ICANEWS

Engineering & Technology

Vehicle Prediction Model for Enhanced MPC Path Tracking in Formula Student Driverless

This research presents a real-time capable prediction model designed for autonomous race cars, specifically within the context of Formula Student Driverless com

Overview

This research presents a real-time capable prediction model designed for autonomous race cars, specifically within the context of Formula Student Driverless competitions. The model aims to improve path tracking performance by addressing the challenges posed by highly nonlinear vehicle behavior and narrow track conditions. It integrates information from previous driving sessions and current driving dynamics to adjust predictions dynamically.

Research Context

Autonomous race cars, particularly those participating in Formula Student Driverless, operate at the limits of their physical handling capabilities. This operational domain results in highly nonlinear vehicle dynamics, which complicates path tracking, especially on confined tracks. Model Predictive Control (MPC) is a common methodology employed for managing these challenges. The efficacy of MPC is directly dependent on the precision of its underlying prediction model. Existing approaches often struggle with adaptability to changing conditions and the efficient incorporation of diverse data sources while maintaining computational tractability.

Approach

The proposed prediction model is structured into three sequential submodels, designed to achieve real-time capability and high prediction accuracy while providing quantitative uncertainty assessment. The architecture begins with a nominal Kinematic Bicycle Model, which serves as a foundational component for vehicle dynamics prediction. Following this, an offline Bayesian Linear Regression (BLR) model is incorporated. This BLR model processes information collected from past runs. The third component is an online Sparse Gaussian Process Regression (SGPR) model, which adapts the predictions based on the current driving situation. This three-tiered approach allows for the efficient integration of all available data without significantly increasing computational cost.

Findings

  • The developed prediction model demonstrated a significant improvement in prediction accuracy, achieving an enhancement of up to 57% when compared to existing methods.
  • The model's design enables quantitative uncertainty assessment from the initial stages of a run.
  • The practical applicability of the model was successfully demonstrated. This demonstration involved integrating the prediction model within an MPC-based path tracking controller.
  • Testing confirmed the model's functionality and performance on a real Formula Student race car.

Why This Matters

The ability to accurately predict vehicle behavior under near-limit conditions in real-time is crucial for enhancing the performance and safety of autonomous race cars. The described model addresses the complexity of highly nonlinear vehicle dynamics inherent in competitive autonomous driving. Its demonstrated improvement in prediction accuracy and real-time operational capability could lead to more robust and precise path tracking in challenging environments.

Key Findings