Modularity

Modularity#

wbc-mjlab separates one shared MDP from robot wiring and paper recipes. Nothing in env/mdp/ should hard-code a platform or a specific paper ablation.

Layer stack#

┌─────────────────────────────────────────────────────────┐
│  Shared WBC MDP (env/) — robot-agnostic term callables  │
│  Actions · Commands · Observations · Rewards · …        │
└───────────────────────────┬─────────────────────────────┘
                            │ make_base_wbc_env_cfg()
┌───────────────────────────▼─────────────────────────────┐
│  Robot entity (<robot>_base_cfg) — must be registered   │
│  MJCF · actuators · sensors · body name tuples          │
└───────────────────────────┬─────────────────────────────┘
                            │
┌───────────────────────────▼─────────────────────────────┐
│  Presets (presets/) — cfg mutators only                 │
│  apply_wbc · apply_zest · apply_binary_failure · …      │
└───────────────────────────┬─────────────────────────────┘
                            │
┌───────────────────────────▼─────────────────────────────┐
│  Tasks (WbcTaskConfig) — CLI ``--task`` ids             │
│  preset stack + experiment_name + build_env_cfg         │
└─────────────────────────────────────────────────────────┘

What each layer owns#

Layer

Responsibility

MDP

Term implementations — same code for every robot and task

Robot entity

Which MJCF, bodies, sensors, and action scales attach to the template

Preset

How rewards, RSI, obs, and terminations are weighted for a method

Task

A named, registered preset stack + log directory for the CLI

Core principle: compose cfg, don’t fork envs. Paper differences are preset + task builders; new robots are extension packages — not copies of env/.

Reference observations are a concrete example: papers and sensing stacks disagree on the actor layout, and per-term noise (as in SONIC-style training) is natural only when reference kinematics are split into modular ref_* slots — not a fixed vector inside the command. Details: Observations.

What not to duplicate#

  • New reward/RSI callables → shared env/mdp/ (needs review)

  • New paper weights / flags → presets/<method>.py

  • New platform → registered robot entity + tasks

  • New --task id → WbcTaskConfig + builder

See Architecture overview for the assembly pipeline and Presets and tasks for how presets relate to tasks.