Time optimal low-thrust rendezvous from an asteroid belt

This regression dataset allows to train models representing the optimal time of flight of a constant acceleration low-thrust trajectory aimed at a randezvous with a target orbiting station (placed at 1.3 Astronomical Units).

The attributes, X, represent the modified equinoctial parameters (p,f,g,h,k,L) of the spacecraft (SI units) at the start of the transfer. L is the true longitude. Different representations of the initial state are likely key to improving any model.

The time of flight, Y,  is given in units of TIME=5022642.890912783s

The dataset was used for value function learning in our paper:

Izzo, D. and Origer, S.: "Neural representation of a time optimal, constant acceleration rendezvous"

where we were more interested on the effects/use of the data augmentation technique called "Backward Propagation of Optimal Examples" than on the accuracy of the resulting neural model. A MAE of ~22 days is obtained on the augmented dataset (~35 days on the non augmented one). Both these numbers can likely be improved considerably, constituting a nice challenge for the community.

The python pickles can be opened as:

with open("filename.pk", "rb") as f:
    X, Y = pkl.load(f)

else , we also provide the corresponding csv files.

* training: contains 3000 items

* test: contains 1000 items

* training_augmented_32_0.0001: contains 96000 items (augmented from the 3000 of training)