---

name: training-and-evaluation description: "Safely construct Humanoid-Gym PPO training and evaluation commands, checkpoint loads, and TorchScript export guidance." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

training-and-evaluation

Use this sub-skill when a future agent needs to prepare or explain Humanoid-Gym PPO training and evaluation workflows without crossing into environment design or MuJoCo deployment.

Use this skill for

• Building safe command lines for humanoid/scripts/train.py and humanoid/scripts/play.py.
• Choosing --load_run and --checkpoint when only a run directory or iteration number is known.
• Explaining task_registry, OnPolicyRunner, PPO, ActorCritic, and RolloutStorage behavior.
• Describing checkpoint layout, logging layout, and TorchScript export behavior.
• Checking backend readiness for Isaac Gym-based training/evaluation.

Route elsewhere

• Robot, terrain, reward, observation, or task-definition changes → environment-customization.
• MuJoCo sim-to-sim rollout and exported-policy deployment → sim2sim-deployment.
• Real robot hardware deployment → out of scope.

Safety boundary

• Native train.py/play.py execution still requires Python 3.8, humanoid==1.0.0, Isaac Gym Preview 4, and the matching CUDA/PyTorch stack.
• If Isaac Gym is missing, stop at static guidance and mark native training/play as BLOCKED_REQUIRED_BACKEND.
• play.py hard-codes export/render/command-fixing behavior unless the source file is edited; the bundled builder only surfaces intent.
• The public registered task in this checkout is humanoid_ppo; the parser default XBotL_free is stale.

Bundled files

• references/workflows.md
• references/api-reference.md
• references/troubleshooting.md
• scripts/build_training_command.py
• scripts/build_play_command.py

Operating rule

This skill is command-oriented. It prints or explains safe launch commands and API facts; it does not launch Isaac Gym training itself.