---

name: runner-and-backends description: "Use MimicKit's shared runner, backend config triad, distributed device flow, logging/video flags, and backend readiness checks." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

runner-and-backends

Use this sub-skill when a user wants to:

• build or sanity-check a MimicKit train/test command
• choose or validate the engine / environment / agent config triad
• expand an --arg_file preset with safe CLI overrides
• reason about --num_envs, --devices, --master_port, --logger, --video, or --out_dir
• check whether a checkout still has the expected runner/layout files without importing simulator backends

Do not use this sub-skill for motion conversion, motion plotting, or algorithm-specific reward/loss tuning. Route those to the sibling sub-skills.

Start with these bundled references:

• references/runner-cli.md
• references/backend-compatibility.md
• references/troubleshooting.md

Use scripts/check_mimickit_layout.py --repo-root <repo-root> when you need a safe layout/config sanity check that does not import Isaac Gym, Isaac Lab, Newton, or Warp.

Use scripts/run_mimickit.py --repo-root <repo-root> -- <runner flags> when a recipe needs to launch the target checkout's runner through a bundled helper instead of invoking a source path directly.

What this sub-skill should extract from a user request

1. Mode: train or test
2. Preset source: direct CLI flags or --arg_file
3. Engine / environment / agent config files
4. Device plan: single device or distributed --devices plus --master_port
5. Runtime toggles: --num_envs, --visualize, --video, --logger, --out_dir, --model_file

Verified boundary

Current verification only confirms PyTorch CUDA works and source imports succeed with repo-style PYTHONPATH. It does not prove that Isaac Gym, Isaac Lab, or Newton are installed in the current environment. Treat simulator-native workflows as external prerequisites and describe them as such.

Expected answer shape

When answering a user, give:

• the exact runner command shape they should use
• which engine / env / agent files belong together
• which backend prerequisites or data assets are still missing
• which output files to expect in training vs testing
• which troubleshooting branch to follow if the command fails