[Paper Review] Rationale-aware Autonomous Driving Policy utilizing Safety Force Field implemented on CARLA Simulator

Ho Suk, Taewoo Kim+4 more|arXiv (Cornell University)|Nov 18, 2022

Autonomous Vehicle Technology and Safety4 citations

TL;DR

This paper proposes a rationale-aware autonomous driving policy using a self-implemented Safety Force Field (SFF) model integrated with the CARLA simulator to enable human-like, consistent driving under varying traffic conditions. By training a deep neural network to predict claimed sets of vehicles and computing safety potential, the method achieves superior safety and efficiency compared to CARLA autopilot and RSS-based systems, with 1.40 arrivals per second and 2,495 accident-free timesteps under high aggression conditions.

ABSTRACT

Despite the rapid improvement of autonomous driving technology in recent years, automotive manufacturers must resolve liability issues to commercialize autonomous passenger car of SAE J3016 Level 3 or higher. To cope with the product liability law, manufacturers develop autonomous driving systems in compliance with international standards for safety such as ISO 26262 and ISO 21448. Concerning the safety of the intended functionality (SOTIF) requirement in ISO 26262, the driving policy recommends providing an explicit rational basis for maneuver decisions. In this case, mathematical models such as Safety Force Field (SFF) and Responsibility-Sensitive Safety (RSS) which have interpretability on decision, may be suitable. In this work, we implement SFF from scratch to substitute the undisclosed NVIDIA's source code and integrate it with CARLA open-source simulator. Using SFF and CARLA, we present a predictor for claimed sets of vehicles, and based on the predictor, propose an integrated driving policy that consistently operates regardless of safety conditions it encounters while passing through dynamic traffic. The policy does not have a separate plan for each condition, but using safety potential, it aims human-like driving blended in with traffic flow.

Motivation & Objective

• To address the liability and safety compliance challenges in SAE Level 3+ autonomous vehicles by implementing a mathematically rigorous, interpretable safety model.
• To overcome the limitations of existing SFF implementations, which are closed-source and platform-specific, by creating a fully open-source SFF implementation compatible with CARLA.
• To develop an integrated driving policy that operates consistently across safe and unsafe traffic conditions without relying on a separate 'panic button' safety module.
• To ensure the policy maintains traffic flow smoothness while guaranteeing safety, aligning with ISO 21448 SOTIF requirements.
• To validate the proposed method through comparative experiments against CARLA autopilot and RSS-based systems under varying levels of traffic aggression.

Proposed method

• Implementing the Safety Force Field (SFF) model from scratch to ensure full transparency and compatibility with open-source frameworks like CARLA.
• Training a deep neural network to predict the 'claimed set' of each vehicle—defining the region of influence for safety calculations—based on vehicle state and trajectory.
• Calculating safety potential using the claimed sets to determine the ego-vehicle’s safety level in real time, enabling dynamic decision-making.
• Integrating the safety potential into a unified driving policy that replaces separate safe/unsafe mode logic, ensuring consistent behavior.
• Using the safety potential as a core component of the planning subsystem, enabling smooth, human-like maneuvers without abrupt interventions.
• Validating the system through controlled experiments in CARLA with four levels of random vehicle aggression, measuring arrival rate and accident-free timesteps.

Experimental results

Research questions

• RQ1 Can a self-implemented, open-source SFF model achieve comparable or better safety and efficiency than existing closed-source SFF and RSS-based systems?
• RQ2 Does integrating SFF into the core planning policy—rather than as a separate override—lead to smoother, more human-like driving behavior?
• RQ3 Can a learned claimed set predictor enable accurate and real-time safety potential computation without relying on proprietary implementations?
• RQ4 How does the proposed policy perform under increasing traffic aggression, particularly in terms of safety and efficiency?
• RQ5 To what extent does the integrated policy maintain traffic flow while ensuring compliance with SOTIF requirements under dynamic conditions?

Key findings

• The proposed SFF-based driving policy achieved 1.40 arrivals per second and 2,495 accident-free timesteps under high aggression conditions, outperforming CARLA autopilot (0.52 arrivals, 1,332 timesteps) and RSS-CARLA (1.00 arrivals, 1,886 timesteps).
• The method demonstrated consistent performance across all aggression levels, showing robustness without requiring separate safety logic for safe vs. unsafe conditions.
• The claimed set predictor successfully inferred vehicle influence regions, enabling accurate safety potential computation and real-time decision-making.
• The integrated policy reduced abrupt interventions compared to traditional SFF-overriding architectures, resulting in smoother traffic flow and human-like behavior.
• The SFF implementation was functionally equivalent to proprietary models, as validated by performance parity with RSS and superior results in key metrics.
• The approach satisfies ISO 21448 SOTIF requirements by providing a mathematically explainable, interpretable safety rationale for all maneuvers.