xbpeng / DeepMimic

Motion imitation with deep reinforcement learning.
https://xbpeng.github.io/projects/DeepMimic/index.html
MIT License
2.34k stars 488 forks source link

Do I need to reset params in character file for different motions or terrains? #51

Open carolrz opened 5 years ago

carolrz commented 5 years ago

@xbpeng Hi, I am training with default configurations in this project. But trained policy only works for some motion.

With the humanoid character file, a good result reproduced for spinkick but not for walking. The final total sample number is around 70 millions (larger than 61e6 in the paper ), but the resulted policy simulator still slipped and fell down.

I want to ask if you changed character params or PDcontroller params for different motions? Or some other trick?

Thank you very much!

xbpeng commented 5 years ago

What does the walking motion look like? We've tested out this code on all the motions, and they should work with the default settings. The PDcontroller parameters should be the same across all motions.

carolrz commented 5 years ago

I just use the given configuration files and motion data:

But I used 32 workers, and the training was not overwhelmed. Does it lead to the poor result?

xbpeng commented 5 years ago

Those settings look right. Do you have a video of what the simulation look like? I've only trained with 16 workers, so not sure if 32 would cause an issue. Might be worth a shot running with fewer workers.

carolrz commented 5 years ago

Thanks for your reply. Sorry that I added some bugs in the code and it worked with the clean one. Thanks again :)