About exploiting wrist view camera

[Eddiesyn]

Very impressive work! We’re currently trying to apply it to our own collected dataset. It seems that this work does not require the use of a wrist-view camera. However, for some fine-grained manipulation tasks, a wrist view could be more beneficial. In the current codebase, is there a way to use both the primary view and the wrist view together, or does it only support appealing to a single view?

[retsuh-bqw]

You can simply encode the addition (wrist) view images with an independent encoder and take as input the corresponding embeddings to only the low-level action deocoder. Please refer to this implementation.