Reinforcement learning-based fusion framework for vehicle sideslip angle estimators using physically guided neural networks

Accurate and robust estimation of the vehicle sideslip angle is crucial for maintaining safety under extreme and variable driving conditions. In this paper, we first propose a reinforcement learning-based framework for estimator fusion in vehicle sideslip angle estimation, where the estimator is constructed using physically guided neural networks incorporating gated recurrent unit (GRU) and self-attention mechanisms. First, the physical knowledge of vehicle kinematics and dynamics is analyzed to design the input configuration of the dynamically and kinematically guided neural network estimators. Second, a GRU-based neural network with self-attention is developed to capture both instantaneous and long-range dependencies in time-series signals, serving as the employed neural network estimator. Third, a reinforcement learning-based estimator fusion framework is proposed to integrate the dynamically and kinematically guided neural network estimators. The estimator fusion is modeled as a Markov decision process (MDP) and implemented using soft actor–critic with auto-entropy, extending reinforcement learning to estimation scenarios where actions do not affect state transitions. Finally, the accuracy and robustness of sideslip angle estimation, as well as the generalizability and adaptability of the reinforcement learning-based estimator fusion framework, are validated through diverse real vehicle experiments using both self-collected and public datasets under normal and extreme maneuvers.