Abstract
Achieving the physical limits of autonomous racing requires highly accurate mathematical models of vehicle dynamics, particularly tire-road interactions and engine torque delivery. This paper presents MI-HPO (Model Identification via Hyperparameter Optimization), a novel approach that treats physical vehicle parameters as hyperparameters to be optimized. By adapting the Hyperband algorithm with Gaussian mutation and annealing, the proposed method identifies parameters over 13 times faster than traditional stochastic methods. The system was validated using a full-scale Dallara AV-21 race car, achieving stable autonomous navigation at speeds up to 217 km/h.
Introduction
To push an autonomous vehicle to its handling limits, controllers must rely on near-perfect models of the car's physics. However, identifying these parameters—such as tire stiffness or engine torque maps—presents a significant challenge. Traditional gradient-based optimization is often sensitive to initial conditions and requires differentiable models, while robust stochastic methods like Particle Swarm Optimization (PSO) suffer from slow convergence. While neural networks offer high representational power, they lack the interpretability and safety guarantees required for high-speed racing.
System Architecture and Methodology
The researchers developed the MI-HPO framework to bridge the gap between robust exploration and computational efficiency. The core innovation lies in treating physical parameters (e.g., Pacejka tire model coefficients) as hyperparameters within an "explore-exploit" optimization loop.
The MI-HPO Algorithm
The framework utilizes several key components to ensure rapid and accurate convergence:
- Hyperband Integration: The algorithm starts with a large pool of random parameter configurations. It tests them over short "budgets" (iterations) and aggressively prunes underperforming sets, allocating more resources to the most promising candidates.
- Gaussian Mutation and Annealing: To avoid local minima, the algorithm introduces a mutation step where selected configurations are slightly perturbed with Gaussian noise. The intensity of this mutation is annealed over time to allow for fine-tuning as the search converges.
- Physics-Based Modeling: Unlike black-box approaches, MI-HPO is applied to interpretable models, such as the Pacejka tire model and a 3rd-order polynomial engine torque model, ensuring the results are compatible with standard control theories like LQR.
Experimental Setup and Data
The study leveraged high-fidelity tire property files, engine dynamometer data, and real-world track logs from the AV-21 autonomous race car. Data was collected during high-speed runs at the Indianapolis Motor Speedway (IMS) and Las Vegas Motor Speedway (LVMS).
Results and Discussion
MI-HPO demonstrated exceptional performance in both simulation and real-world testing.
Convergence and Accuracy
The method reached a 1.0 kN error threshold 13 times faster than the next best stochastic method and was 494 times faster at reaching a 0.5 kN threshold. Ultimately, it achieved a terminal error 24% lower than PSO-500.
High-Speed Validation
The identified parameters enabled the AV-21 to reach a top speed of 217.4 km/h (135 mph). Furthermore, the system demonstrated high agility, successfully performing obstacle avoidance maneuvers at 100 km/h with lateral accelerations reaching 9.0 m/s².
Conclusion
MI-HPO provides a highly efficient and accurate means of identifying the complex physical parameters of high-performance vehicles. By leveraging the resource-allocation logic of hyperparameter optimization, the method overcomes the limitations of traditional parameter estimation. Future work will focus on parallelizing the HPO process to enable real-time, online parameter updates, allowing the car to adapt to changing track conditions such as temperature fluctuations and tire wear during a race.
Summary updated