Rocket Powered Landing Guidance Using Proximal Policy Optimization

Rocket recovery requires advanced guidance algorithms to achieve pinpoint landing while satisfying multiple stringent constraints. In this paper, we design a guidance law based on reinforcement learning for the powered landing phase of vertical take-off and vertical landing reusable rocket. To this end, we apply the proximal policy optimization algorithm to develop a control policy that drives the rocket to land at a specified location. The policy parameterized using a neural network is updated by performing gradient ascent algorithm. After abundant amount of training, the learned policy is evaluated in a simulation of the rocket powered landing scenario considering aerodynamic drag, and the result demonstrates the ability of the proposed guidance method to successfully land the rocket from a random initial state.