VeMo: A Lightweight Data-Driven Approach to Model Vehicle Dynamics

TL;DR

The paper addresses the challenge of modeling vehicle dynamics when detailed physical parameters are unavailable. It proposes VeMo, a lightweight encoder–decoder GRU model with a shared encoder and four output-specific decoders to predict the next state $\mathbf{x}_{n+1}=[a_x,a_y,\dot{\theta},v_x]$ from past states and driver actions. In extensive experiments on a GT3-class racing vehicle using Assetto Corsa data, VeMo achieves accurate one-step predictions and demonstrates robustness to noisy inputs across a range of filtering conditions, all without imposing physical constraints. The findings indicate that a compact, data-driven architecture can deliver physically coherent outputs suitable for integration with control systems and real-time applications in ADAS and Automated Vehicles, with flexibility to extend to other vehicle platforms.

Abstract

Developing a dynamic model for a high-performance vehicle is a complex problem that requires extensive structural information about the system under analysis. This information is often unavailable to those who did not design the vehicle and represents a typical issue in autonomous driving applications, which are frequently developed on top of existing vehicles; therefore, vehicle models are developed under conditions of information scarcity. This paper proposes a lightweight encoder-decoder model based on Gate Recurrent Unit layers to correlate the vehicle's future state with its past states, measured onboard, and control actions the driver performs. The results demonstrate that the model achieves a maximum mean relative error below 2.6% in extreme dynamic conditions. It also shows good robustness when subject to noisy input data across the interested frequency components. Furthermore, being entirely data-driven and free from physical constraints, the model exhibits physical consistency in the output signals, such as longitudinal and lateral accelerations, yaw rate, and the vehicle's longitudinal velocity.

Figures (10)

• Figure 1: Training and test data for the vehicle state. Orange scale color represent training data, and blue scale color represent test data.
• Figure 2: Qualitative representation of the data reshaping process for VeMo architecture. Two sliding windows are used to generate the stack of arrays $\bf X$ and $\bf Y$.
• Figure 3: VeMo neural network architecture and a simplified matrix flow, using B as the batch size. The matrices $W_z$, as described in Appendix \ref{['app:gru']} are schematically represented. The symbol of the curved arrows followed by '$\times100$' indicates that the GRU layer returns the entire sequence of 100 values, while in its absence, it refers only to the last value obtained from the sequential application within the GRU, as shown in Appendix \ref{['app:gru']}.
• Figure 4: Comparison of the power spectral density of each trained model and the reference data for the frequencies under examination.
• Figure 5: Histogram of the relative error for the three outputs, using VeMo trained with data filtered at $0.5\,Hz$, $5\,Hz$, $25\,Hz$ and $45\,Hz$.