Evaluation of Real-Time Machine Learning for Adaptive Steering Control in Steer-by-Wire System

Sammanfattning

This study evaluates the feasibility and implications of applying a real-time Gradient Boosting for adaptation of steering wheel resistance in a Steer-by-Wire system. Using an Unreal Engine-based vehicle simulator and a mixed-method research design, an online learning model was compared against a fixed offline model through technical system logs and participant evaluation. Technical stability was quantified using the Mean Absolute Change indicator between model updates. Quantitative results indicated that while real-time adaptation is technically achievable, the online model exhibited inconsistent stability. Data-rich regimes, such as driving in a city environment, diverged progressively from the baseline, while sparse regimes like highway driving suffered from abrupt step-jumps in torque responses. Furthermore, data starvation at high steering angles resulted in unnatural flattening of the steering wheel torque curves. Subjective evaluations revealed a unanimous preference for the stable offline system, as participants characterized the online model’s feedback as inconsistent and noted that it felt like the system was ”fighting back”. The study concludes that for SbW feedback systems, predictability and technical robustness are more critical for establishing driver trust than rapid real-time adaptation.