motion_preparation.md

Preparing Motions for Tracking

booster_mjlab uses WandB to manage motion datasets for the Mjlab-Tracking-* tasks. Clips can come from GMR (.pkl, the format produced when retargeting GVHMR video reconstructions) or from BeyondMimic (.csv).

Set your WandB entity first:

export WANDB_ENTITY=your-organization-name

GMR motions

upload-motion converts a .pkl clip to the tracking format and uploads it to WandB in one step:

uv run upload-motion --input-file /path/to/motion.pkl --entity your-org

The motion is uploaded as an artifact named <file-stem>-tracking to the motion_upload project (override with --project and --artifact-name). The command prints the artifact's full name, e.g. your-org/motion_upload/motion-tracking:v0, which you pass as --registry-name when training.

BeyondMimic motions

First create a registry collection in your WandB workspace named motions, then run:

MUJOCO_GL=egl uv run python -m booster_mjlab.scripts.csv_to_npz \
    --input-file /path/to/motion.csv \
    --output-name motion_name \
    --input-fps 30 \
    --output-fps 50 \
    --render

This uploads the motion to the motions registry as your-org/motions/motion_name, which you pass as --registry-name when training.

Next steps

See the Motion Imitation section of the README for training and evaluation commands, and use uv run visualize-motions to inspect clips before training on them.