search

xbpeng / DeepMimic

Motion imitation with deep reinforcement learning.
https://xbpeng.github.io/projects/DeepMimic/index.html
MIT License
2.33k stars 489 forks source link

Construction of task goal for jump and backflip #160

Closed y0ngw00 closed 3 years ago

y0ngw00 commented 3 years ago

Your work is so impressive, and I have a question on it.

I wonder what kind of task goal is given for jump and backflip without using phase.

Thank you.

xbpeng commented 3 years ago

There are no goals for imitating the jumping and backflip motions. The only objective was to imitate the ref motion. It should be possible to hand design rewards for those motions too, but could be quite tricky.

y0ngw00 commented 3 years ago

Thank you for your fast reply :)

For adversarial motion prior, is it possible to jump enough using only style reward(reward from discriminator)?

Actually, in my case the character can not jump as high as reference motion. It just pretends to jump or jump lower than reference.

I'm thinking about tuning task goals related to velocity, but I wonder if just style reward is enough to imitate jump or backflip motions.

xbpeng commented 3 years ago

just the style reward should be enough to imitate a jumping clip. Are you training with train_amp_humanoid3d_jump_args.txt, but it's not learning the jump?

y0ngw00 commented 3 years ago

Oh I didn't check args folder. I will try with it. Thank you!