questions about Pre-training

[jbpen]

The pre-training data recipe includes synthetic simulation data, real-world robot data, and human videos.
If using human videos for training, how should qt in the formula in the figure be determined?

[Zeng-Jia]

When training with human videos, only the future image generation loss is computed; the action loss $\mathcal{L}_{\mathrm{act}}$ is omitted.

[jbpen]

Ok, thank you for your reply