Command-line reference

Command-line reference#

Every command, with its options. Run any command with --help for the full list. With pip, drop the uv run prefix.

Command	Does
wbc-mjlab-studio	The whole pipeline in a browser
wbc-mjlab-data-to-npz	Export source clips (CSV / PKL) as NPZ — optional, training reads the sources
wbc-mjlab-data-vis	Play exported NPZ clips (no policy)
wbc-mjlab-list-envs	List robots and tasks
wbc-mjlab-train	Train a task
wbc-mjlab-play	Evaluate a checkpoint; exports `policy.onnx` + `config.yaml`
wbc-mjlab-demo	Play the bundled demo checkpoint
wbc-mjlab-export-tracking-params	Write a task’s deploy `config.yaml`
wbc-mjlab-plot-rsi-bins	Plot saved adaptive-sampling state
python -m wbc_mjlab.studio.train_preset	Train a Studio preset

wbc-mjlab-studio#

uv run wbc-mjlab-studio [--host 0.0.0.0] [--port 8765] [--token SECRET] [--ui web|viser]

`--host`	Interface to serve on (default `0.0.0.0`: the network; `127.0.0.1`: this machine)
`--port`	UI port (default 8765)
`--token`	Require this token from network clients (default: no access control)
`--robot` · `--device`	Robot to start with · torch device for playback and conversion
`--ui`	`web` (default) or `viser`
`--no-open-browser`	Don’t open a browser tab

Guide: WBC Studio.

wbc-mjlab-data-to-npz#

uv run wbc-mjlab-data-to-npz --robot g1 --dataset my_clips

Training does not need it (How the library is built): it writes npz/<clip>.npz — body poses and velocities — for tools that want them, baked with the training code path (batched GPU forward kinematics).

`--robot`	Robot model to run forward kinematics on (default `g1`)
`--dataset` · `--dataset-path`	`data/<robot>/<name>/` · any folder; output goes to its `npz/`
`--input-path` · `--output-dir`	A single file or folder in, an explicit folder out
`--format`	Source format (`default` CSV, `gmr_pkl`); default from the file extension
`--input-fps` · `--output-fps`	Source rate (CSV default 30) · export rate (default 50)
`--batch-size`	Frames per FK batch on the GPU (at least 4096)
`--debias-z` · `--debias-foot-sole-z`	Shift clips so the feet touch the floor · override the robot’s sole height
`--mirror`	Also write left-right mirrored clips (needs a symmetry config)
`--device` · `--render` · `--line-range`	Torch device · step the clips through the sim and render them (frame by frame, batch size 1) · CSV row range

Formats and output: Motion data formats.

wbc-mjlab-data-vis#

uv run wbc-mjlab-data-vis --robot g1 --dataset samples

Plays exported NPZ clips in viser with a timeline (source clips: the Studio). Options: --robot, --dataset, --dataset-path, --motion-file, --fps, --cache-motion-bundle.

wbc-mjlab-list-envs#

Lists registered robots and tasks (built-in, extensions, robots added in the Studio) with their log folders.

wbc-mjlab-train · wbc-mjlab-play#

uv run wbc-mjlab-train --task Wbc-G1 --dataset samples
uv run wbc-mjlab-play  --task Wbc-G1 --dataset samples --viewer viser

wbc-mjlab options (parsed first):

`--task <id>`	Task to train / play (robot taken from it) — Choose a task
`--dataset <name>`	Motion library `data/<robot>/<name>/` (source clips, played online)
`--dataset-path` · `--motion-file`	A dataset folder, a clip or an NPZ anywhere
`--mirror` · `--debias-z` (train)	Add left-right mirrored clips · put each clip’s lowest foot on the ground — Prepare motion data
`--cache-motion-bundle`	Converted-NPZ datasets only: write / reuse `<dataset>/<dataset>.npz`
`--robot <id>` · `--no-state-estimation`	Shorthand instead of `--task`: the robot’s default task · its non-SE variant

Everything else passes through to mjlab (mjlab CLI):

`--env.scene.num-envs N`	Parallel environments (default 8192)
`--agent.max-iterations N`	Training length
`--agent.run-name NAME`	Suffix of the run folder
`--agent.resume True` · `--agent.load-run` · `--agent.load-checkpoint`	Continue a run — Train a policy
`--agent.logger tensorboard\|wandb`	Logger (default TensorBoard)
`--gpu-ids 0 1` · `all`	Train on several GPUs
`--checkpoint-file PATH` (play)	A specific checkpoint (default: latest of the task)
`--viewer viser\|native` (play)	Viewer (viser when headless)

wbc-mjlab-demo#

Plays the bundled checkpoint (demos/wbc_g1) on the bundled samples clips. Options include --agent trained|zero|random, --num-envs, --viewer, --video.

wbc-mjlab-export-tracking-params#

uv run wbc-mjlab-export-tracking-params --task Wbc-G1 --out params/config.yaml [--play]

Rewrites the deploy config.yaml of a task (training and play already write it). Schema: Export.

wbc-mjlab-plot-rsi-bins#

uv run wbc-mjlab-plot-rsi-bins logs/rsl_rl/<experiment>/<run>/params/rsi_bin_stats.npz \
  [--trajectory N] [--output plot.png]

Training a Studio preset#

uv run python -m wbc_mjlab.studio.train_preset --preset <name> --dataset samples [train options]

Registers <base>--<name> from ~/.config/wbc-mjlab/presets/<name>.yaml and trains it; accepts every wbc-mjlab-train option. Single GPU only. File format: Studio preset file.