Code for Mid-Level Visual Representations Improve Generalization and Sample Efficiency for Learning Visuomotor Policies. Arxiv preprint 2018. Alexander Sax, Bradley Emi, Amir R. Zamir, Silvio Savarese, Leonidas Guibas, Jitendra Malik.
I found there are some errors when loading some of the pretrained model. Apparently, it's because the net architecture used for training is different with the net used for reproduction, which refers to the encoder and decoder defined in this repo. We have the same architecture for all the tasks here (except for the output layer). But the trained model use different net architectures for some of the tasks. So maybe it is better to give a clear instruction about which task is supported by this repo.
Besides, I found this repo is pretty similar to the task bank module in the taskonomy repo. Is there any difference between them?