A Tricycle Model to Accurately Control an Autonomous Racecar with Locked Differential

Ayoub Raji, Nicola Musiu, Alessandro Toschi, Francesco Prignoli, Eugenio Mascaro, Pietro Musso, Francesco Amerotti, Alexander Liniger, Silvio Sorrentino, Marko Bertogna

Abstract

In high-speed autonomous racing, the standard single-track "bicycle" model often fails to account for the complex dynamics of vehicles equipped with locked differentials. This paper presents a novel "tricycle" vehicle model that bridges the gap between computational efficiency and physical fidelity. By modeling the rear axle as a double-track system and implementing a micro-step discretization approach within a Model Predictive Control (MPC) framework, the proposed system significantly improves path-tracking accuracy. Experimental results from the Monza Formula 1 circuit demonstrate a reduction in lateral tracking error from 1.22m to 0.23m in tight corners, maintaining stability at speeds up to 270 km/h.

Introduction

Autonomous racing requires vehicle models that can handle the extreme limits of tire grip and high-speed maneuvers. While the bicycle model—which collapses each axle into a single central wheel—is the industry standard due to its simplicity, it assumes an open differential. However, many professional racecars, such as the Dallara AV-21 used in the Indy Autonomous Challenge (IAC), utilize a locked differential or rigid axle. This mechanical configuration forces both rear wheels to rotate at the same speed, introducing a resisting yaw moment that traditional models cannot predict. This paper introduces a hybrid "tricycle" model designed to capture these specific dynamics without the computational burden of a full four-wheel model.

The Challenge of Locked Differentials

A locked differential significantly alters a vehicle's handling characteristics. During turn entry, the rigid axle creates a "nose-out" behavior (understeer) as the wheels resist the difference in path length between the inner and outer tires. Conversely, during turn exit, the system can exhibit "nose-in" behavior.

Standard bicycle models are unable to distinguish between the physics of a tight chicane and a wide-radius curve under these conditions because they lack the lateral dimension on the rear axle. This leads to significant tracking errors or numerical instability when the controller attempts to push the car to its physical limits. For autonomous systems, this inaccuracy often necessitates conservative speed profiles, preventing the vehicle from achieving competitive lap times.

Methodology: The Tricycle Model

Model Architecture

The authors propose the Tricycle Model, a hybrid architecture where the front axle is modeled as a single wheel (similar to a bicycle model) while the rear axle is modeled with two distinct wheels (double-track). This configuration allows the controller to calculate the specific longitudinal forces acting on each rear wheel independently.

Physics of the Rigid Axle

The model estimates the yaw moment ($M_{diff}$) generated by the rigid axle using Pacejka tire models and slip ratio calculations. By accounting for lateral load transfer—recognizing that the inner wheel loses vertical load during a corner—the MPC can accurately anticipate the understeer caused by the differential. This allows the vehicle to "pre-compensate" for the resisting moment of the axle.

Micro-steps Discretization

High-speed vehicle dynamics often result in "stiff" differential equations that become numerically unstable at lower speeds or during rapid direction changes (e.g., in tight chicanes). To maintain a high control frequency (100Hz / 40ms) while ensuring stability, the researchers implemented a micro-step approach. This involves using a smaller 8ms integration step within the larger control loop, allowing for a fast Euler integrator to remain stable even in complex maneuvers.

Experimental Results

The model was validated using a high-fidelity multi-body simulation in Dymola and real-world experimental data from a Dallara AV-21 at the Monza F1 circuit.

Path-Tracking Accuracy

The tricycle model outperformed the standard single-track (ST) model across all sectors of the track, with the most dramatic improvements seen in high-curvature sections. In the "Variante 1" chicane, the tricycle model achieved:

  • Lateral Error: Reduced from 1.22m (ST) to 0.23m.
  • Heading Error: Reduced from 4.14° (ST) to 0.86°.

High-Speed Stability

The system successfully managed the vehicle at speeds reaching 270 km/h and lateral accelerations of 25 m/s² (approximately 2.5G). Unlike the bicycle model, which often requires different parameter sets for different types of corners, the tricycle model provided consistent accuracy across the entire circuit without manual re-tuning.

Conclusion and Future Work

The tricycle model provides a vital "middle ground" for autonomous racing, capturing the asymmetric rear forces of a locked differential with minimal computational overhead. While highly effective for rigid axles, future research will focus on extending this formulation to handle limited-slip differentials (LSD). Additionally, incorporating roll dynamics may further improve accuracy for vehicles with softer suspensions, such as GT cars, where lateral load transfer is more pronounced.

Summary updated

BibTeX
@inproceedings{raji2023tricycle,
  author = {Ayoub Raji and Nicola Musiu and Alessandro Toschi and Francesco Prignoli and Eugenio Mascaro and Pietro Musso and Francesco Amerotti and Alexander Liniger and Silvio Sorrentino and Marko Bertogna},
  title = {{A Tricycle Model to Accurately Control an Autonomous Racecar with Locked Differential}},
  booktitle = {ICSC},
  year = {2023},
  doi = {10.1109/icsc58660.2023.10449744},
}
Figure