Configuration and files

Every file and directory wbc-mjlab reads or writes, and the formats of the YAML files you may edit by hand.

Directory layout

<repo>/
  data/<robot>/<dataset>/       motion clips — sources (*.csv, *.pkl) and npz/ after conversion
  logs/rsl_rl/<experiment>/     training runs (one folder per run)
  logs/studio/jobs/             Studio jobs: command, log, status (*.json, *.log)
  demos/                        bundled demo checkpoint

~/.config/wbc-mjlab/            per user, never in the repo
  presets/<name>.yaml           Studio presets
  modules/<name>.yaml           Studio modules
  robots/<id>/                  robots added from model files (robot.yaml + model)

A training run:

logs/rsl_rl/<experiment>/<timestamp>_<run_name>/
  model_<iteration>.pt          checkpoints
  events.out.tfevents.*         TensorBoard
  params/
    env.yaml, agent.yaml        the configs as trained
    policy.onnx                 deploy policy (actor only)
    config.yaml                 deploy tracking params (wbc_tracking_params_v1)
    rsi_bin_stats.npz           adaptive-sampling state (when persisted)
    studio_preset.yaml          Studio presets only: the preset as trained

<experiment> comes from the task (Wbc-G1 → wbc_g1); a Studio preset <task>--<name> logs to <experiment>--<name>. config.yaml schema: Export. Data layout: Motion data formats.

Environment variables

Variable	Effect
WBC_MJLAB_CONFIG_DIR	Per-user config directory (default $XDG_CONFIG_HOME/wbc-mjlab, i.e. ~/.config/wbc-mjlab)
MUJOCO_GL	Rendering backend; egl on headless machines, disable for smoke tests
CUDA_VISIBLE_DEVICES	GPUs visible to training; --gpu-ids indexes into this list
WBC_MJLAB_SKIP_MJLAB_REGISTER	1 skips task registration on import (docs builds, tooling)

Studio preset file

presets/<name>.yaml — a diff on top of a registered task, written by the Studio (Preset → Save as) or by hand. Registered as task <base>--<name>.

base: Wbc-G1                  # registered task the preset starts from
robot: g1                     # filled in on save
modules: [se]                 # applied in order, before the overrides
description: SE + keybody refs, history 2
overrides:
  observations:
    actor:
      history_length: 2
      enable_corruption: true
      terms:
        ref_body_pos: {add: true, bodies: [pelvis, left_wrist_yaw_link]}
        joint_vel: {noise_scale: 2.0}
        projected_gravity: {enabled: false}
  rewards:
    motion_body_pos: {weight: 2.5, params: {sigma_per_keybody: 0.15}}
    anti_shake: {enabled: false}
  terminations:
    anchor_pos: {params: {threshold: 0.5}}
  command:
    rsi.bin_width_s: 2.0
    env.episode_length_s: 20.0
  events:
    push_robot: {params: {velocity_range.x: [-1.0, 1.0]}}
  terrain:
    curriculum.start_row: 2
    rough.noise_range: [0.0, 0.15]
  agent:
    algorithm.learning_rate: 5.0e-4
    actor.hidden_dims: [512, 256, 128]

overrides sections

Section	Keys
observations	Per group (actor, critic): history_length (0 = none), enable_corruption, and terms.<name> with enabled: false (drop), add: true (add from the robot’s term library), noise_scale (multiplies the term’s noise; 0 = off), history_length, bodies (keybody subset; [] = all)
rewards	<term>: {enabled, weight, params: {<param>: value}}; dict-valued parameters as <param>.<key>
terminations, events	<term>: {enabled, params: {…}}
command	Motion-command fields by dotted path (rsi.bin_width_s, assistive_beta_max, pose_range.x); env.<field> for env fields (env.episode_length_s)
terrain	Procedural-terrain tasks only: curriculum.start_row, curriculum.num_rows, curriculum.difficulty_range, <sub_terrain>.<field> (rough.noise_range, perlin.proportion …)
agent	Runner fields by dotted path (algorithm.learning_rate, num_steps_per_env, actor.hidden_dims)

The Studio’s Preset step lists every key a task accepts, with its default (GET /api/presets/schema?base=<task>).

Studio module file

modules/<name>.yaml — the same overrides without a base, switched on in any preset (modules: [<name>]).

description: actor without the last action
robot: g1
overrides:
  observations:
    actor:
      terms:
        actions: {enabled: false}

The built-in module se is computed from the robot’s <task>-SE / <task> pair and can’t be edited (copy it with Save as module…).

Run snapshot

params/studio_preset.yaml — written when a Studio preset trains: the preset file with resolved_modules (each module’s name and overrides) inlined. Play rebuilds the run from it, so later edits to the preset or its modules don’t change how the run replays.

Robot spec (robot.yaml)

robots/<id>/robot.yaml — a robot added from its model files (Add a robot from model files).

Field	Meaning
id	Robot id: lowercase letters, digits, _; names the tasks (h1 → Wbc-H1)
model	Source MJCF / URDF, relative to the robot folder (wbc_robot.xml is derived from it on save)
anchor_body	Body the reference is expressed in (usually the torso)
keybodies	Tracked bodies (rewards, RSI, terminations)
ee_bodies	End-effectors (EE height termination, the EE task’s command)
feet_bodies	Feet (contact, air time, conversion height correction)
foot_sites	Sites under the soles for the foot-slip reward (empty: reward dropped)
foot_geoms	Named foot collision geoms for friction randomization (empty: not randomized)
foot_sole_z	Foot body height above the floor when standing (m)
init_height, init_joint_pos	Root height (m) and joint defaults (rad) of the initial pose
joints	Per joint: effort (N·m), stiffness, damping, armature (kg·m²)
action_scale	Policy action = action_scale × effort / stiffness (default 0.25)
nconmax, njmax	Contact / constraint buffers per env
description	Shown in task listings

Python API: Robots without code.

See also

WBC Studio (where these files are edited) · Command-line reference · Studio HTTP API