By embedding physics knowledge (differentiable dynamics and adjoint sensitivities) into RL, you can train control policies with 10-100x fewer environment samples and handle high-dimensional systems that standard RL struggles with.
This paper combines reinforcement learning with physics-based optimization to control complex dynamical systems more efficiently. The PEARL algorithm uses automatic differentiation and adjoint methods to compute better policy updates with fewer environment interactions, enabling RL to scale to high-dimensional control problems like navigation in fluid flows.