Agent Skills › rohanpsingh/LearningHumanoidWalking › eval

eval

GitHub

用于评估训练好的模型检查点，支持解析.pt文件、运行目录或日志目录路径。自动验证路径并执行评估命令，打开MuJoCo可视化窗口展示结果，可配置episode长度、随机种子及输出目录。

.claude/skills/eval/SKILL.md rohanpsingh/LearningHumanoidWalking

Trigger Scenarios

用户请求评估已训练的模型检查点需要查看强化学习智能体在环境中的表现

Install

npx skills add rohanpsingh/LearningHumanoidWalking --skill eval -g -y

More Options

Non-standard path

npx skills add https://github.com/rohanpsingh/LearningHumanoidWalking/tree/main/.claude/skills/eval -g -y

Use without installing

npx skills use rohanpsingh/LearningHumanoidWalking@eval

指定 Agent (Claude Code)

npx skills add rohanpsingh/LearningHumanoidWalking --skill eval -a claude-code -g -y

安装 repo 全部 skill

npx skills add rohanpsingh/LearningHumanoidWalking --all -g -y

预览 repo 内 skill

npx skills add rohanpsingh/LearningHumanoidWalking --list

SKILL.md

Frontmatter

{
    "name": "eval",
    "description": "Evaluate a trained checkpoint with visualization",
    "allowed-tools": "Bash, Read, Glob",
    "argument-hint": [
        "path-or-logdir"
    ],
    "disable-model-invocation": true
}

/eval — Evaluate a Trained Checkpoint

Parse the user's request from $ARGUMENTS and run evaluation.

Command Template

uv run python run_experiment.py eval --path <PATH> [OPTIONS...]

Path Resolution

The user may provide:

A .pt file: Use directly (--path /tmp/.../actor_999.pt)
A run directory: Contains actor*.pt files (--path /tmp/.../26-03-07-00-26-36_cartpole/)
A logdir: Contains timestamped run subdirectories (--logdir /tmp/training_runs)

If no path is given, check /tmp/training_runs for the most recent run.

Use Glob to verify the path exists and resolve it before running.

Options

Flag	Default	Description
`--ep-len`	10	Episode length in seconds
`--seed`	None	Random seed for reproducible eval
`--out-dir`	None	Directory to save videos

Instructions

Resolve the model path from the user's input. If ambiguous, list available checkpoints and ask.
Show the user which checkpoint will be evaluated (full path).
Run the eval command. This opens an interactive MuJoCo viewer window — it is NOT a background job.
Report the results when done.