Train problem

[caoran2025]

Hi, All. While I train the teacher policy use the given command,
${ISAACLAB_PATH:?}/isaaclab.sh -p scripts/rsl_rl/train_teacher_policy.py \ --num_envs 1024 \ --reference_motion_path neural_wbc/data/data/motions/stable_punch.pkl
the value function loss does not converge even if the iteration step goes to 8W and the learning rate has converged. Is it normal?

[caoran2025]

I chose model_50000.pt and trained a student policy. The sim2sim result seems no problem. But I'm curious why the model can work event if the value function loss does not converge