Mars powered descent phase guidance law based on reinforcement learning for collision avoidance

Abstract

SummaryThis paper proposes a reinforcement learning‐based guidance law for Mars powered descent phase, which is an effective online calculation method that handles the nonlinearity caused by the mass variation and avoids collisions. The reinforcement learning method is designed to solve the constrained nonlinear optimization problem by using a critic neural network. Specifically, to cope with the position constraint (i.e., glide‐slope constraint) and the thrust force limit constraint, a modified cost function is proposed, and the associated Hamilton‐Jacobi‐Bellman equation is solved online without using an actor neural network, which significantly reduces the computational burden. The convergence of the critic neural network is proven. Simulation results show the effectiveness of the proposed method.