High level policy?

[zengweishuai]

Excited to see your work! May I ask where the arguments for training high-level policy are? BTW, how does the high-level policy get trained? Is it trained in then simulator?

[Yi-Shi94]

Thanks for trying out our work. The arguments are currently in args/. For high-level controllers, it starts with two letters indicating tasks [PH=path, TG=targets, etc.], modes [train, test], and base dataset [lafan1, etc.]. For example, given a pretrained lafan1 base model,

You can train our path following policy with:
python run_env.py --arg_file args/PH_train_amdm_lafan1.txt

You can test our path following policy with: python run_env.py --arg_file args/PH_test_amdm_lafan1.txt

As for the second question, I think you meant the dependency of gymnasium and pybullet. In the implementation of MotionVAE they are used to achieve interactive control and visualization. While these tools are integrated, it's important to note that physics simulation isn't involved, as both our work and MotionVAE are kinematic-based motion generation models.

[zengweishuai]

Many thanks!!