Enhance Generality by Model-based Reinforcement Learning and Domain Randomization

Autonomous driving is becoming more feasible with advances in learning-based decision-making methods. However, generalization to different scenarios remains a major challenge. We propose a model-based reinforcement learning method called reinforced model predictive control (ReMPC), which mimics the scenario-independent property of model predictive control (MPC). ReMPC has the same input-output structure as MPC, but uses a neural network policy to perform offline training and online implementation (OTOI) for computational efficiency. We also use domain randomization to further enhance the generality of the driving policy during offline training. We evaluate our method on path tracking and autonomous driving tasks. Results show that ReMPC can achieve high accuracy by 99% compared to MPC on path tracking and maintain high performance on autonomous driving even in unseen environments. Our approach demonstrates good generality and high potential for real-world applications.