How are previous action sequence as an input condition provided during the 1st inference step

[MLDeS]

Great work!

Could you please clarify, how you provide the previous action sequence as an input condition during the first step of inference? If it is provided as a random noise sequence, doesn't it affect the generated action sequence for the subsequent steps, and the error propagates?

[WhoKnowsssss]

Hi there,

In evaluation, we pad previous action with zeros at the first timestep. This should be fine because action is zero-centered (normalized by the nominal dof positions) and the robot is spawned in the nominal pose as well.