Question about datasets for pretrained AMP and SMP steering models

[jlee52]

Hi, I have a couple of quick questions about the pretrained models provided by MimicKit.

I noticed that the repository provides pretrained models for both AMP and SMP. I'm wondering which motion datasets were used to train the pretrained models for the Steering tasks.

For AMP steering, was the Steering policy trained using the locomotion dataset listed here?
https://github.com/xbpeng/MimicKit/blob/main/data/datasets/dataset_humanoid_locomotion.yaml

For SMP steering, which motion prior was used for the Steering task? Was it smp_prior_locomotion.pt or smp_prior_lafan.pt?

If it was smp_prior_locomotion.pt, was that prior trained using the motions from the same dataset_humanoid_locomotion.yaml dataset, or was a different motion dataset used?

Thank you.

[Murrol]

You can use either smp_prior_locomotion.pt or smp_prior_lafan.pt for steering. and yes, smp_prior_locomotion.pt is trained using the motions from the same dataset_humanoid_locomotion.yaml.

[jlee52]

Thank you for your answer. I understand that a task policy can be generated using different priors. I wanted to clarify which prior was used for the pretrained task policy smp_steering_humanoid_model.pt prvided by author via this link: was it based on the LAFAN dataset, or on the smaller locomotion dataset?

[Murrol]

smp_steering_humanoid_model.pt was trained with a prior from LAFAN.