Repository for the development of my master thesis on control of launch vehicles descent and landing through reinforcement learning actors.
To understand better the algorithms and its working and to get an idea of the amount of time required to train the model a simplified environment is developed using as a reference the 3DOF environment. This is implemented easily by using the class gym.Wrapper. This environment returns as observation a vector containing the height and vertical speed of the rocket. The action is just the thrust as it is unnecessary to control gimbaling of the engine.
The initial attempt to control the 1DOF rocket was executed by testing PPO with a continuous action space, with the engine allowed to throttle between maxThrust and minThrust. The thrust was normalized to lie in the range [-1, +1] as best practice for convergence of the algorithms suggest. The initial results with the default hyperparameters for the PPO policy are shown hereafter.
[ADD IMAGES FROM TENSORBOARD]