SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsisaaclab-randomizing-with-events

isaaclab-randomizing-with-events

Agents & models

View repository

Updated 10-15 22:38

About this skill

Randomizing With Events

When To Use

Use this skill when a user wants to add domain randomization to an Isaac Lab task through event terms. Use ADR when policy success should change the randomization range.

Do not use this skill for unrelated curriculum, command sampling, or reward shaping unless those changes interact with randomization.

Workflow

  1. Identify what should vary: assets, physics properties, observations, initial state, external disturbances, or rendering.
  2. Identify the task workflow: direct (DirectRLEnv or DirectMARLEnv) or manager-based. Both can use EventManager through an events config.
  3. Identify the active target backend: PhysX, Newton, or both through PresetCfg.
  4. Check whether the randomization function has backend-specific behavior or unsupported backends in source/isaaclab/isaaclab/envs/mdp/events.py.
  5. Choose the event mode:
    • Use prestartup events for USD-level properties that must be authored before simulation starts.
    • Use startup events for one-time setup randomization after simulation starts.
    • Use reset events for per-episode randomization.
    • Use interval events for repeated disturbances during an episode.
  6. Check timing limitations before editing. Some USD-stage or topology-level changes are prestartup-only and cannot be safely moved to reset or interval events.
  7. Define event terms in the environment configuration. In direct workflows, assign the event config to the task config's events field. In manager-based workflows, assign it to the manager-based env config's events field.
  8. Use backend-specific PresetCfg event configs when PhysX and Newton need different terms.
  9. Scope each term to the correct scene entities.
  10. Use one clear distribution and range for each randomized quantity.
  11. Validate with a small number of environments and repeated resets on each backend.
  12. Expand ranges only after the baseline randomized task is stable.
  13. For manager-based ADR, validate the final range as a fixed event, then use a task-owned scheduler and modify_term_cfg to interpolate it; see the gravity example.

Validation

Use the plan-validate-execute loop:

  1. List each randomized property, target entity, backend, event mode, distribution, range, and timing limitation.
  2. Check the list against the environment config before editing.
  3. Check CPU/GPU expectations in the implementation. Some PhysX paths use CPU tensors while Newton paths may operate on the environment device.
  4. Run a small reset or rollout smoke test for every targeted backend.
  5. Fix shape, device, backend, and entity-name errors before scaling.
  6. For ADR, test the exact endpoints and confirm policy success changes the next reset's range within its bounds.

For skill changes, run:

uv run --no-project python tools/skills/cli.py check

Maintenance

Keep this skill synchronized with source/isaaclab/isaaclab/managers/event_manager.py, source/isaaclab/isaaclab/envs/direct_rl_env.py, source/isaaclab/isaaclab/envs/direct_marl_env.py, the direct and manager-based environment tutorials, and the managers API docs. For ADR, follow the curriculum guide and Core Lift example.

References