Tuning the feedback controller gains is a simple way to improve autonomous driving performance (2402.05064v1)

Abstract: Typical autonomous driving systems are a combination of machine learning algorithms (often involving neural networks) and classical feedback controllers. Whilst significant progress has been made in recent years on the neural network side of these systems, only limited progress has been made on the feedback controller side. Often, the feedback control gains are simply passed from paper to paper with little re-tuning taking place, even though the changes to the neural networks can alter the vehicle's closed loop dynamics. The aim of this paper is to highlight the limitations of this approach; it is shown that re-tuning the feedback controller can be a simple way to improve autonomous driving performance. To demonstrate this, the PID gains of the longitudinal controller in the TCP autonomous vehicle algorithm are tuned. This causes the driving score in CARLA to increase from 73.21 to 77.38, with the results averaged over 16 driving scenarios. Moreover, it was observed that the performance benefits were most apparent during challenging driving scenarios, such as during rain or night time, as the tuned controller led to a more assertive driving style. These results demonstrate the value of developing both the neural network and feedback control policies of autonomous driving systems simultaneously, as this can be a simple and methodical way to improve autonomous driving system performance and robustness.

• The paper demonstrates that tuning longitudinal PID controller gains in ADS significantly improves driving scores from 73.21 to 77.38 across 16 scenarios.
• It employs a heuristic method by increasing proportional gain and reducing integral gain, validated through extensive CARLA simulations under diverse environmental conditions.
• The findings advocate integrating classical control tuning with neural network advancements to enhance vehicle stability and responsiveness in autonomous driving.

Tuning Feedback Controller Gains for Enhanced Autonomous Driving

Introduction

This paper examines the impact of tuning feedback controller gains within autonomous driving systems (ADS) that integrate neural networks (NN) for trajectory generation. While significant advancements have been achieved in the field of NN-based control algorithms, especially in machine learning and data-driven methodologies, the paper posits that attention must also be given to optimizing the classical feedback controllers that play a crucial role in ensuring vehicle stability and performance. The paper reveals that such tuning can greatly influence the driving score within virtual environments, a finding substantiated through experiments conducted using the CARLA simulator.

Theoretical Foundations and Methods

The research emphasizes that, despite the burgeoning capabilities of NN in controlling ADS, optimization of associated feedback controllers has been neglected. This oversight is problematic as changes in NN often influence the vehicle’s closed-loop dynamics, thereby warranting complementary adjustments in the feedback mechanisms. The paper focuses on tuning the longitudinal PID gains in the Trajectory-guided Control Prediction (TCP) algorithm. Traditionally, feedback gains are inherited from previous research without recalibration, an approach critiqued as insufficient given the evolving dynamics introduced by NNs.

Experimental Setup

Experiments were conducted using CARLA, an open-source driving simulator renowned for its rich environment and realistic portrayal of complex driving situations. The paper leveraged version 0.9.10 of CARLA on Linux to provide a diverse array of testing conditions, encompassing various weather scenarios and urban environments. This enabled rigorous evaluation of tuning paradigms under conditions that approximate real-world driving challenges.

Results

The data indicate that tuning the longitudinal PID controller gains significantly enhances vehicular performance, as evidenced by an improved driving score from 73.21 to 77.38, averaged across 16 simulated driving scenarios. These gains were particularly pronounced under adverse conditions—such as rain and nighttime driving—suggesting that a more assertive driving style facilitated by PID adjustments can better counteract environmental perturbations. The paper's finding underscores the efficacy of a more dynamic control approach in enhancing stability and trajectory fidelity.

Implementational Insights

The tuning process employed a heuristic method to adjust PID controls, specifically by increasing the proportional gain while reducing the integral gain. This configuration fostered a rapid and adaptive response to scenario variability, thereby addressing issues of delayed or incorrect trajectory adherence encountered with the original TCP gain settings. While primitive in execution, these manual adjustments provided substantial performance improvements, indicating potential beyond simple tuning when augmented with advanced methodologies like automatic tuning or Model Predictive Control (MPC) strategies.

Implications for Future Research

Key takeaways include the potential for integrating classical control theory methods with advanced NN algorithms to foster more robust and explainable ADS solutions. This synergy could yield ADS configurations with responsive feedback systems that are both theoretically sound and practically effective under diverse operating conditions. Moreover, the findings advocate a multi-disciplinary collaboration in ADS development—uniting the predictive prowess of NNs with the stability assurances endemic to control theory—thereby advancing both safety and operational excellence.

Conclusion

In sum, the paper illuminates the criticality of re-tuning feedback controllers in tandem with NN developments to improve ADS performance. The research invites further exploration into sophisticated control strategies capable of scaling these preliminary gains to broader, more intricate ADS deployments. The work represents a call to harness the potential of integrated systems approaches, aligning the rapid advancements in machine learning with the depth of control theory to unlock new avenues for ADS enhancement.