Reinforcement learning has become an appealing way to control systems too nonlinear or messy for classical methods. Its cost is appetite: the algorithms are sample inefficient, needing an enormous number of interactions before they arrive at a decent control strategy.

That appetite decides where the method can be used. Every extra sensor and actuator enlarges the space the agent has to explore, and the exploration-exploitation problem grows with it, so applications end up confined to systems with only a handful of sensing and actuation points. The curse of dimensionality is not an abstraction here, it is the reason a promising controller stays in simulation.

The move in this work is to give the learner physics rather than making it rediscover physics from scratch. A system already obeying known dynamics does not need those dynamics inferred from millions of trials, and every constraint supplied in advance is a slice of search space the agent never has to pay to explore.