Physics-Enhanced RL for Real-Time Optimal Control of Nonlinear Dynamical Systems

arXiv·medium signal

Matteo Tomasetto, Nicolò Botteghi and Gabriele Bruni (arXiv 2607.16177, cs.LG/math.OC) embed physical structure into reinforcement learning to make it viable as a real-time feedback control strategy for complex nonlinear systems, where pure model-free RL is too sample-hungry and too unpredictable to deploy. This is part of a steady 2026 trend of hybridizing learned policies with known physics to buy sample efficiency and safety guarantees at once.