Vehicles and car-like robotic platforms operate on diverse and often irregular terrains, where traditional suspension regulators face limitations in maintaining load safety, ride comfort, and adaptability. This work contributes to the analysis and control of active suspension systems under varying road and operating conditions. A machine-learning-based control framework is proposed for nonlinear active suspensions, addressing the requirements of payload protection and performance robustness on changing surfaces. The research develops a nonlinear quarter-vehicle model with progressive springs, hysteretic friction, adaptive damping, load variation, speed-dependent tire stiffness, and actuator dynamics. Road disturbances follow ISO 8608 Classes A–E. The Twin-Delayed Deep Deterministic Policy Gradient (TD3) controller is compared with passive suspension, LQR, and PID across four speeds and five road classes. Performance is evaluated using RMS body acceleration (comfort), suspension travel (safety), and actuator effort (efficiency). The results show substantial improvements relative to conventional methods. The TD3-based controller provides up to 64% reduction in RMS body acceleration compared to passive suspension and achieves an average 42% improvement across all tested road classes. The advantages of the machine-learning approach become more pronounced with increasing terrain roughness and vehicle speed, where nonlinearities and actuator limits degrade the performance of classical controllers.