Sarah Imene Khelil

Phd Candidate, Ampere SAS

Behavioral Cloning for Real-Time Vehicle Control Orchestration: From High-Level MPC to Actuator Allocation

This presentation explores the application of behavioral cloning (BC) for vehicle control orchestration, addressing both high-level Model Predictive Control (MPC) and low-level control allocation (CA). Traditional optimization-based approaches, while effective, suffer from computational complexity that limits real-time implementation in embedded automotive systems. We present a comprehensive machine learning framework that uses BC to imitate expert controllers across multiple hierarchical levels. Our work demonstrates that tree-based models (XGBoost, LightGBM, Random Forest) and recurrent neural networks (LSTM, GRU) can effectively replicate MPC behavior for yaw moment control with R² > 0.999 while achieving inference speeds up to 909× faster than conventional MPC (90,920 Hz vs. 10 ms baseline). For control allocation, stacked LSTM networks successfully learn the nonlinear mapping between yaw moments and individual tire forces, maintaining comparable accuracy to quadratic programming methods. Experimental validation on a Renault Austral prototype equipped with Vehicle Dynamics Control (VDC) and Active Rear Steering (ARS) systems demonstrates that ML-based controllers maintain safety constraint violations below 2.5% while generalizing across friction conditions (μ = 0.7 to 1.0). Critical challenges including constraint enforcement, cross-domain generalization, and embedded implementation are addressed through physics-informed feature engineering and hybrid safe-learning architectures.

Bio