advancement of artificial intelligence (AI) has propelled autonomous driving (AD) from a futuristic concept to a near-term reality. However, as these systems transition from controlled environments to the unpredictable real world, ensuring their safety and robustness is paramount. A significant challenge lies in the lack of standardized evaluation metrics; developers often use proprietary benchmarks, making it difficult for regulators to assess and certify safety-critical AI systems. Furthermore, the push for higher performance often leads to increased model complexity, which can conflict with sustainability goals. To address these issues, we propose the Simulation-Based Robustness Assessment Framework (S-RAF), a tool designed to evaluate AD agents through the lens of Responsible AI (RAI).

The S-RAF Framework

S-RAF provides a standardized, open-source environment for quantifying the trade-offs between an agent's driving performance, its robustness to unexpected conditions, and its environmental impact. The framework operates by selecting top-performing models from initial trials and subjecting them to rigorous assessment within the CARLA Driving Simulator.

Overview of the S-RAF framework, illustrating the pipeline from ML experiments to robustness and emission assessment via the CARLA simulator.

Methodology: Simulating Real-World Stress

The core of S-RAF is its ability to "stress-test" agents across three primary categories of disturbances:

  1. Environmental Disturbances: These include physical obstructions like dirt or snow on sensors and adverse weather conditions such as heavy rain or dense fog.
  2. Sensor Errors: The framework introduces technical failures, such as "salt and pepper" noise in camera feeds, LiDAR channel dropouts, and Gaussian noise in GNSS/IMU data.
  3. Corner Cases and Data Shift: Agents are tested against unpredictable human behavior (e.g., jaywalkers) and environments that differ significantly from their training data.
Visual representation of the various disturbances implemented in S-RAF, including occluded and faulty sensors, weather effects, and corner cases.

To measure these factors, S-RAF utilizes three key metrics:

  • Driving Score (DS): A composite of route completion and infraction penalties.
  • Robustness Driving Score (RDS): The mean DS across all disturbed conditions.
  • Carbon Emission Indicator: Calculated using the codecarbon library to track energy consumption during inference.

Experimental Evaluation

We benchmarked three state-of-the-art agents—LBC (2020), NEAT (2021), and InterFuser (2022)—alongside a classical NPC baseline. Our results indicate a clear hierarchy in performance but also reveal significant vulnerabilities. InterFuser emerged as the most capable agent with a DS of 54.36% and an RDS of 34.55%. However, this performance came at a cost: it emitted 0.0341 kg CO2 Eq/sec, far exceeding its predecessors. In contrast, the NPC baseline had negligible emissions (0.0005 kg CO2 Eq/sec) but lacked the perception necessary for complex navigation.

Discussion: The Robustness-Sustainability Trade-off

Our analysis uncovered a critical "scaling law" conflict: as models become more robust and safe, their computational requirements—and thus their carbon footprints—increase significantly. Furthermore, we found that camera-based perception is particularly fragile; noise or occlusion in camera feeds led to the most severe performance drops across all models. This finding raises important questions for "camera-only" autonomous systems. By testing agents on conditions they likely did not encounter during training, S-RAF provides a "veil of ignorance" assessment, offering a more realistic view of an agent's real-world readiness.

Conclusion and Future Work

S-RAF represents a step toward more responsible and transparent AI development in autonomous driving. While the framework currently focuses on inference-phase emissions and simulated environments, future work will expand to include full lifecycle carbon assessments (including training) and more complex traffic resilience scenarios. By providing a standardized way to measure safety and sustainability, S-RAF helps bridge the gap between AI development and regulatory certification.

Summary updated

Abstract

As artificial intelligence (AI) technology advances, ensuring the robustness and safety of AI-driven systems has become paramount. However, varying perceptions of robustness among AI developers create misaligned evaluation metrics, complicating the assessment and certification of safety-critical and complex AI systems such as autonomous driving (AD) agents. To address this challenge, we introduce Simulation-Based Robustness Assessment Framework (S-RAF) for autonomous driving. S-RAF leverages the CARLA Driving simulator to rigorously assess AD agents across diverse conditions, including faulty sensors, environmental changes, and complex traffic situations. By quantifying robustness and its relationship with other safety-critical factors, such as carbon emissions, S-RAF aids developers and stakeholders in building safe and responsible driving agents, and streamlining safety certification processes. Furthermore, S-RAF offers significant advantages, such as reduced testing costs, and the ability to explore edge cases that may be unsafe to test in the real world.
BibTeX
@inproceedings{omeiza2024s,
  author = {Daniel Omeiza and Pratik Somaiya and Jo-Ann Pattison and Carolyn Ten-Holter and Marina Jirotka and Jack Stilgoe and Lars Kunze},
  title = {{S-RAF: A Simulation-Based Robustness Assessment Framework for Responsible Autonomous Driving}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaaiss.v4i1.31776},
}
Figure