Learning-Based On-Track System Identification for Scaled Autonomous Racing in Under a Minute

TL;DR

The paper tackles the challenge of tire parameter identification for high-speed autonomous racing under varying track and tire conditions by introducing a learning-based on-track system identification method. It couples a nominal Pacejka-based vehicle model with a neural-network residual that corrects model mismatch, and it iteratively updates Pacejka parameters using virtually generated steady-state data to keep predictions in-distribution. Notably, the method identifies tire parameters from as little as 30 seconds of driving data with only 3 seconds of training, achieving up to 3.3× lower one-step RMSE under noise and comparable accuracy to steady-state approaches within critical slip ranges, while avoiding large open-space experiments. The approach is demonstrated on simulation and real F1TENTH hardware, showing data-efficient learning, robust on-track adaptation within 1 second, and practical open-source deployment for reproducibility and extension.

Abstract

Accurate tire modeling is crucial for optimizing autonomous racing vehicles, as state-of-the-art (SotA) model-based techniques rely on precise knowledge of the vehicle's parameters. Yet, system identification in dynamic racing conditions is challenging due to varying track and tire conditions. Traditional methods require extensive operational ranges, often impractical in racing scenarios. Machine learning (ML)-based methods, while improving performance, struggle with generalization and depend on accurate initialization. This paper introduces a novel on-track system identification algorithm, incorporating a neural network (NN) for error correction, which is then employed for traditional system identification with virtually generated data. Crucially, the process is iteratively reapplied, with tire parameters updated at each cycle, leading to notable improvements in accuracy in tests on a scaled vehicle. Experiments show that it is possible to learn a tire model without prior knowledge with only 30 seconds of driving data and 3 seconds of training time. This method demonstrates greater one-step prediction accuracy than the baseline nonlinear least squares (NLS) method under noisy conditions, achieving a 3.3x lower root mean square error (RMSE), and yields tire models with comparable accuracy to traditional steady-state system identification. Furthermore, unlike steady-state methods requiring large spaces and specific experimental setups, the proposed approach identifies tire parameters directly on a race track in dynamic racing environments.

Figures (6)

• Figure 1: Schematic representation of the learning-based on-track system identification method for autonomous racing. The nominal vehicle model is corrected using a residual error model, trained on data collected during on-track, real-world driving. Pacejka parameters are iteratively updated to enhance the overall accuracy of the vehicle model.
• Figure 2: Dynamic single-track model showing lateral forces ($F_{yf}$, $F_{yr}$), slip angles ($\alpha_f$, $\alpha_r$), velocities ($v_x$, $v_y$), steering angle ($\delta$), yaw rate ($\omega$), angle with respect to the positive X axis ($\psi$), and X-Y coordinate system. The symbols $l_r,\,l_f$ indicate, respectively, the distance from the center of gravity to the rear and front axle, and $L$ indicates the full wheelbase length. Note: This image is illustrative; slip angles are exaggerated for visualization purposes.
• Figure 3: The corrected vehicle model structure combining a nominal vehicle model and a to predict the next state variables $v_{y,k+1}$ and $\omega_{k+1}$. Where $f_{\text{model}}(\mathbf{x}_k, \mathbf{u}_k, \mathbf{\Phi_p})$ represents the state transition function based on \ref{['eq:simplified_model_discrete']}, with $\mathbf{x}_k$ and $\mathbf{u}_k$ being the state and input vectors, respectively, and $\mathbf{\Phi_p}$ are the Pacejka parameters.
• Figure 4: Scheme of the iterative approach: Starting with residual error modeling and corrected vehicle model formation, followed by steady-state system identification using virtual data from the corrected vehicle model. Identified parameters are then used as a new nominal model for learning residual errors, repeating iteratively until convergence.
• Figure 5: The of lateral velocity and yaw rate predictions for the models identified using our approach and the on the test set under increasing levels of noise. Experiments were repeated 10 times.