Neural representation of a time optimal, constant acceleration rendezvous

We train neural models to represent both the optimal policy (i.e. thrust direction) and value function time of flight) for a optimal, constant acceleration low-thrust rendezvous. In cases we develop make use data augmentation technique call backward generation examples. are thus able produce work with large dataset fully exploit benefit employing deep learning framework. achieve, in all cases, accuracies resulting successful rendezvous (simulated following learned policy) flight predictions (using function). find that residuals as small few m/s, well within possibility spacecraft navigation $\Delta V$ budget, achievable velocity at also that, on average, absolute error predict from any orbit asteroid belt an Earth-like is (less than 4\%) interest practical uses, example, during preliminary mission design phases.