Robust Design of Low-Thrust Gravity-Assist Trajectories via Reinforcement Learning

This paper presents a reinforcement-learning (RL)-based multiphase robust trajectory design method for low-thrust exploration with gravity assist (GA) under uncertainties. To alleviate the complexity due to the interior-point constraints caused by intermediate GA, a trajectory segmentation strategy is used. The low-thrust trajectory legs before and after GA are segmented into interplanetary transfer phases (ITPs) and approaching phases (APs), respectively, and Markov decision processes with stochastic dynamics are modeled in each phase. Moreover, for the issue of failing to reach the target state of the AP due to the low terminal accuracy of the preceding ITP, a tradeoff factor is defined to modify the nominal initial state of the AP to better guide the exploration of robust policies. In particular, a reachability constraint of the target state of the AP is modeled analytically and incorporated into the reward of the ITP, which can significantly improve the reachability from the end states of RL-based trajectories in the preceding ITP to the target of the AP. Besides, low-thrust robust guidance laws in the AP are trained to deal with the uncertainties over the AP. The promising results in an Earth–Earth–Jupiter mission show that the proposed method can not only effectively deal with various uncertainties but also achieve the desired accuracy for intermediate GA and terminal rendezvous.