Resource-Efficient Path-Following Control for a Self-Driving Car in a Networked Control System

Guillermo Alite, Ángel Cuenca, Julián Salt, Masayoshi TomizukaIEEE Access 2023 · DOI

Abstract

Modern autonomous vehicles rely on complex Networked Control Systems (NCS) where sensors, controllers, and actuators communicate over shared digital buses. However, high-frequency data transmission can lead to network congestion and excessive energy consumption. This paper presents a resource-efficient control framework that significantly reduces in-vehicle network traffic while maintaining high-precision path following. The approach integrates a Non-Uniform Dual-Rate Extended Kalman Filter (NUDREKF) with Periodic Event-Triggered Communication (PETC) and packet-based predictive control. Simulation results demonstrate bandwidth savings of over 90% and robust performance under network delays and packet loss.

Introduction

As autonomous vehicles (AVs) evolve, the internal networks connecting their various components—sensors like LiDAR and GPS, central processing units, and mechanical actuators—are becoming significant bottlenecks. These vehicles are essentially Networked Control Systems (NCS) where resource scarcity and network non-determinism (delays and packet loss) pose critical challenges. Traditional control strategies often assume ideal communication, which can lead to dangerous oscillations or lane departures when real-world network "gremlins" occur. This research addresses these issues by proposing a strategy to slash network traffic and save battery life without compromising vehicle safety or path-following accuracy.

System Architecture

The proposed architecture divides the control loop into three primary domains: the controller side, the in-vehicle network, and the plant side (the vehicle itself).

Figure 1: Proposed resource-efficient control architecture for a self-driving car.

As shown in Figure 1, the system utilizes a feedback loop where sensor data is transmitted across the network to the controller. To minimize traffic, Periodic Event-Triggered Communication (PETC) is employed on both the sensor and controller sides. Data is only transmitted if the change in the vehicle's state exceeds a predefined threshold. To handle the resulting non-uniform data arrival, a specialized Kalman filter is used to maintain a continuous state estimate.

Vehicle Modeling and State Estimation

At the heart of the estimation and control logic is a kinematic bicycle model. This model provides a mathematically efficient representation of vehicle motion, making it suitable for low-power automotive hardware.

Figure 2: The kinematic bicycle model used for state estimation and path-following control.

The model, illustrated in Figure 2, tracks the vehicle's position $(x, y)$, yaw angle $(\psi)$, and velocity $(v)$. The distances from the center of gravity to the front and rear axles are denoted by $l_f$ and $l_r$, respectively.

To bridge the gap between slow, irregular sensor updates and the need for high-frequency control commands, the researchers developed the Non-Uniform Dual-Rate Extended Kalman Filter (NUDREKF). This filter allows the controller to operate at a fast, fixed rate (e.g., 10ms) by predicting the vehicle's state between sensor updates based on the bicycle model. This "gap-filling" capability is essential for maintaining stability when the network is silent.

Resource-Efficient Control Strategy

The framework employs two additional techniques to ensure efficiency and resilience:

  1. Periodic Event-Triggered Communication (PETC): Instead of continuous transmission, the system evaluates the necessity of each packet. If the vehicle's state or the intended control command hasn't changed significantly since the last transmission, the network remains silent, preserving bandwidth and reducing energy consumption.
  2. Packet-Based Control with $h$-step Prediction: To combat network delays and packet loss, the controller transmits a "packet" containing a sequence of predicted future control actions rather than a single command. If a subsequent packet is dropped or delayed, the actuator can continue to execute the pre-calculated sequence from its local buffer, ensuring uninterrupted operation.

Experimental Results and Discussion

The system was validated using a high-fidelity simulation of a 2017 Lincoln MKZ within the TrueTime environment, which accurately models network timing and non-determinism. Testing was conducted on a real-world trajectory from the Richmond Field Station (UCB).

Key Findings:

  • Bandwidth Efficiency: The system achieved a reduction in controller-to-actuator traffic of approximately 92% and sensor-to-controller traffic of 97%.
  • Accuracy Trade-offs: Despite the drastic reduction in data, the average path deviation only increased by 27% compared to a nominal, high-frequency baseline. This represents a highly acceptable trade-off for most autonomous driving scenarios.
  • Robustness: The architecture remained stable and accurate even when subjected to network delays of up to 64ms and packet dropout rates as high as 25%.
  • The Inconsistent Zone: Monte Carlo simulations revealed a critical tipping point; reducing traffic below ~2% of the nominal rate caused performance to become erratic, identifying a clear lower bound for resource optimization.

Conclusion

This research demonstrates that by making autonomous vehicle controllers "network-aware," it is possible to achieve massive gains in communication efficiency with minimal impact on driving precision. Future work will focus on integrating longitudinal control (throttle and braking) and implementing dynamic event-triggering thresholds that adapt to vehicle speed and road curvature to further optimize resource usage.

Summary updated

BibTeX
@article{alite2023resource,
  author = {Guillermo Alite and Ángel Cuenca and Julián Salt and Masayoshi Tomizuka},
  title = {{Resource-Efficient Path-Following Control for a Self-Driving Car in a Networked Control System}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3321269},
}
Figure