Daniel Omeiza,Pratik Somaiya,Jo-Ann Pattinson,Carolyn Ten-Holter,Jack Stilgoe,Marina Jirotka,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. The code for this framework is available here: <a class="link-external link-https" href="https://github.com/cognitive-robots/rai-leaderboard" rel="external noopener nofollow">this https URL</a>
Robotics,Artificial Intelligence,Computer Vision and Pattern Recognition,Computers and Society,Machine Learning
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the robustness evaluation of autonomous driving systems in the face of various environmental interferences, sensor failures, and extreme situations. Specifically:
1. **Robustness under environmental interference**: The paper focuses on how to evaluate the performance of autonomous driving systems in different weather conditions (such as rain, snow, fog, etc.) and sensor occlusions (such as the camera being covered with dirt). These environmental factors may cause distortion or loss of sensor data, thus affecting the decision - making ability of the system.
2. **Robustness under sensor failure**: The paper explores how to evaluate the performance of autonomous driving systems when sensors fail, for example, when the camera has noise or some channels of the LiDAR (Light Detection and Ranging) become ineffective. These problems may cause the system to be unable to accurately perceive the surrounding environment, thereby affecting driving safety.
3. **Robustness in extreme situations**: The paper also studies the performance of autonomous driving systems when dealing with extreme situations (such as pedestrians illegally crossing the road, blurred traffic signs, etc.). Although these situations are not common, they may pose challenges to the safety and reliability of the system.
4. **Environmental sustainability**: In addition to robustness evaluation, the paper also focuses on the carbon emissions generated during the operation of autonomous driving systems. With the development of autonomous driving technology, the consumption of computing resources is continuously increasing, leading to an increase in carbon emissions. Therefore, evaluating the environmental impact of autonomous driving systems is also an important aspect of this research.
To address the above problems, the paper proposes a simulation - based robustness evaluation framework (S - RAF), which uses the CARLA driving simulator to conduct strict tests and evaluations on autonomous driving systems. S - RAF helps developers and stakeholders build safer and more responsible autonomous driving systems and simplifies the safety certification process by quantifying robustness and its relationship with other key safety factors (such as carbon emissions).
### Formula Examples
1. **Image occlusion**:
\[
I_{\text{occluded}}(x, y) = I_{\text{original}}(x, y)\cdot(1 - M(x, y))
\]
where \(I_{\text{original}}(x, y)\) is the original image and \(M(x, y)\) is the occlusion mask, representing the occluded area in the image.
2. **LiDAR occlusion**:
\[
S_{\text{occluded}}(x, y, z)=S_{\text{original}}(x, y, z)\cdot(1 - M(x, y, z))
\]
where \(S_{\text{original}}(x, y, z)\) is the original LiDAR data and \(M(x, y, z)\) is the occlusion mask, representing the occluded part of the data.
3. **Camera noise**:
\[
I_{\text{noisy}}(x, y)=\begin{cases}
0 & \text{with probability }\frac{p}{2}\\
255 & \text{with probability }\frac{p}{2}\\
I_{\text{original}}(x, y) & \text{otherwise}
\end{cases}
\]
where \(p\) is the probability that any pixel value is 0 or 255.
4. **Carbon emission estimation**:
\[
\text{CO}_2 \text{ emissions (Kg CO2Eq.)}=\text{CI}\times E
\]
where CI is the carbon intensity (unit: Kg CO2Eq./kWh) and E is the energy consumed during the model operation (unit: kWh).
Through these methods, S - RAF can comprehensively evaluate the robustness and environmental impact of autonomous driving systems, providing support for the development of safer and more environmentally - friendly autonomous driving technologies.