Question about DM0 training loss (L_AR vs L_FM in code).

[ZWenyue]

Hi, thanks a lot for open-sourcing this awesome project!
I’ve been going through the code and the paper, and had a small question about the training loss.
From the paper, I understand that the total loss is defined as: L_total​=λ*L_AR​+L_FM​

But when I looked into the forward implementation, it seems like only the flow matching part is used.

v_t = self.model.action_out_proj(suffix_out_final)
action_loss = F.mse_loss(v_t, u_t, reduction="mean")
loss = action_loss

I didn’t notice any cross-entropy / AR loss being computed (and labels also doesn’t seem to be used here).

Is the AR loss implemented somewhere else?
Or is this corresponding to a later training stage where only flow matching is used?

[lgautel]

Any updates on this? I'm also facing such issue. Or..Has anyone found a workaround for this yet?

[dexmal2025]

Hi all! Sorry for the late reply. During DM0 pretraining, both the AR loss and flow matching loss are computed. Similar to the openpi, we currently only release the pretrained ckpt for finetuning task with only flow matching loss. The hybrid training version will be released in the near future.

[lgautel]

Hi all! Sorry for the late reply. During DM0 pretraining, both the AR loss and flow matching loss are computed. Similar to the openpi, we currently only release the pretrained ckpt for finetuning task with only flow matching loss. The hybrid training version will be released in the near future.

When the 'hybrid training version' will be released?