Training with image_goal，success rate is 0 ？

[liongliong]

Hi, solid work! I have a question I would like to ask for your guidance. We retrained Octo on new simulation environment data, and when using text as the goal, the success rate reaches 60%. However, when using image as the goal, the success rate drops to 0. Is this a limitation of the Octo model itself? The paper mentions that using images as the goal should perform better than using text, so I am quite confused about this discrepancy.

[gribes02]

Hi, how did you manage to finetune with image goal? I have implemented octo with text as the goal but not image. I don't see what changes I have to make.

[AdriiLosii]

Hi, how did you manage to finetune with image goal? I have implemented octo with text as the goal but not image. I don't see what changes I have to make.

I have the same question, have you been able to finetune Octo with an image goal?