Training quickly diverges with default settings

[dementrock]

This is what I got following the default training command

${ISAACLAB_PATH:?}/isaaclab.sh -p scripts/rsl_rl/train_teacher_policy.py \
    --num_envs 10 \
    --reference_motion_path neural_wbc/data/data/motions/stable_punch.pkl

Zooming in

What should be changed?

[huihuaNvidia2023]

Just to clarify, are you using the retargeted punch.pkl dataset?

[huihuaNvidia2023]

I just followed the tutorial and trained on the punch.pkl locally. Here is the tensor board curve

. It converges quite well even after 2000 steps.

In the Omnih2o homepage, it does provide a dataset via CMU

Here is the command I run

${ISAACLAB_PATH:?}/isaaclab.sh -p scripts/rsl_rl/train_teacher_policy.py     --num_envs 1024     --reference_motion_path neural_wbc/data/data/motions/punch.pkl  --headless

[dementrock]

Thanks @huihuaNvidia2023 ! To confirm did you get the pkl from https://github.com/LeCAR-Lab/human2humanoid/blob/main/legged_gym/resources/motions/h1/stable_punch.pkl?

I'll try your setting override of num_envs=1024

[dementrock]

I think I figured out the reason... It was indeed because of the num_envs option. To avoid confusion, it might be better to use this 1024 default value in https://github.com/NVlabs/HOVER?tab=readme-ov-file#teacher-policy rather than using 10?