Add a task or preset

Add a task or preset#

Turn a new method or setup into a --task id that trains, plays and exports like the built-in ones. For one-off tuning without code, a Studio preset is usually enough.

Decide where the change goes#

You want to…

Add

Change weights, sampling flags, thresholds, observation terms

A preset — presets/<method>.py (or stack existing ones)

Train an existing preset on another robot

A task — a builder passing that robot’s body names; the preset stays as is

A new reward, observation or sampling function

A term in env/mdp/ — shared code, open an issue first (Contributing)

Just try different values

A Studio preset — no code (WBC Studio)

Steps#

  1. Write the preset — src/wbc_mjlab/presets/<method>.py:

    def apply_mymethod(cfg, *, body_names: tuple[str, ...]) -> None:
  """One line on the method, with the paper."""
  cfg.rewards["motion_body_pos"].params["body_names"] = body_names
  cfg.commands["motion"].rsi.strategy = "binary_failure"

    Export it from presets/__init__.py. Robot names are arguments, never constants.

  2. Add the task — a builder and a WbcTaskConfig in robots/<robot>/tasks.py (or your extension package):

    def g1_mymethod_env_cfg():
  cfg = g1_base_cfg()
  apply_mymethod(cfg, body_names=G1_MOTION_BODY_NAMES)
  return cfg

WbcTaskConfig(
  task_id="Wbc-G1-MyMethod",
  robot_id="g1",
  description="MyMethod (arXiv:XXXX): …",
  experiment_name="wbc_g1_mymethod",
  build_env_cfg=g1_mymethod_env_cfg,
)

  3. Check it — the task registers, builds and trains:

    uv run wbc-mjlab-list-envs | grep MyMethod
uv run wbc-mjlab-train --task Wbc-G1-MyMethod --dataset samples --agent.max-iterations 10

  4. Document it — a page under docs/source/tasks/ (what it changes, train & play), a card in Choose a task, and the paper in Research & citations.

See also