train loss and latent_action_accuracy when deploy univla to real world

[BarneetWei]

When deploy univla to real world, train loss convergence to 0.9 and latent_action_accuracy convergence to 0.6;

[BarneetWei]

Is this normal?

[retsuh-bqw]

The latent action accuracy can vary from dataset to dataset. The L1 loss seems acceptable. How many episodes are in your real-world dataset? Also, does the frame interval used for the LAM (which also defines the action chunk size for the decoder) match the frequency of your data?

[BarneetWei]

OK, thanks for your reply. My real-world dataset have 700 episodes, and the frame interval is 20.