Model-Based Assisted Domain-Randomized Reinforcement Learning for Robust Powertrain Vibration Control
Abstract
Automotive powertrain systems exhibit strong nonlinearities and parametric variations that limit the performance of controllers designed from nominal models. While deep reinforcement learning (DRL) can address such complexities, policies trained in simulation often lack robustness under real-world uncertainties, especially when extensive domain randomization is applied. This paper proposes a hybrid control framework that integrates model-based control with domain-randomized DRL for robust vibration suppression of nonlinear powertrain systems. The combined control structure is theoretically formulated within a latent Markov decision process, providing a unified framework for learning under randomized dynamics while preserving a model-based control backbone. A model-based H2 controller generates a baseline control action, while a learning-based policy compensates for nonlinear dynamics and parameter uncertainties. The learning-based component is optimized using Twin Delayed Deep Deterministic Policy Gradient (TD3) with recurrent neural networks, enabling stable policy learning under wide domain randomization. The proposed approach is validated through numerical simulations on a powertrain model with backlash nonlinearity. The results demonstrate improved learning stability and robust vibration suppression under significant parameter variations.