The interface is localized. Skill descriptions and documents are machine-translated by AI into your language; code, commands, paths and technical terms stay in the original, and the author's document governs.
Explore skills
Productivity · SKILL.md

mimic-kit

mimic-kit

Use MimicKit repo-specific guidance for motion imitation, physics-simulator RL training, motion conversion, and SMP prior workflows.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill documentation

GitHub

name: mimic-kit description: "Use MimicKit repo-specific guidance for motion imitation, physics-simulator RL training, motion conversion, and SMP prior workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MimicKit Repo Skill

Use this skill when a task is about MimicKit, a checkout-oriented motion imitation and physics-simulator control framework. It covers the repository's runner, simulator backend choices, motion data tools, DeepMimic/AWR/LCP recipes, AMP/ADD/ASE recipes, and Score-Matching Motion Prior workflows.

Read references/repo-provenance.md before deciding whether this skill is current for a checkout. Read references/troubleshooting.md when the request starts with an install, backend, missing asset, or config-path failure.

First decision

  1. If the user needs to build, launch, or debug a train/test command, load runner-and-backends.
  2. If the user needs motion conversion, motion-file validation, view_motion, DoF diagnostics, or plotting logs, load motion-tools.
  3. If the user asks for DeepMimic, AWR, LCP, vault/static-object, or PPO-style motion tracking, load motion-imitation.
  4. If the user asks for AMP, ADD, ASE, discriminator rewards, task-location, task-steering, or adversarial imitation, load adversarial-control.
  5. If the user asks for SMP, TinyMDM, score-matching priors, GSI, smp_prior_cfg, smp_prior_model, or SMP task policies, load smp.

Important repository facts

  • MimicKit is not packaged by pyproject.toml, setup.py, or setup.cfg in this checkout. Treat it as a checkout-oriented project, not as a pip-installable distribution.
  • Commands and helper scripts should accept an explicit MimicKit checkout root. Do not assume the current working directory is the target checkout.
  • The main runner expects the checkout's root and mimickit/ directory to be importable. The bundled runner wrapper in sub-skills/runner-and-backends/scripts/run_mimickit.py prepares that import layout.
  • Runtime Python dependencies are listed in requirements.txt, but simulator backends are external: Isaac Gym, Isaac Lab/Isaac Sim, or Newton/Warp must be installed separately according to the chosen engine.
  • The repository snapshot used for this skill did not contain downloaded motion clips, pretrained model files, training logs, or several task object XML assets. Preserve that limitation before promising a runnable simulator job.

Minimal safe checks

For a target checkout, start with the non-simulator checker:

python sub-skills/runner-and-backends/scripts/check_mimickit_layout.py \
  --repo-root <mimickit-checkout>

To preview a runner command without launching a simulator:

python sub-skills/runner-and-backends/scripts/run_mimickit.py \
  --repo-root <mimickit-checkout> \
  --dry-run \
  -- --arg_file args/deepmimic_humanoid_ppo_args.txt --visualize false

For motion tools, prefer the bundled helpers in sub-skills/motion-tools/scripts/ instead of source utility paths. For SMP prior/policy checks, use sub-skills/smp/scripts/train_smp_prior.py --dry-run-config and sub-skills/smp/scripts/check_smp_config.py before launching training.

Verification status to preserve

This generated skill was verified for source syntax/import routing, PyTorch CUDA availability, converter parser help, tiny GMR/SMPL conversion fixtures, plotting helper output, and SMP dry-run/config checks. It was not verified by running full simulator-native training/testing because the external simulator packages and downloaded data/model/object assets were unavailable in the production environment.

Do not auto-import or claim full backend validation from this skill alone. If a user needs full native validation, install the selected simulator backend, restore the required data/model/object assets, and run a focused native case after checking the relevant sub-skill.

Saved skills