can amp model here reproduce the result of different tasks of ase?

[xjturobocon]

I try to use the following command to train amp model to complete the task just like location python3 ase/run.py --task HumanoidStrike --cfg_env ase/data/cfg/humanoid_sword_shield_strike.yaml --cfg_train ase/data/cfg/train/rlg/amp_humanoid_task.yaml --motion_file ase/data/motions/reallusion_sword_shield/dataset_strike_amp.yaml

However, the result is worse than which produced by ase. For example, The agent trembles when walking. Do I need to adjust some parameters?

[xbpeng]

Do you have a video of the behavior from the model? The strike task can be a bit challenging, and the behavior of the model can vary quite a bit between different runs. Have you try just training again?

[xjturobocon]

here is a demo of location task. I try to train heading, location and strike, the common problem is the "locomotion" skill of these task, just like the demo.

https://user-images.githubusercontent.com/18169613/220001780-c8290b8d-4816-4c11-9bf7-7bf2ebae1a3b.mp4

[xbpeng]

In principal it should be possible to use AMP for these tasks as well, but I haven't really tried tuning AMP for the tasks in this codebase before. Lookin at the video, it kind of seems like the weigh on the discriminator reward might need to be higher in order to encourage more natural behaviors: https://github.com/nv-tlabs/ASE/blob/de18a567649964d5f12b2bc31f44489d3ef26229/ase/data/cfg/train/rlg/amp_humanoid_task.yaml#L91

[Murrol]

In my case, higher disc learning rate while lower disc update frequency might help, in addition to balance reward weights.

[xjturobocon]

In my case, higher disc learning rate while lower disc update frequency might help, in addition to balance reward weights.

Can you show me a copy of your setting? Thank you.

[xizaoqu]

Maybe dataset_strike_amp.yaml doesn't contain enough locomotion clips? Replacing it by dataset_reallusion_sword_shield.yaml seems work well.

[xjturobocon]

Maybe dataset_strike_amp.yaml doesn't contain enough locomotion clips? Replacing it by dataset_reallusion_sword_shield.yaml seems work well.

I also used all motions to train amp location task, but it also didn't work. Have you try it?

[xizaoqu]

Maybe dataset_strike_amp.yaml doesn't contain enough locomotion clips? Replacing it by dataset_reallusion_sword_shield.yaml seems work well.

I also used all motions to train amp location task, but it also didn't work. Have you try it?

https://user-images.githubusercontent.com/45515569/227081501-475a7036-a544-490e-a955-15c81e0a487c.mp4

[xizaoqu]

[xjturobocon]

Can you give me a contact information(WeChat or something else)? I want to ask you for specific information. Thank you.

[xizaoqu]

.