Torque control strategy for in-wheel motor drive vehicles based on hierarchical deep reinforcement learning considering active safety and energy saving

This study addresses the coordinated optimization of active safety and energy saving for vehicles equipped with in-wheel motors, overcoming the limitations of existing torque distribution methods that exhibit incomplete energy-saving objectives and weak multi-objective coordination. A hierarchical deep reinforcement learning approach based on a soft actor-critic algorithm is proposed, in which electric drive and tire slip losses are uniformly quantified and integrated into a unified objective. The framework uses a two-level architecture: an upper-level agent determines the total driving torque and reference yaw moment for longitudinal and yaw stability regulation, while a lower-level agent optimizes individual wheel torque allocation by adaptively balancing safety and energy saving. Validation shows that hierarchical decomposition substantially improves training stability and convergence versus single-agent methods. Under double lane change maneuvers, the method achieved a longitudinal speed deviation below 0.60 km/h and a maximum lateral displacement error of 0.20 m while effectively constraining the yaw rate and sideslip angle. Step steering tests confirmed robust generalization. Extended validation on a 5-km 3D road demonstrated enhanced active safety while simultaneously achieving energy loss reductions from 6.62% to 16.77%. Hardware-in-the-loop tests under ±10% mass perturbations revealed consistent performance across all dynamic metrics between real-time and simulation environments, with energy saving deviations confined to within 2%. The approach provides a practical foundation for intelligent coordinated control in future electric vehicles.