TLDR: ARGUS is a new runtime framework that significantly improves the safety and resilience of end-to-end autonomous driving systems (ADSs). It continuously monitors for driving hazards like collisions, stop signal violations, and stalling, and proactively takes over control to mitigate risks and prevent accidents, returning control to the ADS once safe. This framework has been shown to improve driving scores by over 150% and prevent nearly two-thirds of violations with minimal performance overhead, demonstrating its effectiveness and efficiency.
End-to-end autonomous driving systems (ADSs) are rapidly advancing, promising safer roads and more efficient transportation. These systems, which handle everything from perceiving the environment to making driving decisions, are attracting significant attention from researchers and industry alike. However, deploying ADSs on public roads introduces a critical challenge: ensuring their resilience against diverse driving hazards that can compromise safety and degrade performance.
Current autonomous driving solutions often fall short in continuously monitoring for potential dangers and adaptively responding to safety violations. Many existing approaches focus on emergency interventions rather than proactive hazard mitigation and quick recovery to maintain robust driving behaviors in complex scenarios. This gap highlights a strong demand for systems that can not only detect but also actively prevent incidents.
Introducing ARGUS: A Proactive Safety Framework
To address this crucial need, researchers have developed a new resilience-oriented runtime framework called ARGUS. This innovative system is designed to mitigate driving hazards, prevent potential safety violations, and ultimately improve the driving performance of end-to-end ADSs. ARGUS works by continuously monitoring the trajectories generated by an ADS. If the vehicle controlled by the ADS (referred to as the EGO vehicle) is deemed unsafe, ARGUS seamlessly takes control through a specialized hazard mitigator. Control is then returned to the ADS only after safety has been reestablished.
The ARGUS framework is composed of three main components: the Takeover Gate, the Hazard Monitor, and the Hazard Mitigator. Each plays a vital role in ensuring the system’s resilience:
-
Takeover Gate: This component acts as the decision-maker, dynamically switching control between the ADS and the Hazard Mitigator. It checks various buffers maintained by the Hazard Monitor to determine if a generated trajectory is safe for execution.
-
Hazard Monitor: Operating asynchronously, the Hazard Monitor is responsible for detecting potential driving hazards. It uses information from the vehicle’s sensors to predict the future movements of the EGO vehicle and surrounding traffic participants (vehicles, pedestrians, static obstacles, and stop signals). Based on these predictions, it evaluates risks such as collisions, stop signal violations, and stalling hazards, updating the takeover and recovery buffers for the Takeover Gate.
-
Hazard Mitigator: When a hazard is detected and control is taken over, the Hazard Mitigator steps in. It generates alternative waypoints to guide the EGO vehicle around obstacles and dynamically adjusts the vehicle’s speed using an Intelligent Driver Model (IDM). This ensures safe navigation by proactively avoiding collisions and adhering to traffic rules, such as stopping at red lights or stop signs.
Also Read:
- Ensuring Data Integrity for Safe Autonomous Driving Systems
- Ensuring Trust in Autonomous AI: A Two-Layered Monitoring Approach for Agentic Systems
Demonstrated Effectiveness and Efficiency
The effectiveness and efficiency of ARGUS have been rigorously evaluated through extensive experiments. The framework was integrated with three state-of-the-art end-to-end ADSs: TCP, UniAD, and VAD. These enhanced systems were tested on widely used benchmarks, Bench2Drive and CARLA leaderboard 2.0, under various driving hazards and conditions.
The results were highly promising: ARGUS significantly enhanced the resilience of ADSs, improving their average driving score by an impressive 150.30%. Furthermore, it prevented 64.38% of violations with minimal additional time overhead, demonstrating its practicality for real-time application. The system also proved highly accurate in making takeover decisions, achieving a high F3 score of 0.899. Even when generalized to modular ADSs like Apollo, ARGUS showed substantial improvements, outperforming existing safety assurance approaches.
This research highlights ARGUS as a crucial step forward in developing safer and more robust autonomous driving systems. By providing a modular, resilience-oriented safety assurance framework, ARGUS helps AI-enabled complex systems navigate the unpredictable challenges of real-world driving. For more detailed information, you can refer to the full research paper: ARGUS: Resilience-Oriented Safety Assurance Framework for End-to-End ADSs.
