Architecture

Architecture#

How a registered --task becomes a runnable ManagerBasedRlEnvCfg. For the layer model (MDP, robot, preset, task), start with Overview.

Config pipeline#

Every task cfg is assembled in the same order:

make_base_wbc_env_cfg()       # shared WBC template (env/wbc_env_cfg.py)
     ↓
<robot>_base_cfg()            # registered robot entity
     ↓
apply_<preset>(...)           # paper / deploy recipe (presets/*.py)
     ↓
[optional] apply_se_actor()   # actor obs swap (wbc / zest / end_effector)
     ↓
WbcTaskConfig.build_env_cfg   # --task id

Preset ↔ task concepts: Presets and tasks. Code examples: Choose a task.

Task registration#

WbcTaskConfig (tasks/config.py) binds:

task_id → CLI --task
robot_id → registered entity + data/<robot>/
experiment_name → logs/rsl_rl/<name>/
build_env_cfg → full env cfg callable

Robots added in the Studio (robots/custom.py) follow the same pipeline: their <robot>_base_cfg is built from robot.yaml (model, PD gains, body names) and their tasks stack apply_wbc (+ apply_se_actor for -SE). Studio presets add one more step after build_env_cfg: modules, then the preset’s own overrides (studio/presets.py), registered as <task>--<name>.

Play / eval mode#

Task builders accept play=True to disable training-only features: long episodes, obs corruption, motion DR, assistive wrench, push events. Same pattern for every registered robot.

WBC Studio#

src/wbc_mjlab/studio/ — a UI-agnostic core with services, and two frontends:

Module	Role
`core.py`	`StudioCore`: selected robot (MuJoCo model + CPU kinematics), scene state, event bus, one background task at a time, a 50 Hz tick loop (playback, job polling)
`robots.py` · `robot_import.py`	Robot catalogue, wiring, pose · add / edit robots from model files
`data.py` · `clips.py`	Datasets, clip index / search, raw + NPZ preview, clip QA, conversion and import
`presets.py` · `train_preset.py`	Studio presets and modules (schema, apply, register, run snapshots) · training entry
`training.py` · `jobs.py` · `tensorboard.py`	Detached train / convert jobs (`logs/studio/jobs`), runs, curves, TensorBoard
`play.py`	A checkpoint on N envs in-process: clip pinning, reference ghost, reward terms
`web/`	FastAPI server (JSON API + binary websocket frames) and the three.js UI
`viser_app.py`	The same core as a viser app