Road-Aware Comfort-Oriented Twin-Delayed Deep Reinforcement Learning for Active Suspension Control of Full-Vehicle Systems

Learning-based active suspension control requires a policy that can adapt to changing road excitations, respect actuator feasibility, and optimize ride comfort in a form consistent with vehicle-vibration evaluation. Existing reinforcement learning controllers for suspension systems are often trained on a narrow set of road profiles, use generic acceleration penalties as rewards, and provide limited evidence on how road-aware state design and comfort-aligned reward design affect cross-road generalization. This study develops RAC-TD3, a road-adaptive comfort-oriented Twin Delayed Deep Deterministic Policy Gradient framework for continuous active suspension control. The proposed formulation uses a feasibility-bounded four-actuator action space, a road-conditioned Markov state that combines vehicle dynamic states with road-excitation descriptors, and a tri-axial weighted comfort reward inspired by ISO 2631 seat-vibration assessment. A 7-degree-of-freedom full-vehicle model with driver-seat dynamics, nonlinear suspension characteristics, ISO stochastic roads, hump excitation, and reconstructed real-road elevation is used for evaluation. Across ISO Class A--F roads, humps, and a KITTI-derived real-road profile, RAC-TD3 achieves the lowest average weighted acceleration RMS among the reinforcement-learning variants, reducing the average RMS by 41.18% relative to baseline TD3. Frequency-domain analysis further shows attenuation near the estimated sprung-body and unsprung wheel-hop frequency bands, indicating that the learned policy suppresses physically meaningful vibration modes rather than only reducing time-domain peaks.