Hello, is there any friend successfully trained teacher policy on whole Amass dataset and got an acceptable result? I just trained it for nearly 100k steps (num_envs=4096) on ACCAD, a subset of Amass. But when i evaluated the trained teacher policy using the following command, the success rate could only reach to 60%.
${ISAACLAB_PATH:?}/isaaclab.sh -p scripts/rsl_rl/play.py
--num_envs 1024
--reference_motion_path third_party/human2humanoid/data/h1/amass_all.pkl
--teacher_policy.resume_path ckpts/teacher/1
--teacher_policy.checkpoint model_100000.pt
I tried serval checkpoints saved between 50k to 100k, but the success rates were all below to 60%.
Hello, is there any friend successfully trained teacher policy on whole Amass dataset and got an acceptable result? I just trained it for nearly 100k steps (num_envs=4096) on ACCAD, a subset of Amass. But when i evaluated the trained teacher policy using the following command, the success rate could only reach to 60%.
${ISAACLAB_PATH:?}/isaaclab.sh -p scripts/rsl_rl/play.py
--num_envs 1024
--reference_motion_path third_party/human2humanoid/data/h1/amass_all.pkl
--teacher_policy.resume_path ckpts/teacher/1
--teacher_policy.checkpoint model_100000.pt
I tried serval checkpoints saved between 50k to 100k, but the success rates were all below to 60%.