[Paper Review] A Learning-Based Tune-Free Control Framework for Large Scale Autonomous Driving System Deployment

Yu Wang, Shu Jiang+6 more|arXiv (Cornell University)|Nov 9, 2020

Real-time simulation and control systems23 references4 citations

TL;DR

This paper proposes a learning-based, tune-free control framework that automates parameter tuning for autonomous driving systems across diverse vehicles and environments. By integrating machine learning for dynamic modeling, open-loop mapping, and Bayesian optimization for closed-loop control (e.g., PID, LQR), the framework achieves significant improvements in control performance and tuning efficiency, validated in both simulation and real-world tests across the US and China.

ABSTRACT

This paper presents the design of a tune-free (human-out-of-the-loop parameter tuning) control framework, aiming at accelerating large scale autonomous driving system deployed on various vehicles and driving environments. The framework consists of three machine-learning-based procedures, which jointly automate the control parameter tuning for autonomous driving, including: a learning-based dynamic modeling procedure, to enable the control-in-the-loop simulation with highly accurate vehicle dynamics for parameter tuning; a learning-based open-loop mapping procedure, to solve the feedforward control parameters tuning; and more significantly, a Bayesian-optimization-based closed-loop parameter tuning procedure, to automatically tune feedback control (PID, LQR, MRAC, MPC, etc.) parameters in simulation environment. The paper shows an improvement in control performance with a significant increase in parameter tuning efficiency, in both simulation and road tests. This framework has been validated on different vehicles in US and China.

Motivation & Objective

• To eliminate manual, human-in-the-loop parameter tuning for autonomous driving control systems.
• To enable rapid and consistent deployment of autonomous driving systems across diverse vehicles and driving environments.
• To improve control performance and tuning efficiency through automated, data-driven parameter optimization.
• To validate the framework on real-world vehicles in both the US and China.

Proposed method

• A learning-based dynamic modeling procedure generates accurate vehicle dynamics for control-in-the-loop simulation.
• A learning-based open-loop mapping procedure automates feedforward control parameter tuning.
• A Bayesian optimization-based closed-loop procedure automatically tunes feedback controllers (e.g., PID, LQR, MPC) in simulation.
• The framework integrates all three procedures into a unified, end-to-end tune-free tuning pipeline.
• The entire tuning process is executed in simulation before real-world deployment, reducing reliance on physical testing.
• The approach leverages machine learning to generalize across different vehicle platforms and environmental conditions.

Experimental results

Research questions

• RQ1 Can a fully automated, human-out-of-the-loop framework effectively tune control parameters for autonomous driving systems?
• RQ2 How does the integration of learning-based dynamic modeling and Bayesian optimization improve tuning efficiency and control performance?
• RQ3 To what extent can this framework be generalized across different vehicles and driving environments?
• RQ4 What is the performance gain in simulation and real-world testing compared to traditional tuning methods?

Key findings

• The framework significantly improves control performance in both simulation and real-world road tests.
• Parameter tuning efficiency is substantially increased, reducing the need for manual tuning iterations.
• The framework has been successfully validated on multiple vehicle platforms in both the US and China.
• The use of Bayesian optimization enables effective and efficient closed-loop parameter tuning without human intervention.
• The learning-based dynamic modeling procedure enables high-fidelity simulation for accurate control tuning.
• The end-to-end automation of tuning across open-loop and closed-loop parameters enables scalable deployment.