search

xbpeng / mcp

Implementation of multiplicative compositional policies.
10 stars 1 forks source link

Corpus of different reference motions Pre-training part #4

Open ruoqizzz opened 3 years ago

ruoqizzz commented 3 years ago

Hi,

I'm curious about how did you get the corpus of different reference motions in the pre-training part?

xbpeng commented 3 years ago

The mocap data is from various public databases: http://mocap.cs.cmu.edu/ http://mocap.cs.sfu.ca/

ruoqizzz commented 3 years ago

Thanks!

One more question:

The paper mentioned in the pre-training part:

To encourage the primitives to learn to transition between different skills, the reference motion is also switched randomly to another motion within each episode.

Does that mean that the target states change every time step in each episode? If so, is it possible that the agent receives some target states it can not achieve in only one time step?

xbpeng commented 3 years ago

we change the reference motion every 5-10 seconds, so we do not change the motion at every timestep. It is possible for the agent to be given a target state that it can't reach in one timestep, especially when switching to a new reference motion. But that's not a big problem, since the policy can just takes a few timesteps to transition between the different motions.