Safe Autonomous Driving with Latent Dynamics and State-Wise Constraints

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous driving systems face challenges in safety and reliability.
• Reinforcement learning (RL) shows promise for autonomous driving but struggles with sample efficiency and safety constraints.

Purpose of the Study:

• To develop a novel framework for safe reinforcement learning in autonomous driving.
• To address limitations of existing RL methods, specifically low sample efficiency and lack of explicit safety constraints.

Main Methods:

• Incorporated a latent dynamic model using bird's-eye view images for efficient learning and synthetic data generation.
• Introduced state-wise safety constraints via a barrier function to ensure safety at each state.
• Utilized the CARLA simulator for experimental validation.

Main Results:

• The proposed framework significantly improved driving performance compared to baseline methods.
• The framework demonstrated enhanced safety, reducing the risk of violations.
• Achieved efficient learning through the latent dynamic model and synthetic data generation.

Conclusions:

• The developed framework advances safe and efficient autonomous driving systems.
• Combining reinforcement learning with explicit safety considerations is effective.
• The approach shows potential for real-world autonomous driving applications.