Barrier Lyapunov Function-Based Safe Reinforcement Learning for Autonomous Vehicles With Optimized Backstepping.

Area of Science:

• Robotics
• Control Theory
• Artificial Intelligence

Background:

• Autonomous vehicle deployment faces challenges in guaranteeing safety and performance, especially during the learning phase.
• Safety-critical systems require robust performance even when utilizing reinforcement learning (RL).

Purpose of the Study:

• To propose a novel algorithm for safe reinforcement learning in autonomous vehicles.
• To ensure system safety and performance during the RL training period.

Main Methods:

• A Barrier Lyapunov Function-based safe RL (BLF-SRL) algorithm is developed for nonlinear systems in strict-feedback form.
• BLF terms are integrated into an optimized backstepping control method to maintain state variables within a safety region.
• An actor-critic framework with iterative updates ensures Bellman optimality and learns optimal control strategies.

Main Results:

• The BLF-SRL algorithm enables safe exploration during the learning process by constraining system states.
• The method optimizes overall system control and reduces performance variance under uncertainty.
• Simulations demonstrate the effectiveness of the proposed BLF-SRL method in autonomous vehicle motion control.

Conclusions:

• The proposed BLF-SRL algorithm effectively addresses safety challenges in autonomous vehicle learning.
• This approach guarantees safe operation and enhances control performance even in uncertain environments.