Inconsistency between your code and paper.

[WuTanKun]

Hi! Thanks for your excellent job! I‘m a little confused about the PatchMoeResNet module obs_encoder.obs_nets.agentview_image.backbone = PatchMoeResNet(k = k, . It seems that the data didn't pass through the PatchMoeResNet, and x = self.nets(inputs) was called instead. In the paper, you mentioned that Vision MoE really improves the model’s ability to handle continuous learning. I don't know whether Vision MoE corresponses to PatchMoeResNet in the code. Or some details I missed make me come through this problem.

[AnthonyHuo]

Thanks for your interest, the vision moe part is used in continue learning, not in multi-task learning. The continue learning part code will be coming soon.