Repository navigation
Question of action key in delta action model training #68
Description
Hi, this is a very nice pipeline for humanoid imitation. However, I have a question about README:
"Note that the only difference between the delta action model training and naive motion tracking training is that delta action model needs a motion file with extra keyname
"action"in the motion file, so that the resulting RL policy we are training is able to use the delta action model to"control the robot"to match the real-world/sim2sim motions."
How can we get "action" data for motion file while training? And shall we collect real world data feedback for this "action" while deploying in real world?
Activity
openhe-hub commentedon Aug 13, 2025
If you could provide some examples will be quite helpful!
Ir1222 commentedon Sep 12, 2025
Same question! I wonder if you solve it now? If you do, I really want to know what should be in “action" in motion file? Is it random sequence data or just numbers like 1,2....... Thanks a lot!
zhang705 commentedon Sep 16, 2025
I think they are outputs of actor network when pretrained policy is deployed to a real robot.