Any plans to release the complete pretraining pipeline

[Yuecheng919]

Hi authors,

Thank you for your great work and for releasing the pretrained checkpoints and training code.

I would like to ask whether there are any plans to release the complete pretraining pipeline in the future.

[Nemo-1024]

Thanks for your interest! The training pipeline released in this repository is the same pipeline we use for pretraining; the main difference is the dataset mixture and training scale.
Our pipeline uses LeRobot 3.0 as the unified data format. We convert datasets from different robot embodiments and sources into LeRobot 3.0, then define weighted dataset mixtures for large-scale pretraining.
The pipeline has two data stages:
LAM pretraining: temporally sampled observation pairs are loaded from the LeRobot 3.0 mixtures. This stage only requires visual observations (and robot states when available), so it can use both robot trajectories and human egocentric videos.

LaWAM policy pretraining: we use action-labelled robot trajectories in the same LeRobot 3.0 format. Each sample provides the current observation, language instruction, action chunk, and future observation. Dataset-specific action/state normalization is handled through the dataset statistics.

Downstream SFT, including LIBERO and RoboTwin, uses exactly the same pipeline. We only switch to the target dataset mixture and initialize from the pretrained checkpoint.
We are cleaning up the LeRobot 3.0 conversion, dataset-mixture, and pretraining configuration code for a more complete release.