Abstract
In recent years, the automotive industry has undergone a rapid evolution, with increasing interest in electric vehicles (EVs). In this context, improving driving comfort (beyond the reduction of internal-combustion-engine related vibrations) and vehicle safety have become crucial factors for the widespread adoption of EVs. This can be achieved thank to the easier modulation of the individual propulsion torques on both axles, which can be exploited for enhancing stability/agility, but also to improve the reaction of the vehicle to external factors, such as road asperities. In particular, longitudinal vehicle vibrations, although often overlooked in comparison to vertical vibrations, have a significant impact on driving comfort. This paper devises a Reinforcement Learning (RL) torque-correction controller aimed at reducing longitudinal vibrations. The proposed controller, based on the Deep Deterministic Policy Gradient (DDPG) algorithm, has been implemented on a validated vehicle model for both the front and rear axles, modulating the torque requested by the driver in response to road unevenness or obstacles. The agent was trained/tested in simulations on various road scenarios, including a variety of obstacles as well as real road profiles sampled experimentally. The results show that the proposed RL controller offers performance comparable to that of a state-of-the-art Nonlinear Model Predictive Controller (NMPC) in terms of vibration reduction, while maintaining a much lower computational load, making it potentially suitable for real-time applications in road vehicles equipped with production automotive-graded ECUs. In this proof-of-concept study, the controller was implemented on a dSPACE MicroAutoBox III device, showing a turnaround time < 1 ms compared with the 6−7 ms required by the considered real-time capable NMPC, on the same unit. Robustness tests were also carried out to assess the controller’s response against different road conditions, road detection accuracy and vehicle parameters, confirming the validity of the proposed approach.