Reinforcement Learning Based Methods for Generating Helicopter Autorotation Trajectories

Autorotation is an emergency flight maneuver in which a helicopter descends safely without engine power by using rotor energy. This paper investigates the use of reinforcement learning (RL) for autorotation trajectory generation and systematically evaluates it against optimal control problem (OCP) solutions. A one-degree-of-freedom powered descent problem is first solved as a surrogate to identify robust hyperparameter settings. The surrogate case results demonstrate that the RL policy closely matches the OCP solution in terms of landing time, confirming its effectiveness. The autorotation problem is then solved under both frameworks, and the resulting Height-Velocity diagrams are compared, with crash behavior in the deadman zone analyzed for each. The RL framework is shown to produce autorotation trajectories comparable to OCP, establishing it as a viable real-time alternative. Warm-starting the OCP with RL-derived solutions improves convergence compared to conventional initialization. Finally, the RL policy's versatility is discussed with an example of varying initial helicopter weight, capturing different fuel states at engine failure.