Deep Koopman Operator Learning for High-Dimensional Stability Control in Autonomous Electric Vehicles

Modern transportation has reached the era of Autonomous Electric Vehicles (AEVs) that are becoming the solution for the future. Hence, we need to keep them safe and efficient by combining stability control. Trajectory deviations are minimized using ‘stability control’ mechanisms while the energy consumption is optimised. Nevertheless, control methods such as Proportional-Integral-Derivative (PID), Model Predictive Control (MPC), and Linear Quadratic Regulator (LQR) suffer from handling nonlinearities of vehicle dynamics, consequently providing suboptimal performance in the dynamic and uncertain environment. Furthermore, because of their high computational complexity and instability during the training of Reinforcement Learning (RL)-based methods such as Deep Networks (DQN), adaptations of adaptive control using them are often fooled. Under these limitations, a need for an advanced control strategy that strikes the right balance in terms of trajectory tracking accuracy, energy efficiency, and computational feasibility is necessary. To tackle such challenges, in this paper, we propose a Koopman Optimal Stability Control framework that makes the highly nonlinear dynamics of vehicles stable via the Koopman operator. It allows for state predictions that are more accurate and permits linear control to be used in a data-driven manner. It is shown that the RMSE is reduced by 57% over PID controllers (0.12m vs 0.28m) and by 13.6% over MPC (950J vs 1100J). The Koopman-based stability control system addresses key problems of existing methods that will pave the way for safer, more energy-efficient autonomous electric vehicles in real-world deployments.