Human-machine shared driving for vehicle collision avoidance based on Hamilton-Jacobi reachability.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Theory

Background:

• Current shared control systems for collision avoidance often disrupt driver intent.
• Existing methods can increase the risk of human-machine conflict.
• There's a need for intelligent systems that intervene minimally and effectively.

Purpose of the Study:

• To develop a shared control framework that aids drivers in critical collision scenarios.
• To minimize human-machine conflicts during autonomous interventions.
• To ensure collision avoidance by intervening only when theoretically unavoidable.

Main Methods:

• Utilized Hamilton-Jacobi (HJ) reachability analysis to define a Collision Avoidance Reachable Set (CARS).
• Employed offline data to precompute reachability distributions and CARS.
• Developed a driver model and an authority allocation strategy to reduce conflicts.
• Trained a Reinforcement Learning (RL) agent to enforce CARS constraints and minimize conflicts.

Main Results:

• The Reachability-Aware RL framework effectively intervenes before the CARS boundary, preventing collisions.
• The system demonstrated improved performance in maintaining the original driving task.
• Robustness analysis confirmed flexibility across different driver attributes.

Conclusions:

• The proposed HJ reachability-guided RL framework offers effective and minimally intrusive shared collision avoidance.
• This method enhances safety by preventing entry into unavoidable collision states.
• The approach shows promise for real-world vehicle platforms and diverse driving behaviors.