Guide · commands verbatim from the official README

Training MotionBricks: three scripts, one pipeline you can verify today

The preview release ships training code for all three model components — VQVAE tokenizer, pose model and root model — and defaults to synthetic data so you can run the whole loop end-to-end before touching the real dataset. Commands below are copied exactly from the README; we did not rewrite or re-test them.

Finish setup first — the training guide assumes a working environment

The three training commands

Run from the motionbricks/ directory of your clone. The README notes the scripts "use synthetic data by default and load model configs from the saved checkpoints in out/":

# Train the VQVAE (motion tokenizer)
python scripts/train_vqvae.py

# Train the pose model (requires pretrained VQVAE checkpoint)
python scripts/train_pose.py

# Train the root model (no VQVAE needed)
python scripts/train_root.py

Order matters once: the pose script needs a VQVAE checkpoint (pretrained in out/, or one you just trained), while the root model trains independently — the README comments say so in as many words. Which component each script corresponds to is covered on the architecture page.

Synthetic data: verify the pipeline first

All current training scripts default to synthetic data (see motionbricks/data/synthetic_dataset.py) "so that the full training pipeline can be verified end-to-end without the real dataset" (README). Practical reading:

You can confirm your environment, configs and the training loop produce checkpoints without downloading anything from BONES-SEED.
Synthetic runs validate plumbing, not quality — a completed synthetic training run tells you the code works on your CUDA box, nothing about motion fidelity.
Reproducibility experiments on the real-data path "are already in flight" per the README.

The real datasets

The README states the datasets used to train the pretrained checkpoints can be downloaded at bones.studio/datasets — that is where the full motion datasets live. The training corpus behind the released model is BONES-SEED (details and number caveats on the assets page).

If you want a lighter reference before committing bandwidth, the checkpoint sizes (~2.2 GB total across four files) are listed in the setup guide.

Your own data or a new robot

Two official docs in the repo handle the customization questions this guide deliberately does not invent answers for:

Motion representation docs

"Details on the motion feature representation, skeleton system, coordinate conventions, normalization, and feature computation pipeline" (README description of docs/motion_representation.md).

Adding your own dataset

"A step-by-step guide to training MotionBricks on your own motion data and adapting it to a new robot" (README description of docs/adding_your_own_dataset.md).

Full release status — what the roadmap actually says

The remaining piece is the complete pipeline integrated with robotics. Quoting the README's roadmap entry verbatim:

"Full training pipeline inside GR00T Whole-Body Control's GEAR-SONIC pipeline — targeted for approximately one month out; reproducibility experiments are already in flight."

Two clarifications this guide insists on: "approximately one month out" is counted from the 2026-04-27 preview announcement, and it is the official roadmap's estimate, not a release commitment — treat any specific date you see quoted elsewhere as somebody's guess. Track the real status on the news page and the official repository; where the training pipeline meets SONIC deployment is mapped on the GR00T-SONIC page.