AVFI: Fault Injection for Autonomous Vehicles

Saurabh Jha, Subho S. Banerjee, Cyriac James, Zbigniew Kalbarczyk, Ravishankar K. Iyer

Abstract

Autonomous vehicle (AV) technology is rapidly becoming a reality on U.S. roads, offering the promise of improvements in traffic management, safety, and the comfort and efficiency of vehicular travel. With this increasing popularity and ubiquitous deployment, resilience has become a critical requirement for public acceptance and adoption. Recent studies into the resilience of AVs have shown that though the AV systems are improving over time, they have not reached human levels of automation. Prior work in this area has studied the safety and resilience of individual components of the AV system (e.g., testing of neural networks powering the perception function). However, methods for holistic end-to-end resilience assessment of AV systems are still non-existent.

Introduction: The Resilience Gap in Autonomous Driving

As autonomous vehicles (AVs) transition from research prototypes to everyday road users, the focus of safety assessment is shifting. While traditional metrics often prioritize the "accuracy" of individual AI components—such as the precision of a pedestrian detection model—real-world safety depends on resilience. Resilience is the capacity of the entire end-to-end system to maintain safe operation despite internal component failures or external environmental stressors.

Current testing methodologies often evaluate AV components in isolation. However, an AV is a complex "system of systems" where a minor hardware glitch or a slight communication delay, when coupled with a degraded sensor feed, can lead to catastrophic outcomes. To address this, researchers have developed AVFI (Autonomous Vehicle Fault Injector), a framework designed for holistic, end-to-end resilience assessment. AVFI moves beyond component-level testing to evaluate how internal system "hiccups" manifest as actual driving failures in a simulated environment.

Methodology: The AVFI Framework

AVFI provides a closed-loop testing environment by integrating with the CARLA simulator, an open-source platform built on Unreal Engine. The framework allows researchers to stress-test an Autonomous Driving Agent (ADA) by injecting faults across the entire compute-to-actuation pipeline.

Overview of the AVFI approach, showing the integration of fault injection into the sensor-compute-actuation pipeline within the CARLA simulator.

The core of AVFI lies in its ability to simulate four distinct classes of faults:

  1. Data Faults: These simulate environmental noise and sensor degradations, such as water droplets on a camera lens, occlusions (objects blocking the view), or specific camera noise models.
  2. Hardware Faults: AVFI can inject single-bit or multi-bit flips into the processing units (CPUs/GPUs) to simulate transient hardware errors or physical degradation.
  3. Timing Faults: These introduce latencies, data loss, or out-of-order delivery in the communication paths between the AI controller and the vehicle's physical actuators.
  4. Machine Learning Faults: Noise is added directly to the neural network's weights to simulate corruption during training or runtime.

By monitoring domain-specific metrics like Mission Success Rate (MSR) and Traffic Violations Per Kilometer (VPK), AVFI quantifies the impact of these faults on actual driving performance rather than just model accuracy.

Experimental Setup and Scenarios

The study utilized an imitation learning-based convolutional neural network (IL-CNN) as the vehicle's "brain." This model was trained to map raw camera inputs directly to control actions (steering, throttle, and braking) based on human driving data. Testing was conducted within CARLA's diverse urban environments, featuring various weather conditions and dynamic actors like pedestrians and other vehicles. This setup allowed for the generation of complex, "on-the-fly" scenarios that challenge the system's resilience more effectively than static benchmark datasets.

Key Insights: The Impact of Latency and Sensor Noise

The experimental results highlight that AV safety is a property of the entire system, not just the AI model. Two major findings emerged:

The Latency Cliff

One of the most significant discoveries was the extreme sensitivity of AVs to output latency. Even if the perception and planning algorithms are functioning correctly, a delay in transmitting those decisions to the wheels can be devastating.

Impact of output delay on safety, showing a sharp increase in traffic violations per kilometer as the delay increases to 30 frames.

As shown in the data, when a delay of 30 frames (equivalent to a 2-second lag at 15fps) was injected, the number of traffic violations skyrocketed from near zero to over 40 violations per kilometer. This "latency cliff" demonstrates that timing is a critical, often overlooked, safety factor.

Sensor Sensitivity and Variance

Injecting faults like "WaterDrop" or "SolidOcc" (occlusions) significantly increased the variance in mission success. In several test cases, the mission success rate dropped from near 100% to below 75%. This suggests that while the system might be robust to some noise, specific environmental conditions can cause unpredictable and dangerous behavior.

Conclusion and Future Directions

AVFI represents a paradigm shift in AV safety testing, emphasizing that resilience must be assessed holistically. The study demonstrates that "Traffic Violations per KM" is a more meaningful metric for resilience than traditional ML accuracy, as it directly captures the consequences of system-wide failures.

Future work will focus on expanding AVFI to evaluate more complex, modular AV architectures that separate perception, localization, and planning. Additionally, researchers aim to use adversarial techniques to automatically discover the specific combinations of faults most likely to cause critical failures, further hardening AV systems for the real world.

Summary updated

Abstract

Autonomous vehicle (AV) technology is rapidly becoming a reality on U.S. roads, offering the promise of improvements in traffic management, safety, and the comfort and efficiency of vehicular travel. With this increasing popularity and ubiquitous deployment, resilience has become a critical requirement for public acceptance and adoption. Recent studies into the resilience of AVs have shown that though the AV systems are improving over time, they have not reached human levels of automation. Prior work in this area has studied the safety and resilience of individual components of the AV system (e.g., testing of neural networks powering the perception function). However, methods for holistic end-to-end resilience assessment of AV systems are still non-existent.
BibTeX
@inproceedings{jha2018avfi,
  author = {Saurabh Jha and Subho S. Banerjee and Cyriac James and Zbigniew Kalbarczyk and Ravishankar K. Iyer},
  title = {{AVFI: Fault Injection for Autonomous Vehicles}},
  booktitle = {DSN Workshops},
  year = {2018},
  url = {https://openreview.net/forum?id=IFaVXeHMq99},
}

Loading related content...

Figure