RDGNN

TODO:

• Improve dataloader returning samples

• Improve code structure, make things easier to use, and more modular.

• Improve object processing/implicit ordering.

• Improve data collection, reduce size

• Fix currently unused self-relations

• Fix data-collection task labeling

• Add offline testing

• Add online planning through ros

• Find out how the one hot encoding encoding network is being trained

• Figure out why/what total_loss += self.bce_loss(outs_decoder_2_edge['pred_sigmoid'][:], x_tensor_dict_next['batch_all_obj_pair_relation'][:, :]) is doing.