name

isaaclab-rl-skillkit

description

Use when developing, customizing, authoring, debugging, reviewing, or porting IsaacLab reinforcement-learning projects that use ManagerBasedRLEnv, DirectRLEnv, DirectMARLEnv, isaaclab_tasks registration, manager terms, direct env hooks, observations, rewards, terminations, events, commands, curriculum, actions, sensors, scenes, RSL-RL wrappers, runner configs, distillation, export, skrl, rl_games, sb3, or version-specific IsaacLab APIs.

isaaclab-rl-skillkit

Required load order

Read the repo's AGENTS.md, local docs, task docs, and training scripts before editing.
Start with selection-guide.md unless the workflow is already proven.
Then load exactly one workflow map:
- manager-based -> manager-based.md
- direct RL or direct MARL -> direct.md
Then load edit-layer-decision-tree.md before choosing files to edit.
Then load exactly one intent recipe.
Then load exactly one source router.
Then load exactly one source leaf doc.
Only load official-docs.md if local code and skill references are insufficient.

Load recipes by intent

Existing manager-based task customization -> manager-customization.md
Existing direct RL or direct MARL customization -> direct-customization.md
New task or variant -> task-authoring.md, then manager-task-authoring.md or direct-task-authoring.md
RSL-RL wrapper, runner, model, loss, distillation, logging, checkpoint, or export -> training-extension.md
Cross-layer env/wrapper/runner/model/registration changes -> architecture-modification.md
Nonstandard downstream repo layout -> repo-portability-guide.md
Debugging or completion checks -> validation-checklists.md
Direct MARL true multi-agent semantics or skrl/rl_games/sb3 boundaries -> scope-boundaries.md

Load source routers by package

source/isaaclab/isaaclab core package -> router.md
source/isaaclab_assets/isaaclab_assets assets package -> router.md
source/isaaclab_rl/isaaclab_rl wrappers and backend integrations -> router.md
source/isaaclab_tasks/isaaclab_tasks task families and registration -> router.md
source/isaaclab_mimic/isaaclab_mimic mimic package -> router.md

Working rules

Treat the local IsaacLab baseline as 2.1.0 when VERSION says 2.1.0.
Prefer local project patterns, task family structure, and vendored rsl_rl code over memory.
Search by symbols when paths differ: gym.register, ManagerBasedRLEnvCfg, DirectRLEnvCfg, DirectMARLEnvCfg, ObsTerm, _get_observations, RslRlVecEnvWrapper.
Do not collapse intentional newer guidance to the checked-out tree. Mark it as newer than the local baseline and verify it against official docs before using it.
If runtime validation is needed, use the repo-intended IsaacLab Python entrypoint, not system Python.

isaaclab-rl-skillkit

Required load order

Read the repo's AGENTS.md, local docs, task docs, and training scripts before editing.

Start with selection-guide.md unless the workflow is already proven.

Then load exactly one workflow map:

manager-based -> manager-based.md
direct RL or direct MARL -> direct.md

Then load edit-layer-decision-tree.md before choosing files to edit.

Then load exactly one intent recipe.

Then load exactly one source router.

Then load exactly one source leaf doc.

Only load official-docs.md if local code and skill references are insufficient.

Load recipes by intent

Existing manager-based task customization -> manager-customization.md

Existing direct RL or direct MARL customization -> direct-customization.md

New task or variant -> task-authoring.md, then manager-task-authoring.md or direct-task-authoring.md

RSL-RL wrapper, runner, model, loss, distillation, logging, checkpoint, or export -> training-extension.md

Cross-layer env/wrapper/runner/model/registration changes -> architecture-modification.md

Nonstandard downstream repo layout -> repo-portability-guide.md

Debugging or completion checks -> validation-checklists.md

Direct MARL true multi-agent semantics or skrl/rl_games/sb3 boundaries -> scope-boundaries.md

Load source routers by package

source/isaaclab/isaaclab core package -> router.md

source/isaaclab_assets/isaaclab_assets assets package -> router.md

source/isaaclab_rl/isaaclab_rl wrappers and backend integrations -> router.md

source/isaaclab_tasks/isaaclab_tasks task families and registration -> router.md

source/isaaclab_mimic/isaaclab_mimic mimic package -> router.md

Working rules

Treat the local IsaacLab baseline as 2.1.0 when VERSION says 2.1.0.

Prefer local project patterns, task family structure, and vendored rsl_rl code over memory.

Search by symbols when paths differ: gym.register, ManagerBasedRLEnvCfg, DirectRLEnvCfg, DirectMARLEnvCfg, ObsTerm, _get_observations, RslRlVecEnvWrapper.

Do not collapse intentional newer guidance to the checked-out tree. Mark it as newer than the local baseline and verify it against official docs before using it.

If runtime validation is needed, use the repo-intended IsaacLab Python entrypoint, not system Python.