Ethics-Aware Safe Reinforcement Learning for Rare-Event Risk Control in Interactive Urban Driving

Area of Science:

• Autonomous Systems
• Robotics
• Artificial Intelligence

Background:

• Autonomous vehicles (AVs) promise fewer traffic fatalities and improved efficiency.
• Ethical reasoning is crucial for AV adoption, especially for protecting vulnerable road users (VRUs).
• Current AV systems often lack robust ethical frameworks for complex scenarios.

Purpose of the Study:

• To develop and evaluate a hierarchical Safe Reinforcement Learning (Safe RL) framework, named EthicAR, for autonomous driving.
• To integrate ethics-aware cost signals into AV decision-making processes.
• To enhance the safety and ethical accountability of AVs in human-mixed traffic.

Main Methods:

• A hierarchical Safe RL framework augmenting standard driving objectives with ethics-aware cost signals.
• A composite ethical risk cost combining collision probability and harm severity for decision-level training.
• Risk-sensitive prioritized experience replay and Temporal Cost Aggregation (TCA) for improved sample efficiency and tail-risk mitigation.
• Polynomial trajectory generation with PID and Stanley controllers for execution-level control.

Main Results:

• EthicAR decreased collision rates by 20-45% compared to baseline methods in simulations.
• Task success rates and comfort metrics were maintained within 5-10% of baselines.
• The framework demonstrated effective protection for vulnerable road users in diverse scenarios.

Conclusions:

• The proposed EthicAR framework offers a viable approach to embedding ethical reasoning in autonomous driving.
• Combining formal control theory and data-driven learning advances ethically accountable autonomy.
• This work provides a benchmark for Safe RL with ethics-aware objectives in real-world traffic simulations.