Question about transition function

[Umesato]

Hello, Thanks for your excellent work! I have some questions about transition function:

1. I noticed tvalue was not used to optimize searching policy (block_assembly task) in backward fine-tuning, and tvalue was not included in the environment code of searching policy. It seems a little different from the paper?
2. The paper mentions that the input of the transition function is terminal state s[T-10:T], but in code input_dim=4 only contains self.camera_view_segmentation_target_rot. Is it an adjustment for the real-world experiment? If I just want to run in sim, should I change the input back to the state of the hand and object?

[cypypccpy]

Hi @Umesato ,

Thanks very much, and we appreciate your detailed observations. The main reason for these problems is that there are so many baselines and settings in our project that some of them are mixed to cause mistakes, and the default configuration in our repo is to train the model used in the BlockAssemblyGUI task. We'll also be cleaning up better code in the future to make it easier to use.

For 1, I accidentally put the baseline setting without tvalue during this release, now I have fixed it. For 2, yeah, this is the setting after distillation (mentioned in the paper) for the real-world. If I just want to run in sim you can change back to the original, just like we did in the searching task.

Hope this can help you.

[Umesato]

Thank you very much!