How to convert RT-1-X action to Octo action

[MasterXiong]

Hi,

I wanted to use successful episodes collected by RT-1-X in the google robot envs as demos to fine-tune an Octo model. I was wondering that how could I convert the raw action generated by RT-1-X to the corresponding raw action of Octo? As I note that the inference function for the two models are different in simpler code, which may indicate that they do not use the same action space and some coordinates transformation is required here? Thanks!

[xuanlinli17]

Just compare the 2 inference code. I think only the gripper is different.

[MasterXiong]

Thanks! But I noticed that for the actions dimensions of orientations, RT-1 uses axis angle representation for google robot tasks, and roll, pitch, yaw representation for widowx tasks (link). And Octo seems to use roll, pitch, yaw for both kinds of robots. So should I transform axangle to euler for google robot?

[xuanlinli17]

I think so.