Integrated Decision and Control: Toward Interpretable and Computationally Efficient Driving Intelligence.

Area of Science:

• Robotics and Artificial Intelligence
• Autonomous Systems Engineering

Background:

• Current automated vehicle decision and control methods face challenges with high computational complexity, poor interpretability, and limited adaptability.
• Existing approaches like functional decomposition and end-to-end reinforcement learning (RL) are insufficient for real-world autonomous driving tasks.

Purpose of the Study:

• To present an interpretable and computationally efficient framework, Integrated Decision and Control (IDC), for automated vehicles.
• To address the limitations of current methods in terms of time complexity, interpretability, and adaptability.

Main Methods:

• The IDC framework decomposes the driving task into hierarchical static path planning and dynamic optimal tracking.
• Constrained Optimal Control Problems (OCP) are formulated and optimized for candidate paths, with the best performing one selected.
• A model-based RL algorithm is proposed for offline solving of OCPs, creating value and policy networks for real-time online decision-making.

Main Results:

• The IDC framework demonstrates an order of magnitude improvement in online computing efficiency compared to baseline methods.
• IDC achieves superior driving performance, enhancing both traffic efficiency and safety.
• The framework exhibits significant interpretability and adaptability across diverse driving scenarios and tasks.

Conclusions:

• The Integrated Decision and Control (IDC) framework offers a computationally efficient, interpretable, and adaptable solution for automated vehicle decision and control.
• IDC significantly improves driving performance, safety, and real-time processing capabilities, outperforming existing methods.