AMASS/OMOMO teacher training

[cjh000102]

Hi, thank you for releasing OpenTrack.

I am trying to train AnyTracker teachers on my own AMASS and OMOMO motions retargeted to Unitree G1, but the tracking performance is not very stable.

My setup:

• Unitree G1, 29 DoF
• AMASS + OMOMO motions retargeted to G1
• qpos: (T, 36), qvel: (T, 35), 50 Hz
• Training with track_mj.learning.train.train_ppo_track
• Tested 6 teacher groups, following the specialist-teacher idea
• Tested 100M and 500M steps

What I observed:

• The released LAFAN1 generalist works well on LAFAN1.
• It does not track AMASS well, so I trained teachers directly on AMASS/OMOMO.
• The teachers learn some tracking behavior, but many motions are still unstable or fall.
• Increasing from 100M to 500M did not clearly improve the result.

Could you share how the AMASS/OMOMO teachers in the paper were trained?

In particular:

1. How did you split AMASS/OMOMO into teacher groups? Could you share the teacher group definitions or the motion lists?
2. What motion grouping/clustering method did you use for AMASS/OMOMO?
3. Which training config was used? G1TrackingGeneral, G1TrackingGeneralDR, or something else?
4. How many timesteps were used for each specialist teacher?
5. Were there any extra preprocessing or filtering steps for AMASS/OMOMO motions?

I would like to know whether the public training pipeline is enough to reproduce the AMASS/OMOMO teacher policies, or whether some additional settings are required.

Thank you!

[zzk273]

Hi, can you share me some visualization results of the unstable tracked motion and the corresponding retargeted motions?