Physics-Informed RL

Physics-informed reinforcement learning studies how to use knowledge of system dynamics and physical constraints to help agents learn better control policies. Common approaches include embedding physics in neural dynamics models and world models, adding physics-based rewards or policy regularizers, and using physical priors to shape actions or enforce safety. These methods are applied to continuous control, robotics, power grids, traffic, and energy systems, with evaluation often spanning simulation benchmarks and real-world transfer. A central challenge is making physical guidance useful when models are imperfect or dynamics involve contacts, uncertainty, and changing conditions.