[Paper Review] Integrated Decision and Control: Towards Interpretable and Computationally Efficient Driving Intelligence

Yang Guan, Yangang Ren+6 more|arXiv (Cornell University)|Mar 18, 2021

Autonomous Vehicle Technology and Safety34 references4 citations

TL;DR

This paper proposes an interpretable and computationally efficient framework called Integrated Decision and Control (IDC) for automated vehicles, which hierarchically decomposes driving into static path planning and dynamic optimal tracking. By formulating path tracking as constrained optimal control problems (OCPs) and solving them offline via a model-based reinforcement learning algorithm to learn value and policy networks, IDC achieves real-time online path selection and control with orders-of-magnitude higher efficiency and superior safety and traffic efficiency compared to baseline methods.

ABSTRACT

Decision and control are core functionalities of high-level automated vehicles. Current mainstream methods, such as functionality decomposition and end-to-end reinforcement learning (RL), either suffer high time complexity or poor interpretability and adaptability on real-world autonomous driving tasks. In this paper, we present an interpretable and computationally efficient framework called integrated decision and control (IDC) for automated vehicles, which decomposes the driving task into static path planning and dynamic optimal tracking that are structured hierarchically. First, the static path planning generates several candidate paths only considering static traffic elements. Then, the dynamic optimal tracking is designed to track the optimal path while considering the dynamic obstacles. To that end, we formulate a constrained optimal control problem (OCP) for each candidate path, optimize them separately and follow the one with the best tracking performance. To unload the heavy online computation, we propose a model-based reinforcement learning (RL) algorithm that can be served as an approximate constrained OCP solver. Specifically, the OCPs for all paths are considered together to construct a single complete RL problem and then solved offline in the form of value and policy networks, for real-time online path selecting and tracking respectively. We verify our framework in both simulations and the real world. Results show that compared with baseline methods IDC has an order of magnitude higher online computing efficiency, as well as better driving performance including traffic efficiency and safety. In addition, it yields great interpretability and adaptability among different driving tasks. The effectiveness of the proposed method is also demonstrated in real road tests with complicated traffic conditions.

Motivation & Objective

• To address the limitations of existing decision and control methods in automated vehicles, such as high computational cost and poor interpretability.
• To develop a unified framework that balances interpretability, computational efficiency, and adaptability across diverse driving scenarios.
• To reduce online computation burden while maintaining high performance in safety and traffic efficiency.
• To enable robust real-time operation in complex real-world driving environments, including human disturbances and sensor noise.
• To provide a structured, modular system that supports generalization across different driving tasks and scenarios.

Proposed method

• The driving task is hierarchically decomposed into static path planning (considering only static obstacles) and dynamic optimal tracking (considering dynamic obstacles and vehicle dynamics).
• For each candidate path, a constrained optimal control problem (OCP) is formulated to ensure safe and smooth tracking under dynamic constraints.
• A model-based reinforcement learning (RL) algorithm is proposed to solve all OCPs jointly offline, learning approximate value and policy networks as surrogate solutions.
• The learned value and policy networks are used online for real-time path selection and control, avoiding expensive online optimization.
• The framework leverages offline training to precompute solutions, drastically reducing online inference time while preserving performance.
• The method ensures interpretability by directly mapping learned value and policy functions to decision-making and control logic.

Experimental results

Research questions

• RQ1 Can a unified framework be designed to simultaneously achieve high computational efficiency, interpretability, and robustness in automated driving decision and control?
• RQ2 How can constrained optimal control problems for multiple candidate paths be efficiently solved offline to enable real-time online path selection and tracking?
• RQ3 To what extent does the proposed model-based RL approach maintain performance under sensor noise and human disturbances compared to traditional methods?
• RQ4 Can the framework generalize across diverse driving tasks and scenarios while preserving safety and traffic efficiency?
• RQ5 How does the interpretability of the learned policy and value functions compare to end-to-end deep RL approaches?

Key findings

• The IDC framework achieves an order of magnitude higher online computing efficiency compared to traditional rule-based methods, significantly reducing real-time computational load.
• The method demonstrates superior driving performance in terms of traffic efficiency and safety, as validated in both simulation and real-world intersection tests.
• The framework is robust to sensor noise up to moderate levels (noise levels 0–3), with only minor increases in state variance and bounds, maintaining stable vehicle dynamics.
• The system effectively recovers from abrupt human disturbances on the steering wheel, quickly correcting vehicle state after takeover and resuming the intended maneuver.
• The learned value and policy networks provide inherent interpretability, as they directly correspond to path selection and control actions, unlike black-box end-to-end models.
• The framework shows strong adaptability across different driving tasks, including complex left-turn maneuvers in real-world intersections with dynamic traffic.