Approaches for Synthesis and Deployment of Controller Models on Automated Vehicles for Car-following in Mixed Autonomy

Amaury Hayat, Rahul BhadaniCps-iot Week Workshops 2023 · DOI

Introduction

Bridging the "valley of death" between theoretical control algorithms and real-world deployment on physical hardware remains a significant challenge in autonomous vehicle research. While laboratory models often simplify vehicle dynamics as point masses and assume perfect communication, real-world systems must contend with complex physical constraints, significant delays, and non-uniform data sampling. Traditional development cycles are often too rigid for the rapid iteration required in modern research. This paper proposes an agile, safety-conscious software design pattern and lifecycle for transitioning automated vehicle controllers from mathematical theory to highway deployment.

Methodology: A Multi-Stage Deployment Pipeline

The researchers developed a robust framework designed to allow non-software experts to safely deploy and refine controllers. This lifecycle consists of several primary stages:

  1. Model-Based Design: Controllers are synthesized in environments like Simulink or WebGME, enabling automated code generation into Robot Operating System (ROS) nodes. This ensures mathematical consistency across simulation and hardware.
  2. Software-in-the-Loop (SWIL): The generated code is tested in simulated environments (e.g., Gazebo) that introduce real-world constraints like finite delays and varied sampling frequencies.
  3. Hardware-in-the-Loop (HWIL) & Ghost-Car Testing: Physical vehicles are tested using a "ghost car"—a virtual leader derived from real-world trajectory data. This allows the controller to interact with actual vehicle dynamics and CAN bus messages without the risk of physical collisions.
  4. Read-Only Testing: The experimental controller runs in live traffic but its commands are intercepted. The vehicle remains under the control of its stock Adaptive Cruise Control (ACC) while researchers compare the experimental "ghost" commands against the stock system's performance.
  5. Deep Learning Integration: For reinforcement learning models, the team utilizes the Open Neural Network Exchange (ONNX) format and a custom ONNX2ROS package to standardize deployment interfaces.
Figure 3: Comparison between (a) standard command flow and (b) read-only mode where experimental commands are intercepted.

Architectural Insights and Safety

A core component of this framework is the decoupling of the controller's "intelligence" from the vehicle's "mechanics." By defining standardized ROS topics for inputs like speed and headway, a single controller can be deployed across different vehicle makes (e.g., Toyota and Nissan) with minimal modification.

Safety is maintained through a supervisory architecture. A dedicated "Safety Controller" (such as the Followerstopper) acts as a final gatekeeper between the experimental model and the vehicle's actuators. If an experimental model—particularly a "black-box" deep reinforcement learning (DRL) model—suggests an unsafe maneuver, the safety layer overrides it to maintain safe following distances.

Figure 1: The control architecture featuring the traffic-smoothing controller and the supervisory safety controller.

Results and Scaling

The framework was validated during the CIRCLES project, which involved deploying 5-10 automated vehicles on the I-24 highway in Nashville. The methodology enabled rapid iteration, allowing for several dozen parameter refinements per week—a process that typically takes months in traditional automotive engineering. The successful integration of DRL models via the ONNX2ROS pipeline demonstrates the framework's versatility. These methodologies are currently being scaled to support experiments involving over 50 vehicles.

Conclusion and Future Work

While the framework successfully mitigates deployment risks, challenges remain in analyzing the internal logic of deep learning models. Future work will focus on scaling the infrastructure to 50+ vehicles and better accounting for the unpredictability of human drivers in mixed-autonomy environments. This research provides a scalable blueprint for moving from theoretical traffic-smoothing algorithms to impactful real-world deployments.

Summary updated

BibTeX
@inproceedings{hayat2023approaches,
  author = {Amaury Hayat and Rahul Bhadani},
  title = {{Approaches for Synthesis and Deployment of Controller Models on Automated Vehicles for Car-following in Mixed Autonomy}},
  booktitle = {Cps-iot Week Workshops},
  year = {2023},
  doi = {10.1145/3576914.3587711},
}
Figure