Hi,
I was working on a similar task like RAM insertion but iam trying to do a cube stacking. i have a fixed blue cube and trrying to stack the red cube on top of it. and Robot (Trossen windowxai) is already grasped the red cube and only randomizing a small workspace. Exactly similar to the RAM insertion.
And the policy did not learned at all in 3 hrs and i have been spending liike 20 days on trying to make hil-serl work and i still trust but i dont know what iam doing wrong with many human interventions. and iam afraid i was doing something wrong.
in my demo dataset i have the masks = 1-dones so we get masks as 0 for last step and for all the other it's 1 but in frnka sim pkl file attachced in guide it has all masks as 1 even last step and i dont know if iam wrong..
and episode max length is 100 so during demo recordings, i used the classifier i trained and each episode has varying lengths. and in the same si franka dataseet every episode has 100 but i dont know if if iam right or wrong.
and. I can share the logs and more technical details if needed

Hi,
I was working on a similar task like RAM insertion but iam trying to do a cube stacking. i have a fixed blue cube and trrying to stack the red cube on top of it. and Robot (Trossen windowxai) is already grasped the red cube and only randomizing a small workspace. Exactly similar to the RAM insertion.
And the policy did not learned at all in 3 hrs and i have been spending liike 20 days on trying to make hil-serl work and i still trust but i dont know what iam doing wrong with many human interventions. and iam afraid i was doing something wrong.
in my demo dataset i have the masks = 1-dones so we get masks as 0 for last step and for all the other it's 1 but in frnka sim pkl file attachced in guide it has all masks as 1 even last step and i dont know if iam wrong..
and episode max length is 100 so during demo recordings, i used the classifier i trained and each episode has varying lengths. and in the same si franka dataseet every episode has 100 but i dont know if if iam right or wrong.
and. I can share the logs and more technical details if needed