memory dimension of RoboDojo

[tuanziqiqi]

Hello, DM0.5 has shown remarkable performance in the memory dimension of RoboDojo. I have replicated it and have a few questions to ask:

1. In the pre-training stage of DM0.5, was the memory only 5 frames?
2. If we do not initialize from the pre-training weights of DM0.5, will the score in the memory dimension of RoboDojo still be so high? That is to say, is it necessary to add memory during the pre-training process?
3. Why are some memory tasks with extremely high success rates, such as cover_blocks 100%. But for some memory tasks, the success rate is extremely low? imitate_sorting_sequence: 0%, swap_blocks: 19%.
4. During training, is it necessary to calculate gradients when extracting 20 historical image tokens?
5. I trained DM0.5 and was able to reproduce the memory on RoboDojo, but it was only about 14%. Could it be that some details were overlooked? Are there any points that need attention?
   Thank you for your reply.