cmjang/mjlab-skillkit

mjlab-skillkit

Skill Suite

• IsaacLab Migration Skill (migrate)
  • Use when there is an IsaacLab source project to port into mjlab-native code.
  • Priority: preserve task behavior while translating APIs, config structure, and registration.
• mjlab Native Skill (author)
  • Use when writing new mjlab-native tasks, configs, managers, sensors, terrain, RL wiring, or mesh workflows directly.
  • Priority: follow mjlab public APIs, existing task patterns, and bundled references.

Previously branded as isaaclab-to-mjlab.

Official Repositories

• Migration pattern (official): https://github.com/mujocolab/anymal_c_velocity
• mjlab repository (read-only reference): https://github.com/mujocolab/mjlab
• IsaacLab repository (source behavior reference): https://github.com/isaac-sim/IsaacLab

Documentation Priority

• Use local docs/ in the target workspace as source of truth first (for example mjlab/docs and IsaacLab/docs).
• Read references/docs-interface-diff.md before implementation and follow the listed API differences.
• Read references/mjlab-api-pack.md for the quick API shortlist.
• Read references/mjlab-api-index.md to choose the exact mjlab API domain files you need.
• When writing new mjlab-native code, also read references/mjlab-authoring-workflow.md.
• For common coding requests, read references/mjlab-authoring-recipes.md and follow the matching minimal recipe instead of exploring broadly.
• Then load only the matching references/mjlab-api-*.md and references/mjlab-mdp-builtins.md files instead of reading the whole API surface at once.
• Do not bulk-ingest the whole mjlab/docs tree or crawl large raw upstream doc sets into context; read only the exact page/signature/example you still need after using the bundled references.
• Fall back to online docs only when local docs are missing or incomplete.

Reference Resolution Order

• Treat raw paths such as mjlab/docs/source/... and mjlab/src/mjlab/... as preferred lookup targets only when a local mjlab/ checkout exists.
• If there is no local mjlab/ checkout, use this skill's bundled references first:
  • references/mjlab-api-*.md
  • references/mjlab-mdp-builtins.md
  • references/mjlab-authoring-workflow.md
  • references/mjlab-authoring-recipes.md
• If the user provides another local mjlab repo path, use that path as the upstream reference root instead of assuming ./mjlab.
• Use online docs / GitHub only as a last resort for exact upstream signatures or examples missing from both the local checkout and bundled references.
• Do not block author mode just because the current workspace has no mjlab/ directory.

Working Modes

• migrate
  • Use when there is an IsaacLab source task/project to port.
  • Goal: preserve behavior while converting to mjlab-native code.
• author
  • Use when writing new mjlab-native code directly.
  • Goal: build tasks/components/configs/scripts that already follow mjlab docs and examples.

If the user does not specify a mode, infer it from the request:

• mentions IsaacLab/source parity/porting -> migrate
• mentions write/build/create mjlab task/config/component -> author

Layout Mode (migration only)

• preserve-layout:
  • Keep original repository structure and module paths.
  • Only migrate API/config semantics to mjlab.
• mjlab-layout:
  • Reorganize into anymal_c_velocity style task package layout (for example src/<task_pkg>/...).
  • Use mjlab.tasks entry points and register_mjlab_task(...) for registration.
• If user does not specify, ask first:
  • Do you want to keep the original project layout, or convert directly to mjlab layout?

Shared Workflow

1. Confirm whether the task is migrate or author.
2. Resolve references in this order: target repo -> local mjlab/ checkout if present -> bundled skill references -> online docs only if still blocked.
3. Read references/mjlab-api-pack.md and lock target APIs.
4. Read references/mjlab-api-index.md, then load only the relevant module references:
   • references/mjlab-api-envs.md
   • references/mjlab-api-scene.md
   • references/mjlab-api-sim.md
   • references/mjlab-api-entity.md
   • references/mjlab-api-actuator.md
   • references/mjlab-api-sensor.md
   • references/mjlab-api-managers.md
   • references/mjlab-mdp-builtins.md
   • references/mjlab-api-terrains.md
   • references/mjlab-api-rl.md
   • references/mjlab-api-viewer.md
   • references/mjlab-api-tasks.md
5. If the request involves importing STL / OBJ / other mesh assets, also read references/mjlab-mesh-import-guidelines.md.
6. Reuse native mjlab patterns from existing tasks in the target repo first, then from a local mjlab/ checkout if present, then from bundled references.

Migration Workflow

1. Confirm migration scope, source path, and target path before editing.
2. Confirm layout mode: preserve-layout or mjlab-layout.
3. Read references/migration-rules.md and enforce all hard constraints.
4. Read references/docs-interface-diff.md for API differences from local docs.
5. Read references/official-migrating-from-isaaclab.md for boundary notes.
6. Read references/migration-gotchas.md for the compressed high-value pitfalls list.
7. If the task is multi-variant / command-heavy / play-eval sensitive, first read references/complex-task-migration-playbook.md.
8. If the task is motion tracking / whole-body tracking / reference-motion tracking, then read references/tracking-case-study.md as a concrete example, not as the only migration pattern.
9. If the target task family has adjacent tests or registry files, treat them as invariants before editing (for example task __init__.py registration files and same-family tests).
10. For complex tasks, first write a variant matrix for base / play / eval / no-state-estimation / low-freq / robot-specific style variants before migrating code.
11. For mjlab-layout, align project packaging/registration with anymal_c_velocity.
12. Read references/mapping.md while replacing imports/fields/term APIs.
13. Read references/patterns.md while implementing EnvCfg/SceneCfg/manager structures.
14. If importing STL / OBJ / mesh assets, read references/mjlab-mesh-import-guidelines.md and explicitly separate visual mesh from collision representation.
15. Run references/checklist.md validation before completion.

Authoring Workflow

1. Read references/mjlab-authoring-workflow.md.
2. Read references/mjlab-authoring-recipes.md.
3. Identify the artifact type:
   • new task package
   • new EnvCfg or scene
   • new manager terms / MDP helpers
   • new sensors / terrain / RL config / registration
4. Choose the closest example before writing:
   • target-repo example first
   • then local mjlab/ example if available
   • then bundled skill references if no local mjlab/ checkout exists
   • velocity-style -> mjlab/src/mjlab/tasks/velocity/
   • tracking-style -> mjlab/src/mjlab/tasks/tracking/
5. Build the config in mjlab-native order:
   • scene / sensors
   • observations
   • actions
   • commands
   • events
   • rewards
   • terminations
   • curriculum / metrics
   • simulation / viewer / episode settings
6. Prefer existing mjlab.envs.mdp helpers before adding custom terms.
7. Prefer the smallest edit surface that satisfies the request:
   • add a term before creating a new subsystem
   • modify robot-specific cfg before cloning a base factory
   • reuse an existing command/reward/event helper before inventing a new class
8. If creating a new task, add RL config plus register_mjlab_task(...).
9. Validate syntax and config wiring before completion.

Authoring Request Routing

• “Add a reward / penalty” -> references/mjlab-authoring-recipes.md + references/mjlab-mdp-builtins.md
• “Add an observation / actor-critic input” -> references/mjlab-authoring-recipes.md + references/mjlab-api-managers.md
• “Add contact / raycast / camera” -> references/mjlab-authoring-recipes.md + references/mjlab-api-sensor.md
• “Add command / curriculum / reset / DR” -> references/mjlab-authoring-recipes.md + references/mjlab-mdp-builtins.md
• “Create a new task / new robot env cfg / register a task” -> references/mjlab-authoring-recipes.md + references/mjlab-api-tasks.md + references/mjlab-api-rl.md
• “Import STL / OBJ / mesh assets” -> references/mjlab-mesh-import-guidelines.md + references/mjlab-api-entity.md
• “Unsure which file to edit” -> first read the “file placement” section in references/mjlab-authoring-recipes.md

Shared Constraints

• Final implementation must be mjlab-native.
• Do not add arbitrary abstractions (extra inheritance/wrappers/major restructuring).
• Do not modify mujocolab/mjlab source code.
• Manager configuration must be dict-based (dict[str, XxxTermCfg]), not manager @configclass.
• Explicitly ban bridge helpers:
  • manager_terms_to_dict
  • AttrDict
  • observation_terms_from_class
• No compatibility layer / adapter shim / transition wrappers.

Migration Constraints

• Preserve behavior equivalence for rewards, observations, actions, commands, reset/events, terminations, and curriculum.
• Keep function boundaries, call order, and config semantics aligned with source unless mjlab API differences require minimal internal changes.
• Prioritize functional/semantic equivalence over literal code-shape equivalence; if mjlab API constraints prevent a one-to-one implementation, use the smallest mjlab-native adaptation that preserves behavior.
• Do not drop source logic steps, config items, or execution order.
• Remove Isaac/IsaacLab API residue (imports, symbols, stale comments, legacy fields).
• Keep source-specific semantic names (for example hack_generator) unless forced field mapping is required.
• Do not keep IsaacLab/Omniverse extension scaffolding by default (ui_extension_example.py, config/extension.toml, omni.* extension files).
• Do not add new raise or assert if the source task has none, unless a minimal check is explicitly required by mjlab/target API semantics for correctness.
• Do not add fallback logic if the source task has none (no broad try/except, no hasattr-style fallback branches, no silent degradations), unless a minimal guard is explicitly required by mjlab/target API semantics for correctness.
• Keep original comments/TODOs. If wording must change, only do minimal mjlab terminology updates while preserving meaning.

Authoring Constraints

• Prefer public mjlab.* modules and local task examples before inventing new helper layers.
• Prefer task factories such as make_xxx_env_cfg() when a reusable base pattern is helpful.
• Reuse mjlab.envs.mdp helper functions/configs whenever they already express the needed behavior.
• Keep new code consistent with existing mjlab task structure, config naming, and registration style.

Execution Notes

• In migrate mode, prefer one-to-one migration, not refactor-oriented rewrite.
• In author mode, prefer nearest mjlab-native example, not a framework-agnostic rewrite.
• For API mismatch, use minimal mjlab-native adaptation instead of bridge code.
• When the question is “which mjlab API should replace this IsaacLab piece?”, choose the target module from references/mjlab-api-index.md first, then read only the matching reference file.
• For mjlab-layout, prefer anymal_c_velocity packaging pattern: standalone task package + mjlab.tasks entry point + register_mjlab_task.
• For preserve-layout, keep directory structure and only migrate API/config/registration wiring.
• Managers should be generated via dict factory functions and initialized with field(default_factory=make_xxx) when dataclass config style is used.
• If inherited config chains are not supported in target style, flatten them via explicit dict merge/override.
• Follow target project registration conventions (prefer project-level registrar; use gym.register only when target project explicitly requires it).