Fear-Neuro-Inspired Reinforcement Learning for Safe Autonomous Driving

Area of Science:

• Neuroscience and Artificial Intelligence
• Autonomous Driving Systems
• Machine Learning Safety

Background:

• Autonomous vehicles face challenges in safety and human-level performance, particularly in critical situations.
• Reinforcement learning (RL) shows promise for complex tasks but lacks safety guarantees for real-world autonomous driving.
• The brain's amygdala plays a vital role in threat detection and defensive responses, crucial for navigating risky environments.

Purpose of the Study:

• To develop a safe reinforcement learning framework for autonomous driving by modeling amygdala functionality.
• To enable autonomous agents to learn defensive behaviors and improve decision-making in safety-critical scenarios.
• To address the limitations of current RL techniques in ensuring safety for autonomous vehicles.

Main Methods:

• A novel fear-neuro-inspired reinforcement learning framework was developed.
• The framework models the functional aspects of the amygdala to elicit defensive responses.
• Experimental tests were conducted on autonomous driving agents using the proposed framework.

Main Results:

• The proposed framework facilitated agents in learning defensive behaviors, leading to fewer safety violations.
• Autonomous driving agents achieved state-of-the-art performance compared to baseline agents.
• The system's performance was comparable to that of certified human drivers in safety-critical scenarios.

Conclusions:

• The fear-neuro-inspired reinforcement learning framework effectively enhances safety in autonomous driving.
• Simulating amygdala function is crucial for advancing RL in safety-critical applications like autonomous vehicles.
• The approach demonstrates a viable pathway toward safer and more reliable autonomous driving systems.